{ "id": "humanoid-robot", "title": "Humanoid robot", "lang": "en", "source_lang": "en", "summary": "A robot with a humanlike body configuration, developed for physical tasks, human interaction, and research into movement and cognition.", "category": "technology", "subcategory": "engineering", "type": "concept", "aliases": [ "Humanoid robotics", "Human-shaped robot" ], "generated_at": "2026-10-14", "model": "pro/gpt-6.1-sol", "url": "https://aiwiki.page/en/humanoid-robot/", "translations": { "zh-Hans": { "url": "https://aiwiki.page/zh/humanoid-robot/", "json": "https://aiwiki.page/zh/humanoid-robot.json" } }, "sections": [ { "heading": "", "text": "A humanoid robot is a robot whose body resembles the human body in its overall configuration or in substantial parts, such as a torso, arms, hands, and head. Humanoid designs include full-body bipedal machines and upper-body systems mounted on fixed supports or other mobile platforms. Their humanlike structure can enable them to handle tools and workspaces designed for people; it does not necessarily imply lifelike appearance or autonomous operation. NASA’s Robonaut, for example, combines anthropomorphic arms and hands with different support and mobility arrangements. (nasa.gov)\n\nHumanoid robotics brings together mechanical design, sensing, control theory, and artificial intelligence. Its central engineering problems include maintaining balance, coordinating movement and manipulation, and acting reliably under changing physical conditions. Research encompasses both model-based methods and learned behaviours, including reinforcement learning for locomotion. Humanoids also provide experimental platforms for studying embodied cognition and human–robot interaction. (arxiv.org)\n\nThe field’s development includes Waseda University’s WABOT-1 in 1973, Honda’s introduction of ASIMO in 2000, and the disaster-response tasks of the 2015 DARPA Robotics Challenge. Applications range from scientific experimentation and space robotics to documented deployments in logistics. Understanding these machines requires distinguishing demonstrated capabilities from proposed uses, and assessing safety within the particular environment and task rather than from humanlike form alone. (humanoid.waseda.ac.jp)" }, { "heading": "Definition, boundaries, and design rationale", "text": "A humanoid robot is a robot whose body arrangement broadly resembles that of a human, particularly through a torso, paired arms, and a head, sometimes combined with two legs. The resemblance concerns both morphology—the placement and articulation of body parts—and functional organization, such as using arms and hands to reach and manipulate objects. It does not require realistic skin, human proportions, or replication of every anatomical joint. NASA’s Robonaut illustrates this distinction: its head contains vision equipment, while its computers occupy its torso; its humanlike organization serves manipulation rather than anatomical imitation. (nasa.gov)\n\nThe category includes several configurations. A full-body humanoid combines an anthropomorphic upper body with two articulated legs; an upper-body humanoid may instead be fixed to a pedestal or another support. A humanoid mobile manipulator combines a humanlike upper body with a mobile base, which need not be bipedal. These descriptions overlap rather than define mutually exclusive classes. Robonaut 2 was initially a torso-only system mounted on a stanchion, and NASA added a mobility platform with two climbing legs in 2014. Earlier Robonaut work also explored mounting the upper body on wheels. Thus, the absence of walking legs does not by itself exclude a platform from humanoid robotics, while a mobile base carrying an ordinary industrial arm is not necessarily humanoid. (nasa.gov)\n\nSeveral neighboring categories describe different characteristics. An android emphasizes lifelike human appearance and, in some research, humanlike movement: Hiroshi Ishiguro’s android-science work explicitly investigates machines whose appearance and motion closely resemble those of people. A social robot, by contrast, is characterized by its interaction with people rather than its body plan; PARO, a seal-shaped robot, provides a non-humanoid example. A powered exoskeleton is worn by a person and mechanically assists, resists, or otherwise modifies the wearer’s movement, rather than constituting a separate human-shaped manipulator. NASA’s X1, developed using technology derived from Robonaut 2, exemplifies this wearable arrangement. (tandfonline.com)\n\nA principal engineering rationale for humanoid morphology is compatibility with tools and interfaces already designed around human bodies. Humanlike arm placement, reach, and grasp geometry can allow a robot to approach equipment without requiring every handle or tool to be redesigned. NASA and General Motors made human-tool operation a design requirement for Robonaut 2. Each hand has 12 degrees of freedom, distributed among five fingers; its developers placed components in the forearm to leave more space for sensing in the fingers and palm. The purpose was to obtain useful grasps and controlled tool actuation, not merely a recognizable silhouette. This rationale also permits selective departures from human anatomy: a humanlike upper body can retain tool compatibility while using wheels or specialized climbing limbs for mobility. (nasa.gov)\n\nMorphology, autonomy, and intended purpose are separate characteristics. “Humanoid” specifies a bodily configuration, not a particular level of artificial intelligence or a guarantee of general-purpose competence. NASA describes Robonaut as supporting both software-controlled task execution and remote operation. In the former mode, an operator assigns a task and software determines the corresponding actions; in the latter, an operator views camera imagery through a headset and uses controls to direct movement. The same humanlike body can therefore serve as an autonomous task-execution platform or as a remotely operated extension of a human operator, without changing its morphological classification. (nasa.gov)" }, { "heading": "Historical development and landmark platforms", "text": "Humanlike automata preceded electronic robotics by several centuries. Among the surviving examples are the Writer, Draughtsman, and Musician, made between 1768 and 1774 by Pierre Jaquet-Droz, his son Henri-Louis, and Jean-Frédéric Leschot. Clockwork supplied their power, while cams encoded their movements; the Writer could be configured to write a text of forty characters. These machines reproduced elaborate human performances, but their mechanically encoded sequences differed from the sensor-based perception and feedback control subsequently integrated into humanoid robots. (mahn.ch)\n\nAt Waseda University in Japan, four laboratories formed a bioengineering group and began the WABOT project in 1970. Completed in 1973, WABOT-1 combined limb control, vision, and conversation systems in a full-scale anthropomorphic machine. It walked on two legs, communicated in Japanese, and measured the distance and direction of objects through its sensory systems. Hands equipped with tactile sensors enabled it to grasp and transport objects. Its historical significance lay in bringing locomotion, manipulation, and communication together within one experimental platform, rather than demonstrating human-equivalent general intelligence. (humanoid.waseda.ac.jp)\n\nWABOT-2, developed from 1980 to 1984, pursued a deliberately specialized task: playing a keyboard instrument. Its visual system read ordinary musical notation, and it could perform pieces of moderate difficulty on an electronic organ. It also conversed with people and accompanied a person singing. Waseda explicitly distinguished this “specialist robot” from the more broadly conceived WABOT-1. Musical performance provided a concrete task linking sensory input to coordinated physical action, rather than merely replaying a mechanically fixed sequence. (humanoid.waseda.ac.jp)\n\nHonda began humanoid-robot research in 1986, initially concentrating on bipedal walking before integrating legs, arms, onboard computation, and power. Its P2, announced in December 1996, stood 1.82 metres tall and weighed 210 kilograms. Computers, motor drives, batteries, and wireless equipment were carried inside its torso, allowing untethered walking, stair climbing, and cart pushing. P3, completed in September 1997, reduced height to 1.60 metres and mass to 130 kilograms through changes to materials and control architecture. These successive platforms established a development path from experimental walking mechanisms toward smaller, self-contained humanoids. (global.honda)\n\nHonda unveiled ASIMO in Tokyo on November 20, 2000. The original version stood 1.20 metres tall, weighed 43 kilograms, had 26 degrees of freedom, and walked at up to 1.6 kilometres per hour. Its predicted-movement control shifted the body’s centre of gravity in anticipation of subsequent motion, supporting smooth changes of direction. Walking commands and gestures could be issued through a workstation or portable controller; the demonstration therefore should not be equated with unrestricted autonomous behaviour. (global.honda)\n\nBoston Dynamics’ Atlas made its public debut on July 11, 2013, as a platform for the DARPA Robotics Challenge. DARPA supplied seven platforms to qualifying teams, which developed their own software and operator interfaces. Initially tethered for power and communications, Atlas was upgraded for onboard power and wireless operation before the finals in Pomona, California, on June 5–6, 2015. Human-supervised tasks included opening doors, operating valves, using tools, and negotiating difficult terrain under constrained communications. This hydraulic research lineage must be distinguished from the fully electric Atlas introduced in April 2024, when Boston Dynamics retired its hydraulic predecessor and announced a redesigned platform intended for industrial applications. (darpa.mil)" }, { "heading": "Mechanical structure, actuation, and power", "text": "A humanoid’s skeleton is commonly modelled as a branching assembly of rigid bodies connected by joints: the torso links two arms, two legs, and a head, sometimes through additional waist joints. Revolute joints provide rotation about an axis; combinations of axes approximate shoulder, hip, wrist, and ankle motion. The number of independent joint coordinates defines the mechanism’s internal degrees of freedom, rather than simply the number of motors. Redundancy is relative to a task: a seven-joint arm generally has an extra degree of freedom when controlling a hand’s six-dimensional position and orientation. This can permit elbow movement while the hand remains fixed, helping avoid obstacles or joint limits, but does not make every hand pose reachable. Link lengths, joint ranges, and self-collision still constrain motion. NASA’s Valkyrie, for example, has seven joints per arm and 44 degrees of freedom overall. (hades.mech.northwestern.edu)\n\nElectric actuation combines an electric motor with a transmission, commonly reduction gearing, to exchange speed for joint torque. Ideally, a reduction ratio \\(N\\) multiplies motor torque by \\(N\\) while dividing output speed by \\(N\\); actual transmissions introduce friction and losses. High reductions allow smaller motors but increase the motor’s reflected inertia at the joint in proportion to \\(N^2\\), making external forces harder to transmit backward through the mechanism. Low-ratio drives reduce this effect but require motors capable of greater torque. Hydraulic actuators offer another route to large joint forces without large gear reductions, using pressurized fluid and controlled valves. Boston Dynamics’ successive Atlas architectures illustrate the distinction: the company retired its hydraulic platform and introduced an all-electric replacement in April 2024, describing the replacement as having a broader range of motion. (fab.cba.mit.edu)\n\nJoint stiffness need not be identical to structural rigidity. A series elastic actuator deliberately places an elastic element between the drive and load; measured spring deflection provides an estimate of transmitted torque and the spring can reduce shock loads on gearing. MIT’s Cog humanoid used this arrangement in its arms, while Valkyrie incorporates both rotary and linear series elastic actuators. Torque control regulates joint effort through feedback, rather than prescribing position alone. Low-friction drives can estimate effort from motor current; highly geared systems may instead use output-side torque sensing. Compliance consequently depends on mechanical elasticity, transmission characteristics, and the control system together. (groups.csail.mit.edu)\n\nPayload capacity depends on posture and acceleration, not just actuator ratings. For a stationary load, its gravitational contribution to joint torque is \\(mgd\\), where \\(d\\) is the perpendicular distance from the joint axis to the load’s line of action; extending an arm therefore increases the torque required to hold the same object. During movement, link inertia and acceleration add further demands. Stronger motors, larger bearings, and reinforced transmissions increase load capacity but also add mass that other joints must support and accelerate. Gear teeth and other transmission components must additionally tolerate impact loading: large reduction ratios increase reflected inertia and can amplify shock forces on the gearing. (manipulation.csail.mit.edu)\n\nOnboard power introduces separate energy, electrical-power, and thermal limits. Battery energy determines approximate endurance through \\(t \\approx E_{\\mathrm{usable}}/P_{\\mathrm{average}}\\); it does not establish whether the electrical system can supply a brief peak load. NASA lists Valkyrie’s battery at 1.8 kWh and describes approximately one hour of battery operation, while also permitting external power. A NASA battery presentation separately specifies continuous discharge-current limits for its two power buses, illustrating that energy capacity and deliverable power are distinct specifications. Motor heating creates another constraint: resistive loss scales as \\(I^2R\\), so producing sustained torque can generate substantial heat even when a joint is stationary and its mechanical output power is zero. Cooling and allowable temperatures therefore constrain sustained effort independently of remaining battery charge. (nasa.gov)" }, { "heading": "Sensing, perception, and state estimation", "text": "Humanoid robots combine internal, or proprioceptive, measurements with external, or exteroceptive, observations. Joint encoders measure angular position; joint velocities can be derived from successive readings and filtered. An inertial measurement unit (IMU), usually attached to the torso or pelvis, measures angular velocity and specific force rather than directly reporting body position. Foot force–torque sensors provide evidence of ground contact and loading. These signals describe the robot’s motion and interaction with its supports. External sensors—including cameras, stereo vision, depth cameras, and laser ranging devices—observe surrounding surfaces and objects. Neither category alone provides a complete description: joint angles do not establish the torso’s location in the room, while an image does not directly establish which foot is bearing weight. (journals.sagepub.com)\n\nState estimation reconstructs quantities such as body orientation, position, velocity, sensor biases, and contact locations from these incomplete measurements. Through sensor fusion, an estimator propagates motion using IMU readings and corrects that prediction using other observations. If a supporting foot is stationary, measured joint angles and the robot’s kinematic model constrain the body’s movement relative to that foot. Contact detection determines when this constraint is applicable; slipping, rolling, and impact weaken it. Extended Kalman filters and factor-graph methods represent measurement uncertainty rather than treating every input as exact. Ross Hartley and colleagues’ 2018 contact-aided invariant filter demonstrated this approach on a Cassie biped. Their analysis also showed that IMU and contact measurements alone cannot determine absolute position or heading about the gravity axis; these quantities require an external reference to prevent unbounded drift. (arxiv.org)\n\nDifferent uses of the estimated state impose different timing requirements. Balance feedback needs a smooth, low-latency estimate, whereas navigation needs continuing alignment with an environmental map. In 2014, Maurice Fallon, Matthew Antone, Nicholas Roy, and Seth Teller described an Atlas architecture separating these functions: a control estimate operated at 333 Hz, while a localization estimate incorporated laser-based map corrections at 40 Hz. This separation prevented discontinuous localization corrections from directly disturbing walking feedback. The study also quantified the benefit of fusion: with the robot standing still, pelvis-velocity standard deviation fell from 7.6 cm/s for raw kinematic measurements to 2.3 cm/s after joint-level filtering and 1.4 cm/s after extended Kalman filtering. These were measurements of stationary noise, not general walking-position accuracy. (groups.csail.mit.edu)\n\nPerception converts external observations into representations usable for physical action. Three-dimensional point clouds can be organized into elevation maps for assessing ground height or occupancy maps for identifying obstacles and free space. Object localization requires more than detecting an object category: reaching toward a valve or tool requires its position, orientation, and relevant interaction geometry in coordinates shared with the robot. Team MIT’s perception work combined automatic and operator-guided fitting of object models to sensor observations, including valve models. Terrain reconstruction likewise has to distinguish observed surfaces from inferred ones. A 2021 study by Maximilian Stölzle and colleagues reconstructed missing elevation-map regions using self-supervised learning; such completion estimates hidden geometry rather than directly measuring it. (groups.csail.mit.edu)\n\nCalibration, occlusion, noise, and latency introduce different uncertainties. Calibration establishes the relationships between sensor measurements and coordinate frames; incorrect relationships misalign reconstructed surfaces with the body. Occlusion leaves unobserved regions, including terrain hidden by foreground objects. Sensor noise, mechanical flexibility, and backlash contaminate motion estimates, while processing delays mean that observations describe an earlier robot configuration. Moving scanners also require motion compensation during acquisition. Fallon and colleagues explicitly accounted for robot motion and scanner rotation when projecting laser returns, and increased measurement covariance around foot-contact transitions. Their system illustrates why closed-loop feedback depends not simply on collecting more sensor data, but on associating observations with the correct time, geometry, contact condition, and uncertainty. (arxiv.org)" }, { "heading": "Balance, locomotion, and whole-body control", "text": "A biped is a floating-base mechanical system: its trunk has six unactuated degrees of freedom—three translations and three rotations—in addition to its articulated joints. Joint motors cannot directly prescribe the body’s motion through space; they must exploit external contact forces to support and accelerate it. This makes locomotion underactuated. The governing constraints change between double support, single support, and flight: a stationary stance foot constrains motion relative to the ground, whereas a swing foot does not. Ordinary ground contacts can push but not pull, and tangential forces are limited by friction. Consequently, a feasible joint trajectory must also produce physically feasible contact forces. (arxiv.org)\n\nOn level ground, the support polygon is the convex hull of the contacting foot areas. Keeping the vertical projection of the centre of mass inside this polygon describes quasi-static balance, but walking also depends on velocity and acceleration. The zero-moment point (ZMP) is the point on the support plane where the resultant moment has no horizontal component. Under the linear inverted-pendulum approximation—constant centre-of-mass height \\(h\\), level support, and negligible change in centroidal angular momentum—the horizontal dynamics satisfy\n\n\\[\\ddot{\\mathbf{x}}=\\omega^2(\\mathbf{x}-\\mathbf{p}),\\qquad \\omega=\\sqrt{g/h},\\]\n\nwhere \\(\\mathbf{x}\\) is horizontal centre-of-mass position, \\(\\mathbf{p}\\) is the ZMP, and \\(g\\) is gravitational acceleration. ZMP-based planning keeps \\(\\mathbf{p}\\) within the support polygon while generating compatible body motion. Shuuji Kajita and colleagues’ 2003 preview controller used future ZMP references to anticipate upcoming steps; its spiral-staircase example was a simulation, not a hardware trial. (underactuated.mit.edu)\n\nThe capture point incorporates motion as well as position. For the same simplified model, it is \\(\\boldsymbol{\\xi}=\\mathbf{x}+\\dot{\\mathbf{x}}/\\omega\\): placing and maintaining the support point there allows the centre of mass to approach rest. Jerry Pratt and colleagues’ 2006 formulation used capture points and capture regions to reason about when and where to step after a disturbance. A reachable capture location outside the current support area suggests stepping rather than relying only on ankle action. Their flywheel extension represented changes in angular momentum produced by trunk and limb motion, broadening the available recovery strategies. These predictions remain conditional on foot reach, timing, and actuator limits. (cs.cmu.edu)\n\nFootstep planning selects landing locations, orientations, and timings subject to terrain geometry and leg reach; swing trajectories must clear obstacles. Model predictive control repeatedly optimizes a finite future sequence of body motion and contact forces, applies its initial command, and replans using updated feedback. Whole-body control then coordinates trunk orientation, swing-foot motion, balance, and other joint tasks through constrained optimization. Higher-priority contact and balance requirements can override lower-priority posture targets. Contact transitions also require load transfer: Donghyun Kim and colleagues’ 2018 controller gradually changed allowable reaction forces to avoid abrupt commands. Its physical stepping experiment used a passive-ankle biped whose torso was lightly held to constrain horizontal movement, distinct from its full-humanoid simulation tests. (underactuated.mit.edu)\n\nUneven terrain introduces changing contact heights, surface inclinations, and uncertain traction. Physical evidence must therefore identify the tested conditions rather than equate simulated robustness with demonstrated capability. In 2024, Ilija Radosavovic and colleagues reported real-world Digit locomotion on surfaces including sidewalks and grass, together with recovery from disturbances applied using a yoga ball, a stick, and backward pulling. These were hardware experiments with a simulation-trained controller; the described manual disturbances do not by themselves establish a standardized maximum recoverable push. (hybrid-robotics.berkeley.edu)\n\nGetting up after a fall requires a different contact sequence from periodic walking, potentially using hands and other body parts before transferring weight onto the feet. Xialin He and colleagues’ 2025 HumanUP study tested a physical Unitree G1 recovering from face-up and face-down configurations across six terrain types, including slippery surfaces and slopes. Its training first discovered recovery motions and then refined them to satisfy deployment constraints such as smoothness and torque limits. The demonstrated scope was recovery from the tested lying configurations, rather than unrestricted recovery from every possible fall posture. (arxiv.org)" }, { "heading": "Hands, manipulation, and tool use", "text": "A humanoid’s end effector need not reproduce a human hand. A task-specific gripper can constrain an object between a small number of jaws, whereas an articulated robotic hand provides additional contact locations and joint motions for varied grasps and object repositioning. Independently actuated fingers increase the number of variables that must be coordinated; underactuated designs instead use mechanical coupling and compliance to let fingers conform to objects with fewer actuators. This is not a strict division between simple grasping and dexterity: Yale’s OpenHand Model Q uses four actuators and can perform precision manipulation and finger-gaiting while retaining adaptive power grasps. (eng.yale.edu)\n\nGrasp selection determines the hand’s approach, finger configuration, and intended contact locations. A secure grasp must also leave the object usable: Robonaut 2’s hand, for example, was designed to grasp tools while retaining the finger motions needed to actuate them. Visual estimates can guide the approach, but uncertainty in object shape makes predicted contacts imperfect. In a 2022 study, Planning Visual-Tactile Precision Grasps via Complementary Use of Vision and Touch, a planner selected fingertip contacts on an estimated surface, then used tactile feedback to correct the motion when actual contact differed from the prediction. Sensor coverage mattered because planned contacts had to reach instrumented fingertip areas. (ntrs.nasa.gov)\n\nForce regulation closes the loop after contact. Insufficient grip permits sliding; increasing grip can arrest it, but excessive squeezing is undesirable, especially for deformable objects. Jasper W. James and Nathan F. Lepora’s 2020 experiments used a three-fingered Tactile Model O hand with optical tactile sensors to detect slip and tighten its grasp. The system also lifted objects of initially unknown mass by establishing light contact, attempting a lift, and increasing grip when slip occurred. Across eleven test objects, two response strategies achieved grasp-recovery success rates of 74% and 73%; the experiments were restricted to objects the hand could already acquire in an antipodal pinch. (arxiv.org)\n\nIn-hand manipulation changes an object’s position or orientation relative to the palm rather than merely moving the entire arm. Fingers may roll or slide over the surface, or alternately release and replace contacts in finger-gaiting. These motions must preserve support while avoiding joint limits and incompatible contact constraints. Mechanical design can simplify the problem: compliant and underactuated fingers can relieve overconstraint, while Yale’s variable-friction gripper changes finger-surface friction to permit selected sliding motions. More independently controlled joints therefore provide possibilities, not an automatic guarantee of more reliable manipulation. (eng.yale.edu)\n\nTwo-arm manipulation adds coordination between the hands and the rest of the body. DLR’s Justin upper-body system supports control at individual end-effector and shared-object levels, allowing both arms to participate in one manipulation task. On a standing or stepping humanoid, interaction forces also affect balance; arms, torso, and support contacts cannot be treated independently. Masaki Murooka and colleagues’ 2025 TACT experiments demonstrated whole-body contact manipulation on the life-size RHP7 Kaleido while maintaining balance and walking. Environmental bracing provides another arrangement: one hand supports the body against the surroundings while the other applies task force, although the supporting hand is then unavailable for grasping the workpiece. (rmc.dlr.de)\n\nRobonaut 2 illustrates human-tool compatibility through specific mechanical choices. NASA’s 2011 specifications give each hand 12 controllable degrees of freedom: four in the thumb, three each in the index and middle fingers, and one each in the ring and little fingers. Its hand-design paper describes forearm-mounted actuators, tendon transmission, coupled finger joints, and force sensing, and illustrates operation of devices including a hand drill. Humanlike dimensions and grasp configurations make existing handles and controls accessible; coupled joints and task-specific demonstrations do not establish the independent finger control or general-purpose competence of a human hand. (nasa.gov)" }, { "heading": "Learning, task planning, and autonomy", "text": "A motion-control policy maps observations and goals to actions, such as target joint positions or torques. Its goal may be a commanded walking velocity or a reference body pose; it need not specify why the robot should move or what task should follow. Higher-level task planning instead selects and sequences skills, such as approaching a table, grasping an object, and placing it elsewhere. Autonomous decision-making additionally requires environmental feedback to determine whether actions succeeded and whether to continue or revise the plan. These functions are separable: a learned controller can stabilize a humanoid while a human supplies its movement goals through teleoperation. (arxiv.org)\n\nImitation learning uses demonstrations to acquire behavior, either by learning observation–action mappings or by training controllers to reproduce reference motions. Human demonstrations cannot simply be copied joint by joint: robots differ in limb proportions, joint arrangements, motion limits, and available torque. Motion retargeting therefore converts human trajectories into robot-compatible targets through geometric fitting and optimization. In the 2024 Human to Humanoid (H2O) system, Tairan He and colleagues fitted a human body model to the robot using 12 corresponding joints, then minimized positional differences to retarget motion. A simulated controller with access to privileged state information subsequently identified sequences it could not track, which were removed before training the deployable controller. The resulting system followed human movements captured by an RGB camera; its learned motion tracking did not replace the human operator’s task decisions. (arxiv.org)\n\nReinforcement learning trains a policy through interaction to maximize expected cumulative reward. For humanoids, the reward can combine task progress or reference-motion tracking with penalties for excessive effort, abrupt actions, or undesirable contacts. The chosen terms and their relative weights define the behavior being optimized, rather than independently establishing that it is useful or reliable. Imitation and reinforcement learning can also be combined: demonstrations supply movement targets while simulated interaction teaches the controller how to realize them under physical constraints. In their April 2024 study of Digit, Ilija Radosavovic and colleagues used proximal policy optimization and a walking reward informed by biomechanics, without a precomputed gait library. (arxiv.org)\n\nSimulation permits many training environments to run in parallel without repeatedly exposing hardware to falls. Transfer to a physical robot nevertheless depends on discrepancies in dynamics, contacts, sensing, and timing. Domain randomization varies simulated conditions so that a policy must succeed across a range rather than exploit one exact model. Radosavovic and colleagues randomized robot dynamics, control parameters, terrain, observation noise, and delays. Their training pipeline also used a privileged-state teacher to guide a student policy operating on observation histories, followed by testing in the manufacturer’s higher-fidelity simulator before hardware deployment. H2O likewise used randomization for simulation-to-real transfer, while reporting that excessive randomization could impede motion learning. (hybrid-robotics.berkeley.edu)\n\nLanguage-conditioned interfaces introduce another distinction between selecting an appropriate action and executing it. SayCan, introduced in 2022 on a non-humanoid mobile manipulator, combined a large language model’s assessment of which skill would advance an instruction with learned estimates of that skill’s feasibility in the current state. Its PaLM-based version reported 84% planning success but 74% execution success, illustrating that a suitable sequence of skill descriptions is not equivalent to successful physical performance. The authors also identified limitations in feedback after skill failures or environmental changes. (say-can.github.io)\n\nLearned perception–action systems can instead connect visual observations, language instructions, and robot state directly to action generation. The 2025 GR00T N1 research model used a vision–language module coupled to a diffusion-transformer action module and trained on mixtures of real-robot trajectories, human video, and synthetic data. Such architectures investigate how multimodal learning can support instruction-conditioned manipulation across embodiments. Their task-specific physical and simulated evaluations must be distinguished from unrestricted autonomy: interpreting an instruction, producing an action trajectory, detecting failure, and recovering over a long sequence remain separate capabilities requiring separate evidence. (arxiv.org)" }, { "heading": "Human–robot interaction and embodied cognition", "text": "Human–robot interaction investigates how people interpret and coordinate with robots through gaze, gesture, posture, and speech. These channels convey complementary information: eye and head orientation can indicate an object of attention, pointing can identify a referent, and body movements can accompany spoken instructions. Their timing matters as much as their presence. Researchers synchronize recordings of gaze, movement, and speech to examine how signals precede an action or coincide with a conversational turn. Turn-taking is therefore studied not simply as alternating speech, but as coordination involving pauses, changes of gaze, and indications that a speaker intends to continue or yield. A 2021 study, Speech Driven Gaze in a Face-to-Face Interaction, combined eye tracking with speech and dialogue-act annotations to model gaze behavior associated with speaking, thinking, and turn management. (frontiersin.org)\n\nControlled robot experiments can isolate signals that are difficult to suppress in human partners. In the 2012 study I Reach Faster When I See You Look, the iCub robot verbally identified target cubes under three conditions: coordinated eye–head gaze, head movement with its eyes hidden by sunglasses, and a stationary head with no eye movement. Five participants completed ten rounds per condition. With visible eye–head cues or head cues alone, participants began reaching before the verbal target description ended; without these cues, reaching began afterward. Gaze affected movement onset rather than the subsequent interval from hand lift to cube contact. The result demonstrates anticipatory use of a robot’s directional signals in this particular cooperative task, rather than a general understanding of the robot’s intentions. (frontiersin.org)\n\nThe child-sized iCub also provides an experimental platform for embodied cognition: the investigation of cognition as developing through bodily action and interaction with the environment. The original RobotCub specification described a 94-centimetre robot with 53 degrees of freedom, articulated eyes and head, dexterous hands, and visual, vestibular, auditory, and haptic sensing. Its developmental research programme proceeds from coordinating gaze, head orientation, and reaching toward manipulation, imitation, and communication. For example, manipulating an object provides sensorimotor evidence about what actions that object permits, rather than requiring its properties to be supplied solely as symbolic descriptions. RobotCub’s openly distributed platform and open-source software support the integration and comparison of experimental perception, action, and cognitive architectures; the project concluded in January 2010. (robotcub.org)\n\nThe uncanny-valley hypothesis, proposed by Masahiro Mori in 1970, concerns a different aspect of interaction: whether affinity decreases when an artificial figure approaches, but does not fully achieve, human likeness. It is an empirical hypothesis, not a rule that every increasingly realistic robot necessarily becomes disturbing. Experiments distinguish perceived human likeness from eeriness, attractiveness, and acceptance. Matthieu Destephe and colleagues’ 2015 study analyzed responses from 69 French and Japanese participants, measuring these dimensions separately and asking about the acceptability of possible robot occupations. In that experiment, attractiveness and perceived humanness predicted occupational acceptance, whereas reported eeriness did not—showing that discomfort and rejection were not interchangeable outcomes. (frontiersin.org)\n\nAppearance judgments also differ methodologically from observations of movement or repeated encounters. In a 2015 live-interaction study, Jakub Złotowski and colleagues compared the humanlike Geminoid HI-2 with the more machine-like Robovie R2 across three interactions, while varying friendly or unfriendly behavior. Geminoid received higher explicit eeriness ratings, and positive behavior increased likeability. The overall analysis detected an effect of interaction number on explicit eeriness, although the corrected first-to-third comparison was not statistically significant; an implicit association measure showed no significant effects. These distinctions matter because questionnaire responses, implicit associations, and behavior need not change together. (frontiersin.org)\n\nExperimental design consequently determines the scope of conclusions. Ratings after viewing a stimulus do not establish willingness to collaborate over extended periods, and comparisons between two robots cannot independently separate every difference in appearance, voice, and movement. Participant recruitment, cultural background, task, exposure duration, and measurement reliability further constrain generalization. The French–Japanese study used unequal national samples, while the repeated-interaction study deliberately introduced mismatches between Geminoid’s appearance, speech, and movement. Their findings describe those participants and experimental configurations, rather than universal acceptance thresholds for humanoid robots. (frontiersin.org)" }, { "heading": "Applications and deployment contexts", "text": "In scientific use, humanoid robots serve as experimental instruments rather than primarily as labour-saving equipment. The European Commission-funded RobotCub project, completed in January 2010, developed iCub as an openly distributed platform for studying cognition, with shared hardware documentation and open-source software. Its articulated body and visual, auditory, vestibular, and touch sensing allowed researchers to investigate learning through physical interaction, including object manipulation and imitation. The relevant output was experimental evidence and reusable research infrastructure, not warehouse throughput or manufactured goods. RobotCub also used the platform for advanced education: its summer schools offered ten consecutive days of hands-on programming, integrating perception and control modules for the head, hands, arms, legs, and whole body. (robotcub.org)\n\nSpace-maintenance research placed a humanoid in a working spacecraft while retaining an experimental mission. Robonaut 2, developed jointly by NASA and General Motors, reached the International Space Station aboard Discovery in February 2011. Initially a torso mounted on a support post, it demonstrated actions such as pressing buttons, operating switches, and turning knobs in microgravity. A mobility upgrade added two climbing limbs in 2014; NASA reported that they were installed in August. Each limb had seven joints and a gripping end effector designed to engage station handrails and sockets. Unlike terrestrial walking, this arrangement was intended to let the robot reposition itself while keeping at least one limb anchored, leaving both hands available for work. These were technology demonstrations toward astronaut assistance, not evidence of routine autonomous station maintenance; NASA explicitly stated that exterior work would require further upper-body upgrades. (nasa.gov)\n\nDisaster-response applications were investigated through the DARPA Robotics Challenge, whose finals took place in Pomona, California, on June 5–6, 2015. The event tested human-supervised robots in a simulated disaster environment, rather than deploying them to an active emergency. Within a one-hour run, machines were asked to drive a vehicle, get out, pass through a door, operate a valve, cut an opening with a power tool, complete a surprise manipulation task, negotiate obstructed terrain, and climb stairs. Wireless communications between robots and remote operators were deliberately degraded, requiring onboard execution to continue through interruptions rather than depending on uninterrupted remote control. Program manager Gill Pratt identified this combination of human supervision and partial autonomy as a central challenge. The tasks therefore investigated access to human-built infrastructure under hazardous conditions, without establishing readiness for routine emergency service. (defense.gov)\n\nCommercial logistics provides a different deployment context: a narrowly specified physical task integrated into an existing production workflow. On June 27, 2024, GXO and Agility Robotics announced a multiyear Robots-as-a-Service agreement following a proof-of-concept pilot in late 2023. The deployment concerned Digit robots at a GXO-operated SPANX facility in Flowery Branch, Georgia. The announced workflow involved taking totes from other mobile robots and placing them on conveyors, coordinated through Agility’s Arc software. GXO’s announcement described the arrangement as an industry-first formal commercial humanoid deployment; that characterization was a company claim, distinct from the documented agreement and task description. (investors.gxo.com)\n\nThe pilot, contract, and subsequent operating record represent different kinds of evidence. The pilot tested feasibility; the agreement established a commercial service arrangement; repeated tote transfers supplied evidence of operation beyond a demonstration. In a subsequent deployment update, Agility reported that Digit had moved more than 100,000 totes at the Flowery Branch facility, with work also including stacking totes at another floor location. This supplier-reported cumulative count documents a specific material-handling application, but does not by itself establish an hourly transfer rate, availability, intervention frequency, or cost advantage over alternative automation systems. (investors.gxo.com)" }, { "heading": "Safety and responsible interaction", "text": "Humanoid robots combine the hazards of powered manipulators with those of mobile, potentially unstable machines. Moving arms can strike people; fingers and joints create pinch points; and a person can become trapped between the robot, its payload, and fixed structures. Tools and carried objects introduce hazards beyond those of the robot’s body. Bipedal systems additionally present a falling hazard: loss of balance can move the torso, limbs, and payload outside their normal operating envelope. Stopping is therefore not simply equivalent to removing power. A 2021 study of humanoid safety explicitly distinguished rapidly interrupting a task from stopping locomotion while maintaining balance. (osha.gov)\n\nRisk assessment addresses the complete application rather than the robot in isolation. ISO 12100:2010 establishes a machinery-safety methodology covering hazard identification, risk estimation and evaluation, risk reduction, and verification throughout relevant lifecycle phases. For a humanoid application, operating limits include its tasks, payloads, surroundings, and foreseeable human access; setup, fault recovery, cleaning, and maintenance also require examination. OSHA’s technical guidance identifies programming, environmental conditions, malfunctions, and worker activities as assessment inputs. Consequently, changing a gripper, adding a sharp tool, or introducing a conveyor changes the application being assessed, even when the robot itself is unchanged. (iso.org)\n\nProtective measures can include inherently lower-energy designs, guarded workspaces, monitored separation, and safety-related limits on speed, force, or power. Separation monitoring can slow or stop motion as a person approaches; contact-limited operation instead requires assessment of the forces and pressures produced by the robot, end-effector, and workpiece together. These are distinct protective strategies, not consequences of humanlike appearance. Maintenance introduces a further distinction between a controlled operating stop and isolation of hazardous energy: stored or supplied energy must be addressed during servicing, rather than treating a stopped robot as necessarily de-energized. (osha.gov)\n\nISO 13482:2014, published in February 2014 and confirmed in 2020, remained the published edition on October 15, 2026; its proposed replacement, ISO/FDIS 13482, was still under development. The 2014 standard concerns earthbound personal-care robots, particularly mobile servants, physical assistants, and person carriers, including applications involving physical human–robot contact. It excludes robots travelling faster than 20 km/h, toys, waterborne and flying robots, industrial robots, robots that are medical devices, and military or public-force applications. It is therefore not a universal humanoid-robot standard. Industrial-robot safety is addressed separately by ISO 10218-1:2025 and ISO 10218-2:2025, distinguishing robot requirements from application and cell integration. Part 1 explicitly excludes mobile-platform mobility hazards, illustrating why an industrial manipulator standard alone does not resolve every hazard of a walking machine. (iso.org)\n\nAnthropomorphic presentation also affects judgments about appropriate interaction, but those judgments are not capability measurements. In a 2015 experiment, Matthieu Destephe and colleagues analysed responses from 69 participants—47 French and 22 Japanese—who watched videos of WABIAN-2R performing different walking patterns. Perceived attractiveness correlated with occupational acceptability (\\(r_s = 0.347\\)); perceived humanness showed a smaller correlation (\\(r_s = 0.209\\)), while eeriness was not significantly associated with occupational acceptability. These questionnaire results concern perceived suitability, not demonstrated competence or safe physical collaboration. The distinction is important when interpreting expressive movement: ratings of human likeness or acceptance do not establish that a machine understands a situation, can execute the imagined occupation, or can safely tolerate unintended contact. (frontiersin.org)" }, { "heading": "Performance assessment and unresolved engineering problems", "text": "Task-level assessment distinguishes completing useful work from reproducing an individual movement. Relevant measures include the proportion of attempts completed, elapsed execution time, recovery after disturbances or failed actions, and the amount of human supervision required. Completion criteria must specify the final state: grasping a component, for example, is not equivalent to placing it correctly. Intervention records distinguish autonomous retries from operator commands, remote control, and physical resets. These measures describe different properties: a robot may complete tasks reliably while demanding substantial operator attention. NIST’s evaluation framework accordingly separates component capabilities, functional tasks, and operational scenarios, and treats interaction time and robot attention demand as distinct from task-completion time. (tsapps.nist.gov)\n\nMeaningful comparisons also require a documented operating envelope. Terrain geometry and friction, payload mass and placement, object properties, sensing arrangements, power supply, and communication conditions can change the difficulty of nominally identical tasks. The DARPA Robotics Challenge finals, held on June 5–6, 2015, illustrate how procedures define the capability being measured. Robots had an hour to attempt eight disaster-response tasks, including driving, opening a door, operating a valve, cutting a wall, traversing difficult terrain, and climbing stairs. They operated on onboard power without fall-arrest tethers, while deliberately degraded wireless communications tested the division of responsibility between robots and human supervisors. Performance therefore represented a human–robot system under specified constraints, not unrestricted autonomous competence. (tsapps.nist.gov)\n\nReproducibility and generalization require different evidence. Repeating a scripted demonstration establishes repeatability in its prepared setting; it does not establish transfer to unfamiliar objects or environments. Simulation permits controlled comparisons and repeated trials, but physical experiments test whether the model captures consequential properties of the machine and its surroundings. In their 2024 study, Ilija Radosavovic and colleagues trained a Digit locomotion controller using reinforcement learning in randomized simulations and transferred it to hardware without additional training. They reported no falls during a week of full-day outdoor testing. Their quantitative controller comparison, however, used ten simulated runs per terrain with 95% confidence intervals; unstable-plank terrain was not tested physically because of possible hardware damage. The controller also lacked cameras and could become trapped by obstacles. These results support specific locomotion capabilities rather than general autonomous navigation. (hybrid-robotics.berkeley.edu)\n\nIntegrated locomotion and manipulation remain difficult because contact decisions, balance, and object motion must be solved together. A controller may need to change foot placement while maintaining a grasp or applying force to the environment. Adding impact, rolling, or deformable-contact models increases computational complexity; formulations that choose contact sequences automatically introduce difficult numerical constraints. Dexterous hands further enlarge the action space and make informative physical training data expensive to collect. A 2025 research survey identified real-time optimization, uncertainty in contacts, and insufficient force-rich datasets as continuing obstacles to whole-body task execution. (arxiv.org)\n\nEnergy efficiency is measured differently from peak speed or payload. DARPA’s January 2015 description of the upgraded hydraulic Atlas specified a 3.7-kilowatt-hour lithium-ion battery and a potential hour of mixed operation involving walking, standing, and tool use. That estimate was explicitly workload-dependent, not a universal endurance rating. Long-horizon reliability similarly requires sequences of work rather than isolated successes: recovery, supervisory effort, and the ability to resume an interrupted task become part of operational performance. NIST’s baseline benchmark proposal, published in April 2026 and updated in May, addresses part of this measurement problem through reproducible apparatus for mobility, dexterity, coordinated locomotion–manipulation, and confined-space tasks. It describes a benchmark under development, rather than an established certification of general-purpose capability. (darpa.mil)" } ], "links": [ "robot", "control-theory", "artificial-intelligence", "reinforcement-learning", "degrees-of-freedom", "feedback", "rigid-body", "electric-motor", "torque", "heat", "camera", "object-detection", "calibration", "measurement-uncertainty", "friction", "center-of-mass", "angular-momentum", "mathematical-optimization", "policy-reinforcement-learning", "large-language-model", "multimodal-learning", "open-source-software", "experimental-design", "cognitive-science", "education", "infrastructure", "logistics", "automation", "force", "reproducibility", "generalization", "lithium-ion-battery" ], "sources": [ { "title": "Robonaut2 - NASA", "url": "https://www.nasa.gov/robonaut2/" }, { "title": "What Is Robonaut? (Grades 5-8) - NASA", "url": "https://www.nasa.gov/learning-resources/for-kids-and-students/what-is-robonaut-grades-5-8/" }, { "title": "Robonaut 2 Fact Sheet", "url": "https://www.nasa.gov/wp-content/uploads/2016/09/robonaut2_508.pdf" }, { "title": "Robotic Technology Lends More Than Just a Helping Hand - NASA", "url": "https://www.nasa.gov/technology/robotic-technology-lends-more-than-just-a-helping-hand/" }, { "title": "NASA's Ironman-Like Exoskeleton Could Give Astronauts, Paraplegics Improved Mobility and Strength", "url": "https://www.nasa.gov/technology/nasas-ironman-like-exoskeleton-could-give-astronauts-paraplegics-improved-mobility-and-strength/" }, { "title": "The Jaquet-Droz automata", "url": "https://www.mahn.ch/en/expositions/automates-jaquet-droz" }, { "title": "MOUVEMENTS", "url": "https://www.mahn.ch/fileadmin/mahn/EXPOSITIONS/EXPOSITIONS_ACTUELLES/Mouvements/MVTS_Textes_Expo.pdf" }, { "title": "Automate \"L'Ecrivain\"", "url": "https://collections.mahn.ch/fr/notice/aa-2-automate-l-ecrivain-4c56a476-8b6e-431f-aa2a-b90ed1abd7a5" }, { "title": "Humanoid History -WABOT", "url": "https://www.humanoid.waseda.ac.jp/booklet/kato_2.html" }, { "title": "History of robotics development", "url": "https://global.honda/en/ASIMO/history/?from=asimo_top" }, { "title": "Honda Debuts New Humanoid Robot \"ASIMO\"", "url": "https://global.honda/en/newsroom/news/2000/c001120b-eng.html" }, { "title": "Atlas Robot Debuts", "url": "https://www.darpa.mil/about/innovation-timeline/atlas-robot-debut" }, { "title": "Human-robot Teams Compete June 5 at DARPA Finals", "url": "https://www.defense.gov/News/Article/604732/human-robot-teams-compete-june-5-at-darpa-finals/" }, { "title": "An Electric New Era for Atlas", "url": "https://bostondynamics.com/blog/electric-new-era-for-atlas/" }, { "title": "Atlas’ Evolution From Research Robot to Industrial Humanoid", "url": "https://bostondynamics.com/blog/atlas-evolution-from-research-robot-to-industrial-humanoid/" }, { "title": "MODERN ROBOTICS", "url": "https://hades.mech.northwestern.edu/images/9/90/MR-tablet-temporary.pdf" }, { "title": "R5 - NASA", "url": "https://www.nasa.gov/technology/r5/" }, { "title": "Motors_and_Actuators", "url": "https://fab.cba.mit.edu/classes/865.24/topics/motors_and_actuators/index.html" }, { "title": "Drake: JointActuator< T > Class Template Reference", "url": "https://drake.mit.edu/doxygen_cxx/classdrake_1_1multibody_1_1_joint_actuator.html" }, { "title": "Ch. 2 - Let's get you a robot", "url": "https://manipulation.mit.edu/robot.html" }, { "title": "An Electric New Era for Atlas", "url": "https://bostondynamics.com/blog/electric-new-era-for-atlas" }, { "title": "Cog Project Overview", "url": "https://groups.csail.mit.edu/lbr/humanoid-robotics-group/cog/hardware.html" }, { "title": "Ch. 2 - Let's get you a robot", "url": "https://manipulation.csail.mit.edu/robot.html" }, { "title": "Series Elastic Actuators", "url": "https://groups.csail.mit.edu/lbr/humanoid-robotics-group/cog/cog-publications/mattw-nelastic.pdf" }, { "title": "MRV and Robotic", "url": "https://ntrs.nasa.gov/api/citations/20150008171/downloads/20150008171.pdf?attachment=true" }, { "title": "Optimal Actuator Design", "url": "https://biomimetics.mit.edu/research/1a9ba04b-d200-4743-b8fc-bf231f3231f0/" }, { "title": "Contact-Aided Invariant Extended Kalman Filtering for Legged Robot State Estimation", "url": "https://arxiv.org/abs/1805.10410" }, { "title": "Legged Robot State-Estimation Through Combined Forward Kinematic and Preintegrated Contact Factors", "url": "https://arxiv.org/abs/1712.05873" }, { "title": "Drift-Free Humanoid State Estimation Fusing Kinematic, Inertial and LIDAR Sensing", "url": "https://groups.csail.mit.edu/robotics-center/public_papers/Fallon14a.pdf" }, { "title": "Perception Methods for Continuous Humanoid", "url": "https://groups.csail.mit.edu/robotics-center/public_papers/Marion16a.pdf" }, { "title": "Reconstructing occluded Elevation Information in Terrain Maps with Self-supervised Learning", "url": "https://arxiv.org/abs/2109.07150" }, { "title": "Ch. 5 - Highly-articulated Legged Robots", "url": "https://underactuated.mit.edu/humanoids.html" }, { "title": "Biped Walking Pattern Generation by using Preview Control of Zero-Moment Point", "url": "https://people.csail.mit.edu/katiebyl/kb/DW2008/papers_of_tangential_interest/kajita03.pdf" }, { "title": "Capture Point: A Step toward Humanoid Push Recovery", "url": "https://www.cs.cmu.edu/~cga/legs/Pratt_Goswami_Humanoids2006.pdf" }, { "title": "Computationally-Robust and Efficient Prioritized Whole-Body Controller with Contact Constraints", "url": "https://arxiv.org/abs/1807.01222" }, { "title": "Real-world humanoid locomotion with reinforcement learning", "url": "https://hybrid-robotics.berkeley.edu/publications/ScienceRobotics2024_Learning_Humanoid_Locomotion.pdf" }, { "title": "Learning Getting-Up Policies for Real-World Humanoid Robots", "url": "https://arxiv.org/abs/2502.12152" }, { "title": "Dexterous Manipulation", "url": "https://www.eng.yale.edu/grablab/research_dexterous.html" }, { "title": "Yale OpenHand Project", "url": "https://www.eng.yale.edu/grablab/openhand/" }, { "title": "Planning Visual-Tactile Precision Grasps via Complementary Use of Vision and Touch", "url": "https://arxiv.org/abs/2212.08604" }, { "title": "Slip detection for grasp stabilisation with a multi-fingered tactile robot hand", "url": "https://arxiv.org/abs/2010.01928" }, { "title": "A humanoid upper body system for two-handed manipulation", "url": "https://rmc.dlr.de/rm/de/staff/extcms/images/rmc/users/hillenbd/Borst_etal_07.pdf" }, { "title": "TACT: Humanoid Whole-body Contact Manipulation through Deep Imitation Learning with Tactile Modality", "url": "https://arxiv.org/abs/2506.15146" }, { "title": "Brace Yourself: Task-Conditioned Environmental Bracing for Forceful Humanoid Manipulation", "url": "https://arxiv.org/abs/2609.25486" }, { "title": "The Robonaut 2 Hand – Designed To Do Work With Tools", "url": "https://ntrs.nasa.gov/api/citations/20110023122/downloads/20110023122.pdf?attachment=true" }, { "title": "Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning", "url": "https://arxiv.org/html/2501.02116v1" }, { "title": "Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation", "url": "https://arxiv.org/abs/2403.04436" }, { "title": "SayCan: Grounding Language in Robotic Affordances", "url": "https://say-can.github.io/?trk=public_post_main-feed-card-text" }, { "title": "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances", "url": "https://say-can.github.io/assets/palm_saycan.pdf" }, { "title": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", "url": "https://arxiv.org/html/2503.14734v1" }, { "title": "I Reach Faster When I See You Look: Gaze Effects in Human–Human and Human–Robot Face-to-Face Cooperation", "url": "https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2012.00003/full" }, { "title": "Speech Driven Gaze in a Face-to-Face Interaction", "url": "https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2021.598895/full" }, { "title": "RobotCub Science", "url": "https://www.robotcub.org/index.php/robotcub/results/robotcub_science.html" }, { "title": "RobotCub – An Open Framework for Research in Embodied Cognition", "url": "https://www.robotcub.org/index.php/robotcub.html" }, { "title": "Walking in the uncanny valley: importance of the attractiveness on the acceptance of a robot as a working partner", "url": "https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2015.00204/full" }, { "title": "Persistence of the uncanny valley: the influence of repeated interactions and a robot's attitude on its perception", "url": "https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2015.00883/full" }, { "title": "The RobotCub Summer School", "url": "https://www.robotcub.org/index.php/robotcub/dissemination/the_robotcub_summer_school.html" }, { "title": "A Step Up for NASA’s Robonaut: Ready for Climbing Legs", "url": "https://www.nasa.gov/technology/a-step-up-for-nasas-robonaut-ready-for-climbing-legs/" }, { "title": "NASA’s Robonaut Legs Headed for International Space Station", "url": "https://www.nasa.gov/general/nasas-robonaut-legs-headed-for-international-space-station/" }, { "title": "Crew Works Multitude of Advanced Robotics", "url": "https://www.nasa.gov/blogs/spacestation/2014/12/17/crew-works-multitude-of-advanced-robotics/" }, { "title": "Upgraded Atlas Robot to Go Wireless as the Stakes Are Raised for the DARPA Robotics Challenge Finals", "url": "https://www.darpa.mil/news/2015/atlas-robot-wireless" }, { "title": "Twenty-five Teams From around the World to Participate in DARPA Robotics Challenge Finals", "url": "https://www.darpa.mil/news/2015/robotics-challenge-finals" }, { "title": "GXO Signs Industry-First Multi-Year Agreement with Agility Robotics", "url": "https://investors.gxo.com/news-releases/news-release-details/gxo-signs-industry-first-multi-year-agreement-agility-robotics/" }, { "title": "Digit Moves Over 100,000 Totes in Commercial Deployment", "url": "https://www.agilityrobotics.com/content/digit-moves-over-100k-totes" }, { "title": "OSHA Technical Manual (OTM) — Section IV: Chapter 4", "url": "https://www.osha.gov/otm/section-4-safety-hazards/chapter-4" }, { "title": "A behavior-based framework for safe deployment of humanoid robots", "url": "https://link.springer.com/article/10.1007/s10514-021-09978-5" }, { "title": "Advancing Capabilities of Industrial Robots Through Evaluation, Benchmarking, and Characterization", "url": "https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=926493" }, { "title": "Humanoid Robot Baseline Performance Benchmark", "url": "https://www.nist.gov/el/intelligent-systems-division-73500/humanoid-robot-baseline-performance-benchmark" } ], "license": "https://creativecommons.org/licenses/by/4.0/" }