Physical AI

Physical AI refers to artificial intelligence systems that perceive, reason about and act in the physical world through embodied machines such as robots, autonomous vehicles and drones. Unlike software-only models that process text or images, these systems are coupled to sensors and actuators, so their decisions change the state of real objects, machines and people. The term became common in industry and policy discussion around 2024–2026, although its intellectual roots lie in cybernetics and early mobile robotics.

The field matters because robotics has long been limited by brittle, hand-engineered controllers that work only in narrow settings. Large models that combine vision, language and action, trained on growing collections of robot data, have begun to generalize across tasks and hardware platforms. Physical AI sits at the intersection of machine learning, control theory and physics, and it raises practical questions about automation of physical work, ethics, and safety when learning takes place where errors can cause irreversible harm.

This entry first defines the term and separates it from related concepts, then traces its history and describes its technical components: perception and world models, action models, learning methods, physics-informed approaches and hardware. It then covers application domains, industry and investment, safety and governance, labor and social effects, and open research problems.

Definition and terminology

Physical AI refers to artificial intelligence systems that sense their surroundings, decide what to do, and act through a physical body such as a robot, vehicle, drone or industrial machine. In the working definition used here, the defining feature is that the system's outputs are physical motion, force or control signals rather than information alone. This separates it from software-only AI, including most large language models and many generative systems, whose errors are usually informational and can be corrected by rewriting an answer. A language model can describe how to pick up a cup; a physical system must locate the cup, apply a suitable grip force, and carry it without dropping it. Many practitioners describe the core as a closed loop of sensing, modelling, deciding and acting, repeated continuously, with the results of each action feeding back into the next observation.

The term also needs to be distinguished from automation. A robot vacuum that follows fixed rules, such as reversing when a bump sensor fires, senses and moves but does not generate its own behaviour. A common dividing line in industry usage is whether the control policy is written out in advance or learned from data. On this view, physical AI begins where hand-written rules end, though the boundary is blurred in practice because many deployed systems mix learned components with rule-based safety layers.

Several overlapping terms are used, and their boundaries are not fixed. Embodied AI is the older label, associated with an academic research tradition dating to the 1980s, which holds that intelligence depends on a body and on sensorimotor interaction with the world. Physical AI is the newer label, which became common in industry from about 2024 and functions as a broad umbrella. Some sources treat embodied AI as a narrower subset that focuses on agents that learn through their bodies, while others regard the two as heavily overlapping. World models are learned internal representations that predict how an environment changes in response to actions; they are better understood as a technical component used to build physical AI than as a synonym for it. "Physical intelligence" is used both as a general phrase and as the name of specific research and company efforts, which adds to the confusion.

A central unresolved question is whether simulated agents count as embodied. One position requires an actual physical presence, so that only machines with sensors and actuators interacting with matter qualify. The opposing position accepts virtual or simulated bodies, arguing that a digital character or agent in a physics engine also perceives and acts through a body whose constraints shape its behaviour. This entry adopts the criterion of physical constraints rather than "real world" as the defining property. A system qualifies when its decisions are bound by physical laws and limits, such as inertia, friction, material properties, latency, wear and energy budgets, and when mistakes have consequences that propagate through those laws. The phrase "real world" is avoided because many systems are trained in simulation and deployed in reality, and because a simulator that faithfully enforces physical dynamics is doing physical work even if no hardware is present. A game character that ignores physics does not meet the criterion, regardless of how lifelike it appears.

Why the physical setting changes the problem

Physical actions exchange energy and momentum with the environment. A robot arm that pushes an object changes that object's position and may damage it, and the same force also acts on the arm itself. This coupling means that actions are not isolated outputs: each one alters the state the system will face next, and the resulting feedback is part of the learning signal. Standard machine learning often assumes that training examples are independent and that the data distribution stays fixed, and that assumption breaks down when the agent's own behaviour determines what data it collects.

Physical actions also have lasting and often irreversible effects. A text model that produces a poor answer can be queried again at negligible cost, but a robot that collides with a shelf, drops a fragile part, or overheats a motor cannot simply rewind the episode. This rules out the large-scale trial and error on which many successful reinforcement learning results in games and simulations rely, and it makes exploration itself a safety question. It also raises the cost of each data point, which is one reason real robot data remains scarce compared with text and images.

Timing adds a further constraint. Physical control often requires decisions within milliseconds, since a vehicle braking for an obstacle or a gripper compensating for a slipping object cannot wait for a slow response. Inference speed, sensor synchronisation and the reliability of communication links therefore become part of the learning problem, not just engineering details. Taken together, these properties mean that methods developed for software-only AI cannot be transferred directly: the system must be evaluated against physical limits, learn from costly and partly irreversible experience, and remain stable in a world that changes as it acts.

Historical origins

Physical AI draws on two lineages. One is control engineering and cybernetics, which treated machines as parts of closed loops with their surroundings. The other is robotics research from the 1960s onward, which tried to give machines the ability to reason about and act in physical space. Early work in both lineages established the tensions that later learning-based systems inherited.

Cybernetics and feedback

Norbert Wiener's wartime research on anti-aircraft fire is the starting point. During World War II, Wiener was asked to solve a critical problem: how to shoot down enemy aircraft more effectively. A gun that aimed at a plane's predicted position missed evasive pilots, so Wiener reframed aiming as a continuous correction task, in which the gun kept comparing expected and observed trajectories and adjusted itself. He borrowed from electrical engineering, where amplifiers used negative feedback to stabilize signals.

His 1948 book generalized this into cybernetics. Wiener defined cybernetics in 1948 as the science of control and communications in animals and machines. The core model separates measurement from action: a monitor compares a system's current state to a desired standard, and a controller adjusts the system's behavior accordingly. For physical AI, the important claim is that control depends on repeated information exchange between a system and its environment, so intelligence cannot be cleanly separated from interaction with the world. The field's reach was wide. Wiener's program influenced the work of an entire generation of renowned scientists, including Warren McCulloch, Walter Pitts, Claude Shannon, John von Neumann, and Alan Turing.

Shakey and the sense–plan–act pipeline

The best-known early attempt to combine perception, reasoning and action in one machine was Shakey, built at SRI. The subject of SRI's Artificial Intelligence Center research from 1966 to 1972, Shakey could perform tasks that required planning, route-finding, and the rearranging of simple objects. The robot was a mobile robot with cameras and touch sensors, controlled by a large remote computer. Operators could phrase goals in ordinary language, such as pushing a block off a platform, and the system worked out its own sequence of actions. Although not a practical tool, it led to advances in AI techniques, including visual analysis, route finding, and object manipulation.

Shakey's architecture is usually described as a sense–plan–act pipeline: sensors build a model of the world, a planner searches for an action sequence against that model, and actuators carry it out. Planning relied on STRIPS, a formal planner that could update its world model from sensing. The project was ambitious in scope. SRI wanted to integrate in one system representation and reasoning, planning, machine learning, computer vision, natural language understanding, even speech understanding, for the first time. It was funded through ARPA and ended in 1972, not for lack of exciting ideas to pursue, but because the funding climate had changed. The pipeline it embodied dominated robot design for roughly two decades.

Critique and behavior-based robotics

By the 1980s, researchers trying to run symbolic, planning-heavy systems on physical robots encountered serious difficulties. Classical planners typically assumed that sensor data was clean and complete and that only the robot's own actions changed the world. Real environments violated both assumptions, and the resulting systems were brittle.

Rodney Brooks at MIT became the most prominent critic. Brooks argued that the sense-plan-act paradigm used in some of the first autonomous robots was in fact detrimental to the construction of real working robots. He also held that building world models and reasoning using explicit symbolic representational knowledge at best was an impediment to timely robotic response. His 1986 subsumption architecture was proposed in opposition to traditional symbolic AI. Its layers operate more or less independently and in parallel, with the higher levels relying on successful operation of the lower levels, and higher levels could change the behaviour of the lower levels. Brooks's central premise was that the coupling of perception and action gives rise to all the power of intelligence. The approach shifted attention from high-level abstract problem solving to low-level grounded intelligence. It also reached the market: in 1990, Rodney Brooks founded the company iRobot to develop and market the Roomba vacuum cleaning robot.

From hand-coded to learned perception

Through the 1990s and 2000s, robot perception still relied largely on hand-engineered features and calibrated geometric models, which required expert tuning for each sensor and environment. The turning point in computer vision came in 2012, when AlexNet won the ImageNet competition with a deep learning model trained on large labelled datasets. Learned visual representations then spread into robotics for grasping, pose estimation and manipulation, replacing many hand-built pipelines.

Learned vision alone did not resolve the original problem, however. Benchmarks used static images, while a robot must close the loop between what it perceives and what it does next. That gap motivated work on reinforcement learning, imitation from demonstrations, and eventually models that link perception directly to action. The phrase "physical intelligence" became a mainstream label around 2024, when companies and researchers began using it for general-purpose robot models. Physical Intelligence, a startup founded that year, is a prominent example of this usage.

Perception and world models

Physical AI systems must first turn raw sensor streams into a usable description of their surroundings and their own bodies, and then predict what their actions will do. Perception therefore has two linked jobs: estimating the current state, and modelling how that state changes. The four topics below cover the sensing hardware, the spatial and physical reasoning built on it, the learned predictive models known as world models, and the probabilistic methods that handle uncertainty.

Robots combine several kinds of sensing. Exteroceptive sensors such as cameras, depth sensors and LiDAR observe the external scene, while proprioceptive sensors measure the machine's own state, such as inertial measurement units, joint encoders and force or torque sensors. Tactile sensors add contact information that vision cannot reliably supply. Sensor fusion combines these streams into a single state estimate. In legged robots, for example, proprioceptive data arrive at high frequency but accumulate drift, while slower external sensors can correct that drift; a common approach is an extended Kalman filter that fuses inertial, kinematic and stereo-vision or LiDAR measurements into a pose estimate. Recent work on tactile–proprioceptive fusion in a skin-covered arm combines pneumatic tactile pads, which indicate where contact occurs, with motor-current measurements that estimate force, and uses a temporal convolutional network to correct friction effects during motion.

Perception in physical settings also requires three-dimensional spatial understanding. A robot must infer geometry from depth and multi-view data, recognise where an object's surfaces lie when other objects block part of it, and estimate contact points and interaction forces. Some research has shown that fusing visual tracking with force and torque data through factor-graph inference can keep estimates robust under severe occlusion. Scene representations used in robot learning increasingly include RGB-D inputs, three-dimensional scene flow and Gaussian-based models of dynamic scenes, which allow future states to be predicted in space rather than only in the image plane.

A world model, in the robotics sense, is a learned predictive model of how an environment and the agent interacting with it evolve over time. The concept dates to Jürgen Schmidhuber's 1990 use of the term in machine learning, and it was revived by Ha and Schmidhuber's 2018 work on learned generative environments. Formally, the model maps a current state and an action to a predicted next state, but real robots rarely observe the true state directly, so modern systems usually predict the next observation, such as an image or depth map, from a history of observations and actions. The most useful variant in embodied control is the action-conditioned model, because a predicted future that looks plausible but does not respond correctly to the candidate action gives little help in closed-loop decisions. Predictions may be made in pixel space, in a learned latent space, or in a feature space built from pretrained visual embeddings, which is cheaper and less sensitive to noise than predicting every pixel. Once learned, such a model supports planning by evaluating candidate action sequences, and supports "imagined" rollouts in which a policy is improved without physical trials. Some architectures decouple the process: a generative model predicts a future visual sequence, and a separate module infers the actions that would produce it.

Sensor data are always noisy and incomplete, so perception is inherently probabilistic. Bayesian filters maintain a belief over the state and alternate between a prediction step driven by a motion model and an update step driven by measurements. The Kalman filter, introduced by Rudolf Kalman in 1960, is the optimal estimator for linear systems with Gaussian noise. Its extended and unscented variants handle nonlinear sensor models, but they assume the posterior is roughly Gaussian and can diverge if the motion model is badly wrong. When the belief is multimodal, for example after a robot is lost or when landmarks are ambiguous, particle filters and Monte Carlo localization represent the posterior with weighted samples. They are more robust in such cases but require more computation. In practice, designers tune process and measurement noise: if measurements are assumed much noisier than the model, the estimator smooths heavily; if the reverse, it trusts the sensors more. Learned world models add a further layer of uncertainty, and their predictions are usually evaluated against the estimated belief rather than treated as certain.

Action models and robot foundation models

Vision-language-action models

A vision-language-action (VLA) model takes camera images and a natural-language instruction as input and outputs robot actions from a single network. Spatial reasoning, language understanding, and action prediction are implicitly learned within a unified representation space, benefiting from large-scale pretraining. Earlier robot software usually split perception, planning and motor control into separate modules. High-level reasoning and low-level manipulation systems played an imperfect game of telephone to operate the robot. VLA models try to remove that hand-off.

The label is most often traced to Google DeepMind's RT-2, announced in July 2023. RT-2 was the first to explicitly propose the Vision-Language-Action model paradigm. Its central technique is to treat motor commands as language. Robot actions are represented as a second language that can be converted into text tokens and trained alongside large-scale vision-language datasets. The model was built on existing vision-language backbones; Google's existing Vision-Language Models such as PaLM-E and PaLI-X served as its backbone network. Its robot data came from the RT-1 collection, gathered with 13 robots over 17 months in an office kitchen environment. According to DeepMind, the transfer from web data made it possible to show improved generalisation and semantic and visual understanding beyond the robotic data it was exposed to.

Representative systems and design choices

RT-2 established the pattern of co-training on web data and robot trajectories, and its outputs are discrete action tokens. Because it inherits web-scale semantic knowledge while emitting executable end-effector deltas, it is often cited for generalization to novel objects and multi-step tasks. RT-2 was not released as open weights.

OpenVLA was introduced in 2024 as an open alternative. It is a 7B-parameter open-source VLA trained on a diverse collection of 970k real-world robot demonstrations. Its visual encoder fuses two pretrained feature extractors, and it builds on a Llama 2 language model combined with a visual encoder that fuses pretrained features from DINOv2 and SigLIP. Continuous actions are handled by quantization: each dimension of the action space is discretized into 256 bins, producing integer-valued tokens for each action dimension. In reported comparisons, OpenVLA outperforms closed models such as RT-2-X (55B) by 16.5% in absolute task success rate across 29 tasks and multiple robot embodiments, with 7x fewer parameters. Its authors also report that joint training on cross-embodiment datasets enables efficient robot-specific fine-tuning, requiring only 10 to 150 trajectories per robot to optimize performance. These results come from specific benchmarks, and their transfer to other settings is not established.

NVIDIA's GR00T N1, announced in March 2025, is the first model in a family that NVIDIA describes as fully customizable models that NVIDIA will pretrain and release to worldwide robotics developers. Its main design is a dual-system architecture. "System 1" is a fast-thinking action model, mirroring human reflexes or intuition. Powered by a vision language model, System 2 reasons about its environment and the instructions it has received to plan actions. Training data mixes egocentric human videos, real and simulated robot trajectories, and synthetic data. The developers state that the model supports multiple robot embodiments, and it has cross-embodiment support from tabletop robot arms to dexterous humanoid robots.

Physical Intelligence's π0, released in 2024, couples a pretrained vision-language backbone with a separate action module that is trained to generate continuous action sequences via flow matching. Its design aims at general-purpose manipulation across tasks and robot types.

Action generation: flow matching and diffusion

Discrete action tokens are one way to output robot motion; continuous generative methods are another. Diffusion-based policies learn to convert noise into an action sequence through iterative denoising, and this makes them suited to multimodal demonstrations. OpenVLA was reported to outperform expressive from-scratch imitation learning methods such as Diffusion Policy by 20.4% in one set of fine-tuning experiments, which illustrates that the choice of action representation strongly affects results. Flow matching is a related approach that learns a vector field transporting samples from a simple distribution to target actions, and it is often used to produce smooth trajectories in fewer integration steps. Many recent systems predict short sequences of future actions, known as action chunks, rather than single steps, which reduces jitter and the frequency of model calls. The flow matching and diffusion families therefore compete mainly on smoothness, speed and how well they handle multimodal behavior.

Inference speed is a practical constraint. The open-source OpenVLA codebase added a new action tokenizer that allows action chunks to be compressed into fewer tokens, speeding up inference by up to 15x when using discrete robot actions.

Cross-embodiment training and hierarchical control

Cross-embodiment training pools demonstrations from robots with different arms, grippers and sensors into one model. The Open X-Embodiment collection, which OpenVLA draws on, is a large example of such pooling. The approach aims to transfer skills across bodies, though differences in kinematics, camera placement and control frequency still require adaptation. Fine-tuning on a new platform with a modest number of trajectories is the usual method for closing that gap, as noted in the OpenVLA results above.

Hierarchical designs split the problem between a slower planner and a faster controller. GR00T N1's System 1 and System 2 follow this pattern, with the planner interpreting instructions and the controller executing motion at high frequency. Hierarchy can make the system easier to inspect and to swap components, but the interface between levels becomes a point where information can be lost, which is the failure mode that end-to-end VLA designs were intended to avoid. Both approaches are active research directions, and no consensus has been reached on which is preferable across tasks.

Learning from interaction and data

Imitation learning from demonstrations

Imitation learning trains a policy by fitting it to recorded examples of a task being performed, usually by a human. Teleoperated demonstrations have been found to narrow the mismatch between training data and deployment conditions, and collecting them can be scaled with relative ease. Teleoperation interfaces range from leader-follower arms to virtual-reality rigs and exoskeletons, but they carry their own costs. Teleoperation remains limited by restricted operator feedback, system complexity and latency. Humanoid platforms are harder still, because human operator commands cannot be easily mapped onto a robot with a different body.

The main weakness of the method is that a policy is bounded by its teacher. Supervised imitation makes the quality of the demonstrations cap the performance of the resulting policy, so demonstrations act as an upper bound that the policy cannot exceed. A related difficulty is covariate shift: small errors move the robot into states that never appeared in the training set, where the policy has no learned response. Narrow demonstrations, such as those covering one object or one table layout, therefore tend to produce policies that generalize poorly. Such approaches require a lot of high-quality data and show minimal generalization.

Reinforcement learning on hardware and in simulation

Reinforcement learning avoids the ceiling of imitation by letting the agent improve from its own experience. A reward function supplies feedback on performance, and the agent uses that feedback to optimize its behavior. The method is data-hungry: it allows free exploration of the state-action space, but often takes a long time to converge. Reward design is the central engineering problem. A reward that is too sparse gives little learning signal, while one that is too loosely specified can be satisfied by behavior the designer did not intend.

On physical hardware, every trial has a price. Failed grasps wear gears, break fingers and can damage the workspace. Physical hardware is needed both for safe human-robot interaction and for evaluating learning outcomes, which makes dexterous manipulation an especially costly research area. For this reason most large-scale reinforcement learning for locomotion and manipulation is run in simulation first, with the results then transferred to real machines.

Sim-to-real transfer and domain randomization

Simulation offers cheap, parallel and safe trials. Robotic manipulation tasks benefit from simulation-based training because of its advantages in safety, cost and data availability. The obstacle is the reality gap between simulated and physical dynamics. This gap poses a significant challenge for transferring a learned policy.

Domain randomization is the most widely used response. Simulation parameters are sampled from broad ranges during training so that the real world looks like one more variation. Parameters such as mass, friction or camera angle are randomized to make the resulting policy more robust. Randomization can be applied to visuals, to dynamics, or to both. Visual randomization and dynamics randomization are the two main kinds. In one example, researchers randomized object dimensions, link masses, surface friction coefficients, joint damping and actuator force gains to learn in-hand manipulation on a physical five-fingered hand.

The method has known limits. The ranges are typically set by trial and error. The randomization is usually tuned empirically, which makes it inefficient. Choosing the ranges also requires expert knowledge. Randomization cannot compensate for effects that the simulator does not model at all, such as complex contact, deformation of soft materials and some kinds of friction, which is why these are listed among the open problems below.

Scarcity of real robot data and large open datasets

Robot learning lacks the corpora that drive language and vision models. Real robot trajectories must be collected on hardware, one at a time, and they are costly to produce. To address this, the Open X-Embodiment effort pooled data from several laboratories. It contains more than one million real robot trajectories spanning 22 robot embodiments, and it was built by pooling 60 existing robot datasets from 34 robotic research labs. The collection covers more than 500 skills and 150,000 tasks. The project is community-driven, and its contributors hope to broaden participation and grow the dataset over time.

Large datasets do not remove the need for task-specific data. Evaluations of a generalist model fine-tuned on one embodiment suggest that extensive pre-training primarily enhances the model's capacity to execute more complex tasks, but contributes less to reducing the data volume or diversity needed for effective fine-tuning. In-context learning, in which a model adapts to a new task from a single demonstration supplied at inference time without gradient updates, is an active attempt to reduce this dependence. Its reliability across varied tasks and hardware remains an open question.

Physics-informed and generative physical intelligence

Physics-informed AI

Physics-informed approaches place known physical knowledge inside the learning process rather than expecting a model to infer all regularities from examples. The best-known form, physics-informed neural networks (PINNs), is a class of neural networks that writes differential equations into the loss function, so training favours solutions that are consistent with the underlying physics incorporate physical laws described by differential equations into their loss functions to guide the learning process toward solutions that are more consistent with the underlying physics. Derivatives of the network output with respect to its inputs are obtained through automatic differentiation, and the governing equations (for example partial differential equations) are enforced at sampled points alongside any measured data.

Because the equations supply information that labelled measurements would otherwise have to provide, such models can need less data. One review of PINNs in physiological signal analysis reports that PINNs require fewer data points and labels to learn compared to a traditional data-driven approach. Constraints can also be built into the architecture itself rather than applied as a penalty. One 2025 variant, Ψ-NN, first trains a teacher PINN, then extracts parameter relations such as weight symmetries and rebuilds a student network with the physics hard-wired into it, covering symmetries and conservation laws automatically discovers and embeds physical structures directly into the network architecture, e.g., symmetries, conservation laws.

Soft penalties do not guarantee compliance. A black-box model can still break conservation of mass even when the governing PDE is included as a soft constraint in the loss, which is why hard constraints or hybrid designs are pursued in some work. Hybrid "grey-box" designs combine first-principles equations with machine-learning components, so that the physics structure ensures physically plausible extrapolation.

Limits of purely data-driven models

Unconstrained learning models frequently generate outputs that contradict known physics. Without explicit constraints, they are prone to make physically implausible predictions, violating the governing physical laws. Extrapolation is the central weakness. Most deep learning models interpolate reliably within the parameter ranges seen during training but struggle beyond them. In battery prognostics, data-driven forecasts can exhibit what researchers call "physical hallucination", such as predicting spontaneous capacity recovery, which is thermodynamically impossible they frequently suffer from "physical hallucination" predicting thermodynamically impossible phenomena such as spontaneous capacity recovery or inverted temperature dependence when extrapolating beyond training data. Generative models of physical systems show similar failures: a sample may look statistically plausible while drifting in phase over long trajectories or violating invariants and boundary conditions a generated sample may remain statistically plausible while developing phase drift in long-time trajectories.

The remedy is not automatic, and it depends on knowing the relevant physics. In automated driving, purely data-driven models struggle with robustness and out-of-distribution scenarios, especially for rare edge cases which are underrepresented in the training data. Physics-informed variants have been reported to extrapolate more accurately, but only a few studies have applied such networks to the use case automated driving because of lack of reliable physical models for the complex human-driving patterns.

Generative physical intelligence

Generative physical intelligence uses generative models to produce states, trajectories, environments and scenarios that respect physical regularities, chiefly to create training and test data. Synthetic data from simulated environments or Digital Twins offers scalable, risk-free training opportunities covering multiple edge cases. The physical grounding usually comes from simulators. Modern simulation platforms, powered by high-fidelity physics engines such as PyBullet, MuJoCo, CoppeliaSim, NVIDIA Omniverse and Unity, enable accurate modeling of physical interactions, including friction, contact, and collisions.

Generative models offer an alternative route. Unlike simulations, generative models can produce realistic data without relying on physics engines. That flexibility means physical plausibility rests on what the model has learned rather than on an explicit law, so pipelines often combine both approaches. NVIDIA's workflow renders randomized simulated scenes and then uses generative models to augment, curate, and annotate the generated data to scale a single scenario into hundreds. Transfer to hardware remains the test. Simulated and real environments differ, and researchers have developed techniques such as domain randomization to address this, introducing variability in visual and physical properties. In one reported pipeline for assistive robot tasks, learned policies reached success rates exceeding 80% and demonstrated resilience to unscripted human motion in zero-shot transfer.

Scientific applications and consistency methods

The same family of methods is used in science. PINNs have been applied across diverse engineering domains, including fluid dynamics, structural mechanics, heat transfer, biomedical modeling, materials science, and geophysics. Climate modelling is a further case where accuracy under new conditions matters, since purely data-driven approaches in this field face a struggle with out-of-distribution situations, which is the same weakness seen in engineering.

Several techniques keep learned predictions consistent with physical laws. Residual formulations use a physics model as a baseline and let the network learn only the correction; in this sense residual models use physics equations to improve generalization. Constraint penalties, architectural constraints and hybrid grey-box models are the other main approaches. Extrapolation to long time horizons, unseen parameters or new geometries remains difficult, as such physical extrapolation remains a central difficulty for data-driven methods, especially when the target regime involves long-time evolution, unseen system parameters, or geometries absent from the training set.

Hardware and onboard computing

Hardware sets the outer limits of what a physical AI system can do, and many of the constraints that decide commercial deployment come from the body and its power supply rather than from the model. The first choice is the platform. Wheeled bases are mechanically simple and stable on flat floors, which suits warehouses and many service tasks. Humanoid platforms are built for human-scale spaces, such as stairs, shelves and tools designed for people, but they need more joints, more balance control and more power.

Actuators and dexterous hands

Most joints are driven by electric motors working through gearboxes or, in some newer designs, through direct or quasi-direct drives. Each option trades off strength, speed and compliance. High gear ratios produce torque but add friction and backlash, and make a joint harder to back-drive, so the robot cannot easily sense external force through its own joints. Boston Dynamics' 2026 hand for its Atlas robot illustrates the newer approach: four fingers, 13 degrees of freedom, a single actuator type, and actuators enclosed so that no cables cross the joints. The company stresses backdrivability, which lets the controller rely on proprioception, and adds dense pressure sensors on the fingertips and palm.

The human hand remains the benchmark. It offers more than 20 degrees of freedom and can exert grasping forces above 400 N at roughly 0.4 kg, and no commercial dexterous hand matches that combination. Designers therefore move the bulky motors out of the hand and into the forearm or torso, driving fingers through tendons or linkages. This lowers distal mass but introduces cable friction, wear and maintenance problems. Many hands are also built around a compromise of finger count: fewer fingers simplify the mechanism and improve durability, while more fingers improve in-hand manipulation.

Energy and runtime

Battery capacity is one of the tightest limits on mobile robots and vehicles. The runtime of a battery-powered machine can be estimated as

where $E_{\text{usable}}$ is the usable stored energy and $P_{\text{avg}}$ is the average power draw, which includes actuators, sensors and computing. In 2026, most humanoid robots run for two to four hours on a charge, against industrial shifts of eight to twelve hours. The Unitree H1 carries a battery of about 0.864 kWh, while the Figure 03 is reported with a 2.3 kWh pack and up to five hours of runtime under stated conditions, a figure that depends heavily on walking, lifting and manipulation.

Adding cells is not a free fix. A heavier pack must be carried and powered by the same actuators, which partly cancels the extra energy. Mainstream humanoid cells use high-nickel NMC or NCA chemistries, with energy densities around 250–300 Wh/kg, while lower-cost LFP cells, at roughly 160–200 Wh/kg, are used where endurance matters less. Solid-state designs are widely presented as the next step, but their advertised energy densities are forecasts and vendor targets rather than established figures. Mobile robots and cars face the same arithmetic: the computing load also draws from the battery, so an onboard inference choice is also an energy choice.

Onboard computing and the edge–cloud trade-off

Robots can carry only so much compute. Weight, cost and power limit them to embedded GPUs of the Jetson class rather than data-centre accelerators. A control loop that must respond in milliseconds cannot wait for a network reply, but a large vision-language-action model may not fit on a small board either. Recent measurements show the trade-off clearly. One study of vision-language-action inference found that a server with a high-end GPU could give up to about three times lower end-to-end latency than an on-robot embedded GPU, even after paying the network cost, and that on-robot inference could cut battery life and operating time by up to roughly 45 percent. Another comparison of a navigation model found a remote server slightly faster than an onboard Jetson Thor, at about 196 ms versus 204 ms. These results depend on network quality. On-device execution remains preferable on slow cellular links or distant data centres. Practical systems therefore split the work: reflexes, balance and collision avoidance run on the robot, while heavier planning and fleet tasks may run on a nearby edge server, with a local fallback that keeps the robot safe when the connection drops.

Co-design and economics

Hardware and software are designed together because each layer constrains the others. The compute budget limits model size, the power budget limits compute, and the battery limits the power budget. Sensor selection follows the same logic. Cameras, depth sensors, tactile arrays and joint encoders each have a unit cost, a failure mode and a set of environmental limits such as dust, rain or lighting. A cheaper sensor that fails often can cost more in downtime than a dearer, more reliable one. Whether an application is economically viable therefore depends on the combination of unit price, uptime, runtime per charge and the labour cost of the task being replaced.

Application domains

Autonomous vehicles and robotaxis

Road vehicles are the most visible commercial use of physical AI. An autonomous vehicle must turn camera, lidar, radar and satellite-positioning data into a continuously updated picture of traffic, then choose steering, braking and acceleration in real time. Deployed stacks have traditionally been split into perception, prediction, planning and control modules. Some newer systems replace parts of this chain with a single learned network that maps sensor input more directly to driving commands, which can improve smoothness but makes the decisions harder to audit.

Robotaxi service has grown quickly in the United States. Waymo was providing 500,000 paid robotaxi rides every week across 10 U.S. cities in March 2026. By September, the count of driverless paid metros had risen further: one report says the company will begin offering paid robotaxi rides to the public in Denver, San Diego, and Tampa, Florida, bringing the number of US cities where its driverless ride-hail service is available to 14, while a later tracker counts 15 metros after Las Vegas opened. The company's stated goal is 1 million weekly rides by the end of 2026. The hardest part of deployment is the long tail of rare situations, such as unusual road users, construction zones and adverse weather, and this is what determines how fast a service can enter a new city.

Warehouse and logistics

Warehouses are among the most mature sites for robot manipulation, but the task that consumes the most labor is still the hardest to automate. Automation has long handled pallets and uniform containers well, whereas picking individual items from shelves or bins remains difficult. The main obstacle is variety. Stock keeping units differ in size, weight, surface and fragility, and a single order may combine a glass bottle, a soft package and a small electronic part. Suction arms handle a growing share but still fail on irregular, fragile and deformable items, which is why humans remain at picking and packing stations in even the most automated buildings.

Deformable goods are a particular problem. Vacuum cups are efficient for many boxes and bags but struggle with porous, perforated or heavily wrinkled surfaces. Rigid-looking objects can also be difficult: a shoebox has defined edges, yet its loose lid can shift during handling because of small differences in overlap, orientation or friction. Systems that succeed in this area combine vision, grasp planning and tool choice, and they are usually paired with human workers who handle exceptions such as damaged items and jammed equipment.

Manufacturing

Factory work is more structured than field work, which makes it an early target for physical AI, but the tasks differ considerably in difficulty. Machine tending, in which a robot loads and unloads a fixed machine from a repeatable fixture, is comparatively tractable. Assembly is harder because it involves contact-rich operations such as inserting connectors, fastening parts and holding tight tolerances, where small force errors can cause failure. Inspection benefits from computer vision and can be deployed incrementally. High-mix, low-volume production is the most demanding setting, because frequent product changes require re-fixturing, reprogramming and new training data, and the economic case depends largely on how quickly a line can be changed over.

Agriculture

Farming is a clear example of an environment that is difficult to automate. A farming environment is unstructured, cluttered and heavily occluded, in contrast to a factory. Fruit and vegetable plants vary in shape, lighting changes through the day, and leaves hide the target. Labor scarcity and seasonal demand make harvesting an economically attractive target, yet there is still no commercially available autonomous harvester, according to a 2023 survey. Research prototypes show the scale of the problem. One imitation-learning pepper harvester trained on 300 demonstrations reported a harvesting success rate of 28.95% with a cycle time of 31.71 seconds, which demonstrates feasibility but remains far from commercial speed and reliability. Soft grippers are often proposed for delicate produce.

Construction, mining and service robots

Construction sites change daily, with unfinished structures, uneven ground, moving equipment and frequent changes to the plan. Tasks such as drilling, bricklaying or drywall installation are repetitive in principle, but each instance has different geometry, and robots must localize themselves without reliable satellite positioning indoors. Mining has similar localization and communication challenges underground, together with dust, poor lighting and harsh conditions for sensors and actuators. Open-pit haulage is more advanced than underground work because the environment is larger and more predictable.

Service and social robots operate among people in homes, hospitals and public buildings. Their environments are cluttered, varied and designed for human use rather than for machines, so stairs, furniture and moving people all count as hazards. Household manipulation, including handling cloth or dishes, involves deformable and fragile objects, and safety around people constrains speed and force. Cameras in these spaces also raise privacy questions, which limit where such systems can be deployed.

Industry, markets and investment

Market size estimates

Forecasts for physical AI diverge sharply, and the spread mostly reflects how each firm draws the boundary of the category. A narrow definition gives modest figures. One projection for AI robots, which counts hardware and software together, grows from USD 6.11 billion in 2025 to USD 33.39 billion by 2030, at a CAGR of 40.4%. A separate analysis of the embodied-AI segment alone estimates growth from USD 4.44 billion in 2025 to USD 23.06 billion by 2030. Broader definitions that include autonomous vehicles, smart infrastructure and semiconductors produce much larger totals. One report values the market at USD 30.1 billion in 2025 and expects USD 641.6 billion by 2035, while another gives USD 81.4 billion in 2025 and USD 1145 billion by 2035. A further estimate expects approximately $383 billion in 2026 to reach $3.26 trillion by 2040.

Humanoid-specific projections are similarly unstable. Goldman Sachs estimates the humanoid market will reach USD 38 billion by 2035 as a standalone segment, whereas other investment banks have projected the global humanoid market could hit USD 5 trillion in the coming decades. These figures are commissioned or published by research firms with differing scope, base years and assumptions, so they are best read as scenario estimates rather than measured revenue.

Capital concentration in foundation and world-model developers

Venture funding has concentrated in companies building general-purpose models for the physical world, often described as world models. Six relatively pure world-model companies announced about $3.05 billion of major rounds in 2026, with World Labs and AMI Labs alone accounting for roughly two-thirds of that total. AMI Labs, co-founded by Yann LeCun after he left Meta, raised $1.03 billion at a $3.5 billion pre-money valuation. Its investor list includes Headline, NVIDIA, Temasek Holdings, Cathay Innovation, and Samsung Electronics, which shows the mix of financial and strategic backers typical of this segment. The company's CEO has also warned that the label is becoming a fundraising device: "In six months, every company will call itself a world model to raise funding."

The financing pattern resembles that of frontier artificial-intelligence labs more than that of ordinary software start-ups. Large seed and early rounds, milestone-linked tranches and compute-heavy plans are common. Adding Wayve's $1.2 billion Series D, Waabi's $750 million Series C and Runway's $315 million Series E pushes the total above $5.3 billion before counting milestone-based future commitments. Chip and platform suppliers also invest directly. NVIDIA's venture arm and Sapphire Ventures joined the cap table of Reactor, a company whose total funding stood at $74 million, and the investment gives NVIDIA a stake in the infrastructure for running interactive models.

Layered industry structure

The sector is organised in layers. At the top are model and platform providers that supply pre-trained perception, world-model and control systems, simulators and onboard compute. Applied firms sit below them and combine hardware, software and integration to deploy robots in specific sites. NVIDIA illustrates the platform layer, since it has developed its own Cosmos world models, alongside Isaac tools for robot simulation and Jetson computers that run AI on robots. Licensing pre-trained models to hardware makers lets applied firms avoid building a full model stack, which lowers entry costs for them but concentrates bargaining power in the model layer. Operating models also differ. Robotics-as-a-Service (RaaS) models enable small and medium-sized enterprises (SMEs) to access robotics without the need for heavy capital investment in robot fleets. Analysts also see a regional division of labour: the U.S. excels in software, China in manufacturing, Japan in deployment, and Europe in industrial physics. That characterisation is contested, but it captures why supply chains are split across borders.

China's public-market expansion and manufacturing base

Chinese humanoid makers are moving quickly toward public listings. China has fast-tracked approval for humanoid company Unitree to list in Shanghai, while Hong Kong processes applications from more than 500 companies across sectors. Unitree's stock then made a strong debut: Unitree's shares closed up 460% on the first day of trading, which the NBC report attributes to investor enthusiasm for domestic AI beneficiaries. AgiBot has been reported to eye a valuation of HK$40 billion to HK$50 billion, with the company expected to issue 15%-25% of its shares. Smaller firms are also seeking capital. LimX Dynamics raised $200 million in a pre-IPO round that values the company at $2.21 billion, and Robotera has raised over $340M in 2026 while planning a Hong Kong listing. Listing has become part of the financing sequence rather than its end point, and analysts expect competition to intensify. Morgan Stanley noted that more industrial and collaborative robot companies could come to IPO, so competitive pressure is likely to persist.

The same domestic base is central to the global supply chain. EngineAI, for example, opened a 12,000-square-metre factory in Shenzhen and says the line can turn out a humanoid every 15 minutes, and is geared for 10,000 units. Policy friction is also rising. Washington banned imports of new foreign-made robots over national security concerns, which may push purchasers toward separate supply chains.

Gap between demonstrated capability and commercial scale

Announced capacity, funding and valuation run well ahead of verified deployments. Warehouse and surgical systems are regarded as reliable, while general-purpose humanoids are not yet ready for mass adoption. Companies frequently report pilot placements and targets instead of audited revenue or unit counts. Robotera, for instance, reports that it has already placed robots across more than 10 logistics centers and projects deliveries of more than 1,000 units in Q2 2026, which are projections rather than settled totals. A trade report on EngineAI's listing put the issue bluntly: "The supply is real; the demand is not proven."

The same gap appears in world-model finance. Wayve and Waabi can already show world models inside serious autonomy programs; the broader labs still need to prove that their simulations give robots or agents a large, repeatable advantage across unrelated environments. Until deployment counts, repeat orders and buyer outcomes are reported on a consistent basis, valuations and market forecasts will remain weaker evidence of commercial scale than the funding figures suggest.

Safety, evaluation and governance

Safety risks in physical AI come from the way learning systems improve. A policy that acquires skills through trial and error may act unsafely before its errors are detected. In a simulator, a failed grasp costs nothing. On a factory floor or in a home, the same failure can pinch a worker, damage equipment, or cause a fall. Some actions cannot be reversed once they occur, and a single misjudgment can cascade: a shifted object changes the sensor input, which triggers a wrong correction. Established robotics has handled these risks mainly through physical safeguards. The technical specification ISO/TS 15066 was introduced in 2016 to support practitioners in ensuring safety in human–robot interaction. The 2025 revision of ISO 10218-2 moved the focus from the robot to the application: only the application can be designed, verified, and validated as collaborative, rather than the robot. Standards for humanoids and other dynamically stable machines are still being written. One practical guide notes that ISO/CD 25785‑1 for dynamically stable robots is in draft, while ISO 13482:2025 addresses service robots and emphasizes sensor reliability, uncertainty management protocols, and decision verification through multiple sensing modalities.

Evaluation is a second weak point. Benchmark results in simulation are attractive because they are repeatable, but they do not always predict physical performance. The advantage of using simulation is that reproducibility is guaranteed for evaluation, and a large number of tasks can be used. However, due to the sim-to-real gap, the performance in simulation cannot simply be transferred to the real world. Visual fidelity is a particular problem: many widely used simulation frameworks lack high-fidelity visuals, causing a distribution shift that pushes vision-based policies trained on real data out of distribution. Real-world testing is the alternative, but it is expensive. Measuring generalization on hardware can require hundreds of trials, which is hard to sustain and to repeat. Newer benchmarks try to close this gap by fixing the physical setup. RoboDojo, for example, standardizes the hardware setup, scene reset procedure, evaluation protocol, and deployment interface, so that policies are tested under consistent conditions. Reproducibility across different robot bodies, lighting, and objects remains an open problem, and results obtained on one platform are often not comparable with those from another.

Verification and interpretability address a different question: how to show that a learned controller will behave acceptably across the cases it will meet. Classical control offers analytical guarantees that neural policies generally lack. Proposed approaches include formal verification of controller behavior within bounded conditions, runtime monitors that override unsafe commands, and explainable AI methods that help auditors understand why a policy chose an action. Certification regimes tend to specify what must be documented rather than how a model must reason. Under the EU AI Act, high-risk systems must meet requirements that include risk management, data governance, technical documentation, transparency, human oversight, accuracy, robustness, cybersecurity, and post-market monitoring. Whether a given robot falls into that category is still contested. The Commission's classification rules under Article 6 were, as of the latest available sources, being drawn up; the consultation deadline has been extended to 23 July 2026, with the final version expected by the end of 2026.

Liability and legislation are developing mainly in the European Union. The revised Product Liability Directive treats software and AI as products, which means that compensation can be sought without proving fault; EU member states have until 9 December 2026 to transpose the directive; it applies to products placed on the market or put into service after that date. By contrast, the earlier proposal for a dedicated AI liability rule was dropped: the European Commission formally withdrew its proposed AI Liability Directive (COM(2022) 496) on 6 October 2025. The legislative framework for machinery is also changing. The Machinery Regulation will replace the old Machinery Directive on 20 January 2027 and, for the first time, contains explicit requirements relating to AI/self-learning functions, cybersecurity, human-robot collaboration and autonomous mobile machines. Overlapping obligations create friction: developers of autonomous inspection robots have said their navigation systems could require certification under both the Machinery Regulation and the AI Act, leading to double assessments and higher costs. Cybersecurity is also part of the picture, since connected robots and AI-enabled controllers face a September 2026 deadline under the Cyber Resilience Act. Liability questions in other jurisdictions are generally handled through existing product-safety and tort rules, and the allocation of responsibility among manufacturers, integrators, software providers and operators, together with the ethical principles behind it, remains an unsettled subject across most legal systems.

Labor and social effects

Displacement and task exposure

The most detailed evidence on job loss concerns industrial robots rather than humanoids or learned controllers. Acemoglu and Restrepo's study of US labor markets estimates that one more robot per thousand workers reduces the employment-to-population ratio by 0.2 percentage points and wages by 0.42%, for the period 1990 to 2007. Their model separates a negative displacement effect on workers who lose their jobs from a positive productivity effect across the wider economy. Other settings give different results. A robot meta-analysis finds that at the industry level, a 20% increase in robot installations reduces employment by about 3.2%, while firm-level data show a slightly positive effect. In Germany, Dauth and colleagues found no adverse effects of industrial robots on aggregate employment, as workers displaced in manufacturing reallocated to services.

The harm is concentrated in routine work. Robots are best suited to routine manual tasks such as welding, assembly, handling materials, and packaging, so exposure is more accurately assessed at the level of tasks than of occupations. Meta-analytic work also cautions that aggregate figures may hide the underlying dynamics of job displacement in robot-adopting sectors and the compensatory reallocation elsewhere in the economy.

Working conditions

Displacement is not the only labor effect. A study of warehouse robots reports that severe injuries decline by 40%, but non-severe injuries such as sprains, strains, and repetitive-motion problems rise by 77%. The same research notes that pick rates sometimes two to three times higher can accompany automation, which raises the risk of burnout. The authors argue that automation must be paired with better job design, task rotation, and realistic performance goals. Automation can therefore shift risk between categories of injury rather than remove it.

Reskilling and transitions

Evidence on how workers move into new roles is thinner. A GAO review describes a California agricultural company that automated its fruit packing and sponsored in-house training to retain about 90 percent of their affected employees, who learned how to operate, clean, and maintain the new machinery. A study cited in the same review warns that without reskilling, a quarter of at-risk workers could see their income drop. A small case study of four automotive firms in Portugal found no negative impact on employment, since the destruction of some jobs was compensated by the creation of new ones. Its sample is too small to generalize, but it illustrates that reskilling programs shape outcomes.

Productivity and deployment conditions

The productivity case depends on deployment speed, cost and task fit. A recent arXiv review argues that the risk of direct displacement is greatest where work is repetitive, predictable, measurable, physically demanding, performed at high volume, and embedded in a controlled environment, and is lower where tasks require irregular movement, tacit knowledge, responsibility for ambiguous outcomes, social trust, or frequent adaptation to exceptional circumstances. The same review notes that even an apparently autonomous system depends on installation, remote supervision, cleaning, repair, data annotation, software updates, energy, and human intervention. Productivity gains from automation thus come with new forms of human work, which affects how much of the gain reaches economic growth in practice.

Privacy and surveillance

Robots that rely on cameras and depth sensors capture far more than their task requires. In workplaces, the same sensing used for navigation or quality control can record individual pace and movement, creating monitoring concerns. In homes, sensors may capture interiors, family members and visitors. In public spaces, bystanders who have not consented may be recorded. Key questions include what data is retained, who may access it, and whether it may be reused for purposes other than the original one. These issues are addressed through data privacy rules and sector-specific legislation, which differ widely between jurisdictions.

Trust and anthropomorphic design

Whether human-like design helps is contested. A study of a collaborative industrial robot found that attribute-based trust measures revealed no beneficial effect of anthropomorphism but even a negative impact on the perceived reliability of the robot. In that study, trust behavior showed a relatively steeper decrease after participants experienced a failure of the robot, and the study clearly reveals a distracting effect of anthropomorphic robot design. By contrast, an experiment that added an anthropomorphic mock driver to an automated guided vehicle increased the reported trust in the interaction of participants with the platform. A broader review of the human–robot interaction literature reports that some studies find anthropomorphism builds trust, while others find that it distract users from the primary task. For workplace deployment, the evidence suggests that appearance choices should be tested against safety and attention outcomes, not only user preference.

Open problems and research directions

Most open problems in physical AI trace back to one constraint: physical interaction produces far less training data than the text and image corpora that drove language models, and each unit of robot data costs time and money to collect. The same constraint shapes the research agenda on data efficiency, generalization, simulation fidelity and compute.

Data efficiency is the most immediate bottleneck. Robot data has to be gathered through labor-intensive processes, and one survey estimates that the Open X-Embodiment corpus contains roughly one two-hundred-thousandth as many tokens as large language model training sets. Unlike language and vision data, robot data cannot be scraped from the internet and must be collected by hand. Teleoperation, the main collection method, is costly, and the resulting trajectories often cover only a narrow set of environments and tasks. Researchers are therefore pursuing ways to extract more from each demonstration, such as using unstructured play data, augmenting images through in-painting, and learning from unlabeled human video (see data-augmentation and self-supervised-learning).

Data heterogeneity compounds the shortage. The Open X-Embodiment collection pools material from 21 institutions, yet most constituent datasets are tied to a single robot in a fixed environment, risking overfitting to robot–scene combinations. The distribution is also lopsided: over 85% of real trajectories come from just four robots. Because the "embodiment gap" makes data transfer difficult between different robot platforms due to variations in kinematics, sensors, and action spaces, cross-embodiment transfer remains a central goal. Evidence that such transfer helps exists: in one comparison, RT-2-X outperformed RT-2 by roughly threefold on a generalization test, which suggests that pooled data can extend the range of tasks even for a robot that already has abundant data. Generalization to unseen objects, environments and long action sequences remains weaker than in-distribution performance, and hierarchical designs that pair a planner with a controller, together with world-model-ai approaches, are often proposed as partial remedies.

The sim-to-real gap is a second major problem. Most simulators are built mainly around rigid-body dynamics, which limits how accurately they capture contact. To keep computation affordable, simulation engines trade off the fidelity of multi-contact modeling, soft-material deformation, friction variation and actuator stiffness. High-resolution tactile simulation is especially expensive, which forces a trade-off between simulation speed and accuracy, often resulting in brittle policies. Motor backlash, static and dynamic friction, and response delays add further mismatch (see friction). Domain randomization narrows the gap, but a policy may still meet materials or geometries outside the randomized range. There is also no single metric for the gap: it is typically judged relative to the natural variation between instances of the same physical system, and it depends on the task. Work on deformable objects such as cloth shows that small mismatches in friction or compliance can change force signals enough to break transfer.

Scaling is an open empirical question. A meta-analysis of 327 papers reported that the performance of robotic models improves with increased resources, following a power-law relationship, and that the improvement in robotic task performance scales notably faster than language tasks, although current results are often moderate-to-poor. Other analysts argue that each new data point in robotics produces a smaller performance gain than in language modelling, owing to heterogeneous data quality and the scarcity of action-labelled contact-rich interactions. A further view holds that the axes of robot scaling laws will likely not be the same axes as language model scaling laws. Whether neural-scaling-law behavior transfers to physical action therefore remains undecided.

Energy and compute costs form a final cluster of issues. Training large vision-language-action models requires substantial graphics-processing-unit capacity, and onboard inference must meet strict latency and power limits, which keeps pushing designers toward a split between edge and cloud-computing. Simulation of contact-rich scenes is itself computationally heavy, so the cost of generating training data can rival the cost of training the policy. Published accounts of the energy use per deployed robot or per training run remain sparse, which makes it hard to assess the environmental trade-offs of scaling embodied models.

References

1. What is Physical AI — and why it matters nowey.com
2. What is Physical AI? Definition and Exampleswherobots.com
3. What Is Physical AI? - Ciscocisco.com
4. What Is Physical AI?coreweave.com
5. What is Physical AI? Definition, Examples, and Usesencord.com
6. Physical AI: What It Is and What It Is Nottowardsdatascience.com
7. Physical AI vs. Embodied AI — Futures Gardengarden.johanneskleske.com
8. The Three A’s of Embodied AI: Autonomy, Accuracy, Accountabilityxavor.com
9. Embodied AI vs. Physical AI: From Digital Humans to Real Worldmimicproductions.com
10. Cosmos-Reason1: From Physical Common Sense To Embodied Reasoningarxiv.org
11. Embodied AI: Emerging Risks and Opportunities for Policy Actionarxiv.org
12. Norbert Wiener and the Birth of Cybernetics: The Science of Feedbackmedium.com
13. Cyberneticsbritannica.com
14. Principles and Practice of Deep Representation Learning: or a Mathematical Theory of Memoryarxiv.org
15. Shakey the Robot - SRI Internationalsri.com
16. Shakey - CHM Revolution - Computer History Museumcomputerhistory.org
17. Shakey: From Conception to History - Stanford AI Labai.stanford.edu
18. A BRIEF INTRODUCTION TO BEHAVIOR-BASED ROBOTICSbaibook.epfl.ch
19. Subsumption architectureen.wikipedia.org
20. PCT and Beyond: Towards a Computational Framework for `Intelligent' Communicative Systemsarxiv.org
21. Cybernetics: Or Control and Communication in the Animal and the Machineen.wikipedia.org
22. SRI’s AI Center celebrates 60 years of AI impact - SRIsri.com
23. 12 AI Milestones: 1. Shakey The Robotforbes.com
24. World Model for Robot Learning: A Comprehensive Surveyarxiv.org
25. World Model for Robot Learning: A Comprehensive Surveyarxiv.org
26. World-Action Models for Robot Learning and Control: A Surveyarxiv.org
27. PlayWorld: Learning Robot World Models from Autonomous Playarxiv.org
28. Toward Unified Robot Learning: Bridging Representation, Vision-Language-Action, and World Modelsarxiv.org
29. DyWA: Dynamics-adaptive World Action Model for Generalizable Non-prehensile Manipulationarxiv.org
30. Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directionsarxiv.org
31. DSSP: Diffusion State Space Policy with Full-History Encodingarxiv.org
32. World model (artificial intelligence)en.wikipedia.org
33. Tactile-Proprioceptive Sensor Fusion for Contact Wrench Estimation in Whole-Body Physical Human-Robot Interactionarxiv.org
34. (PDF) Proprioceptive Sensor Fusion for Quadruped Robot State Estimationresearchgate.net
35. Multi-Sensor Fusion for Quadruped Robot State Estimation using Invariant Filtering and Smoothingarxiv.org
36. Artificial Intelligence-Enabled Space Robot Operations: Technologies, Challenges and Prospectsarxiv.org
37. Kalman Filter: Historical Overview and Review of Its Use in Robotics 60 Years after Its Creation - Urrea - 2021 - Journal of Sensors - Wiley Online Libraryonlinelibrary.wiley.com
38. A Review of Radio Frequency Based Localization for Aerial and Ground Robots with 5G Future Perspectivesarxiv.org
39. Map as The Hidden Sensor: Fast Odometry-Based Global Localizationarxiv.org
40. Mobile Robot Localization Using Fuzzy Neural Network Based Extended Kalman Filterarxiv.org
41. (PDF) Kalman Filters and Robot Localizationresearchgate.net
42. Building Ears for Robots: Machine Hearing in the Age of Autonomyarxiv.org
43. Vision-Language-Action Models: Experimental Insights from a Real-World UR5 Platformarxiv.org
44. What is RT-2? Google DeepMind’s vision-language-action model for roboticsblog.google
45. RT-2（A Vision-Language-Action model released by Google DeepMind in July 2023.）_Baiduwikibaike.baidu.com
46. Google DeepMind Researchers Introduce RT-2: A Novel Vision-Language-Action (VLA) Model that Learns from both Web and Robotics Data and Turns it into Action - MarkTechPostmarktechpost.com
47. RT-2: New model translates vision and language into action — Google DeepMinddeepmind.google
48. Google RT-2learn.engineering.vips.edu
49. [2406.09246] OpenVLA: An Open-Source Vision-Language-Action Modelarxiv.org
50. A Survey on Vision-Language-Action Models: An Action Tokenization Perspectivearxiv.org
51. NVIDIA Announces Isaac GR00T N1 — the World’s First Open Humanoid Robot Foundation Model — and Simulation Frameworks to Speed Robot Developmentnvidianews.nvidia.com
52. NVIDIA Isaac GR00T N1: An Open Foundation Model for Humanoid Robotsresearch.nvidia.com
53. GR00T N1: An Open Foundation Model for Generalist Humanoid Robotsarxiv.org
54. GitHub - openvla/openvla: OpenVLA: An open-source vision-language-action model for robotic manipulation. · GitHubgithub.com
55. Vision Language Action Models (VLA) & Policies for Robotslearnopencv.com
56. RT-2: Vision-Language-Action Models for Generalizable Robotic Control: A Comprehensive Reviewmadison-proceedings.com
57. OpenVLA: An Open-Source Vision-Language-Action Modelarxiv.org
58. Embodied Operators and Benchmarking: Toward Reusable and Deployable Embodied Intelligence Systemsarxiv.org
59. Deep Imitation Learning for Humanoid Loco-manipulation through Human Teleoperationarxiv.org
60. ExoStart: Efficient learning for dexterous manipulation with sensorized exoskeleton demonstrationsarxiv.org
61. Interactive Imitation Learning for Dexterous Robotic Manipulation: Challenges and Perspectives -- A Surveyarxiv.org
62. AI Robotics: A Field Report on Imitation Learning with LeRobotml6.eu
63. Robot learning by demonstration - Scholarpediascholarpedia.org
64. Frontiersfrontiersin.org
65. Sim-to-Real Transfer in Robotic Manipulation Using Domain Randomization and Meta-Learningresearchgate.net
66. ADR: Train Hard, Transfer Smart. Learned domain randomization selects…medium.com
67. Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: a Surveyarxiv.org
68. Markerless Camera-to-Robot Pose Estimation via Self-supervised Sim-to-Real Transferarxiv.org
69. Open X-Embodiment: Robotic Learning Datasets and RT-X Modelsrobotics-transformer-x.github.io
70. Scaling up learning across many different robot types — Google DeepMinddeepmind.google
71. Open X-Embodiment: Robotic Learning Datasets and RT-X Modelsarxiv.org
72. Domain Randomization in Sim-to-Real Transferemergentmind.com
73. What Are Physics-Informed Neural Networks (PINNs)? - MATLAB & Simulinkmathworks.com
74. Physics-informed neural networks for physiological signal processing and modeling: a narrative review - PMCpmc.ncbi.nlm.nih.gov
75. Physics Informed Neural Networks in Engineeringneuralconcept.com
76. Physical vs. Data-Driven Modeling for Industrial Process Optimization — PatSnap Insightspatsnap.com
77. Learning dynamical systems from data: An introduction to physics-guided deep learningpnas.org
78. Hybrid physics-informed machine learning framework for calibration-free degradation prediction of lithium-ion batteriesnature.com
79. Least-Action-Guided Diffusion for Physical Extrapolationarxiv.org
80. Physics-informed deep learning for robust trajectory prediction in automated drivingnature.com
81. A Comprehensive Review of Generative Physical Artificial Intelligencearxiv.org
82. Generative Artificial Intelligence in Robotic Manipulation: A Surveyarxiv.org
83. What is Physical AI?nvidia.com
84. Generative Simulation for Policy Learningin Physical Human-Robot Interactionarxiv.org
85. Blue Sky Ideas: Towards Physics-Guided Foundation Modelsarxiv.org
86. Physics-Informed Neural Network - an overviewsciencedirect.com
87. Physics-Informed Neural Networks and Extensionsarxiv.org
88. Physics-constrained machine learning for scientific computing - Amazon Scienceamazon.science
89. Scale Synthetic Data and Physical AI Reasoning with NVIDIA Cosmos World Foundation Modelsdeveloper.nvidia.com
90. Humanoid handen.wikipedia.org
91. Boston Dynamics unveils new dexterous hands for Atlas humanoid robot - Robotics 24/7robotics247.com
92. Boston Dynamics drops pinkie on new humanoid handtherobotreport.com
93. Antagonistic Bowden-Cable Actuation of a Lightweight Robotic Hand: Toward Dexterous Manipulation for Payload Constrained Humanoidsarxiv.org
94. Humanoid Robots Move Toward Commercialization; Solid-State Batteries Set to Break Through Power Bottlenecks, Says TrendForcetrendforce.com
95. Batteries for Humanoid Robots: Global Market 2027-2040 - Advanced and Emerging Technology Market Researchfuturemarketsinc.com
96. Humanoid Robot Battery Selection: Battery Life, Capacity, and Future Trendscmbatteries.com
97. Solid-State Batteries for Humanoid Robots: The Next Power Revolution - HYXinbatteryhyxinbattery.com
98. Why Are Humanoid Robots Becoming the First Major Market for Solid-State Batteries? - HYXinbatteryhyxinbattery.com
99. How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-Perfarxiv.org
100. Efficient Vision-Language-Action Management and Serving for Robot Factoriesarxiv.org
101. Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation Systemarxiv.org
102. Microsoft tests when robots should think beyond the machineblackscarab.ai
103. Waymo’s skyrocketing ridership in one charttechcrunch.com
104. Waymo Expands Paid Robotaxi Rides to Denver, San Diego and Tampa - Bloombergbloomberg.com
105. Robotaxi Status October 2026 — Waymo, Tesla & Zooxthechargeport.com
106. What warehouse robots actually replace, and what they still cannot dowashingtontimes.com
107. Warehouse Picking Robots: A Practical Guidephysicalailab.io
108. Towards Active Robotic Vision in Agriculture: A Deep Learning Approach to Visual Servoing in Occluded and Unstructured Protected Cropping Environmentsarxiv.org
109. Advances in soft grasping in agriculturearxiv.org
110. Autonomous Robotic Pepper Harvesting: Imitation Learning in Unstructured Agricultural Environmentsarxiv.org
111. One of Warehouse Automation's Hardest Problems - Logistics Newslogisticsbusiness.com
112. Artificial Intelligence (AI) Robots Market Report 2025 - 2030 [260 Pages & 178 Tables]marketsandmarkets.com
113. Physical AI Statistics: Robotics And Automation Future (2026)sci-tech-today.com
114. Physical AI Market Size, Share, Trends & Growth Report 2026–2035kaisoresearch.com
115. Physical AI Market Report 2027-2040: Robots, Autonomous Vehicles & Smart Systemsfuturemarketsinc.com
116. Rise of Physical AI: Market Growth, Global Landscape, and Future Outlookcervicornconsulting.com
117. World models: who are the top startups today?signalsinbox.com
118. Yann LeCun's AMI Labs raises $1.03B to build world modelstechcrunch.com
119. The 16 Largest Global Startup Funding Rounds of March 2026alleywatch.com
120. Nvidia backs startup Reactor as buzz grows for world modelsfortune.com
121. Physical AI Market Set to Surpass $430 Billion by 2030, Driven by Nine Key Vertical Sectorsfinance.yahoo.com
122. 'Listing is a must': Chinese humanoid startups are rushing to launch IPOscnbc.com
123. China’s backflipping robots leap to blockbuster stock market debut in latest sign of AI boomnbcnews.com
124. Exclusive-Chinese robot maker AgiBot plans Hong Kong IPO next year, sources sayfinance.yahoo.com
125. Robotera plans IPO in Hong Kong as humanoid robot funding hits overdrivecryptobriefing.com
126. Humanoid robot maker EngineAI files for a Hong Kong IPOthenextweb.com
127. 23 New World Model startups emerge in China in 7 months; 2026 sees rapid growthkucoin.com
128. Agibot IPO: everything you need to knowcapital.com
129. Physical Human-Robot Interaction: A Critical Review of Safety Constraintsarxiv.org
130. Safety Standards for Humanoid and General-Purpose Robots: A Practical Guidesres.ai
131. SCENEREPLICA: Benchmarking Real-World Robot Manipulation by Creating Replicable Scenesarxiv.org
132. REALM: A Real-to-Sim Validated Benchmark for Generalization in Robotic Manipulationarxiv.org
133. RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policiesarxiv.org
134. ARC: Autonomous Robotics Compliance A Three-Layer Governance Architecture for Deployed Autonomous Systemsarxiv.org
135. Physical AI: embodied AI gives rise to new legal requirementscms.law
136. Trustworthy Agentic AI: Failure Modes, Mitigation Strategies, and a Lifecycle Framework for Autonomous LLM Systemsarxiv.org
137. One robot, two laws: Industry warns EU’s double regulation in the way of progress - EU Perspectiveseuperspectives.eu
138. EU Machinery Regulation: the safety baseline for AI-enabled and robotic machinerytransition-pathways.europa.eu
139. A Pathway Study For Future Humanoid Standardstherobotreport.com
140. Robots and Jobs: Evidence from US Labor Marketsideas.repec.org
141. Robots and jobs: Evidence from the UScepr.org
142. Robots vs. Workers: Evidence From a Meta‐Analysis - Guarascio - 2025 - Journal of Economic Surveys - Wiley Online Libraryonlinelibrary.wiley.com
143. Robots, Natives and Immigrants in US local labor markets - ScienceDirectsciencedirect.com
144. Warehouse automation hasn’t made workers safer—it’s just reshuffled the risk, say researchersinsights.bu.edu
145. WORKFORCE AUTOMATION Insights into Skills and Training Programs forgao.gov
146. Automation and the Importance of Reskilling Workers: A Case Study in the Automotive Industrypapers.academic-conferences.org
147. Artificial Intelligence as an Economic, Environmental, Geopolitical, and Social Transformationarxiv.org
148. Impact of Anthropomorphic Robot Design on Trust and Attention in Industrial Human-Robot Interactiondl.acm.org
149. The Effect of Anthropomorphism on Trust in an Industrial Human-Robot Interactionarxiv.org
150. Impact of Anthropomorphic Robot Design on Trust and Attention in Industrial Human-Robot Interactionresearchgate.net
151. Mapping trust and cultural dimensions in Human-Robot Interaction: A scoping review approach - ScienceDirectsciencedirect.com
152. OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learningarxiv.org
153. Open X-Embodiment: Robotic Learning Datasets and RT-X Modelsalphaxiv.org
154. Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulationarxiv.org
155. Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulationarxiv.org
156. Neural Scaling Laws in Roboticsarxiv.org
157. The economics of Physical AI: Why data quality beats scaleblog.robotiq.com
158. What’s Missing for Robot Foundation Models? - by Ted Xiaoagentic.substack.com
159. Foundation Models in Robotics: Applications, Challenges, and the Futurearxiv.org
160. Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectoriesarxiv.org
161. Mind to Hand: Purposeful Robotic Control via Embodied Reasoningarxiv.org
162. Bridging the Sim-to-Real Gap with Dynamic Compliance Tuning for Industrial Insertionarxiv.org
163. (PDF) Benchmarking the Sim-to-Real Gap in Cloth Manipulationresearchgate.net
164. Examining the simulation-to-reality gap of a wheel loader digging in deformable terrainlink.springer.com
165. Open X-Embodiment: Robotic Learning Datasets and RT-X Modelsarxiv.org
166. (PDF) Open X-Embodiment: Robotic Learning Datasets and RT-X Modelsresearchgate.net