{
  "id": "motion-planning",
  "title": "Motion Planning",
  "lang": "en",
  "source_lang": "en",
  "summary": "Motion planning computes feasible movements for robots and other systems while accounting for obstacles, physical constraints, uncertainty, and performance objectives.",
  "category": "technology",
  "subcategory": "engineering",
  "type": "concept",
  "aliases": [
    "Robotic motion planning",
    "Robot motion planning"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/motion-planning/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/motion-planning/",
      "json": "https://aiwiki.page/zh/motion-planning.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Motion planning is the computational problem of determining how a physical or simulated system can move from an initial condition to a desired goal while respecting constraints. In robotics, these constraints commonly include avoiding collisions, obeying joint limits, maintaining balance, and staying within available forces or torques. Related problems arise in autonomous vehicles, manufacturing, spacecraft operations, computer animation, and the analysis of molecular motion. Depending on the application, a solution may be a geometric path, a time-dependent trajectory, or a policy that selects actions from observations.\n\nThe central difficulty is that feasible motion depends on the configuration of an entire system, not merely the position of a representative point. A robot arm may have to move its elbow away from an obstacle before its gripper can approach a target; a car cannot move directly sideways; and a walking robot must preserve appropriate contact with the ground. Motion planning therefore combines geometry, search, mathematical optimization, and models of physical evolution. Practical systems also connect planning to perception and feedback control, because maps, object positions, and executed motions are never perfectly known."
    },
    {
      "heading": "Problem formulation and levels of planning",
      "text": "A motion-planning problem begins with a description of the system, its environment, an initial condition, and a goal. For a rigid mobile robot, the configuration might consist of planar position and orientation. For an articulated manipulator, it typically consists of joint coordinates. When velocity, momentum, or other evolving quantities affect what can happen next, the planning state must include them as well. This distinction separates geometric planning from planning under differential constraints.\n\nA geometric path is usually written as a continuous mapping\n\n\\[q:[0,1]\\rightarrow \\mathcal{C}_{\\mathrm{free}},\\]\n\nwhere \\(\\mathcal{C}_{\\mathrm{free}}\\) is the set of collision-free configurations. The parameter need not represent time. A trajectory instead assigns states to physical times and must satisfy the system’s evolution equations. A path that curves sharply around an obstacle may be geometrically valid but impossible to follow at a specified speed. Conversely, changing only the timing of a path may resolve acceleration violations or avoid an obstacle that moves through the route.\n\nFeasibility and optimality are different requirements. A feasible solution satisfies the stated constraints; an optimal solution minimizes a specified objective among feasible alternatives. Common objectives include travel time, path length, energy expenditure, clearance penalties, and combinations of these quantities. A general trajectory objective is\n\n\\[J=\\phi(x(T))+\\int_0^T L(x(t),u(t))\\,dt,\\]\n\nwhere \\(x(t)\\) is the state, \\(u(t)\\) the control input, \\(L\\) a running cost, and \\(\\phi\\) a terminal cost. These choices are not interchangeable: the shortest route can require frequent braking, while a longer route may consume less energy.\n\nMotion planning occupies a particular level in an autonomous system. Task planning chooses actions such as picking an object, opening a door, or visiting a workstation. Motion planning determines the movements that realize those actions. Feedback control adjusts commands during execution to follow the selected motion despite disturbances and modeling errors. The boundaries can overlap, especially when a controller repeatedly solves a planning problem over a short future horizon.\n\nAn important theoretical distinction concerns the planner’s guarantees. A complete algorithm finds a solution whenever one exists and correctly reports failure otherwise, under its modeling assumptions. Resolution completeness provides such a guarantee only relative to a chosen discretization. Probabilistic completeness means that, for a solvable problem satisfying specified conditions, the probability of finding a solution approaches one as computational effort increases. None of these properties by itself guarantees short runtime, good path quality, or safe behavior under an inaccurate model."
    },
    {
      "heading": "Configuration spaces and collision geometry",
      "text": "The configuration space provides the standard geometric language for motion planning. It represents every possible arrangement of a system as a point in an abstract space. A planar translating body has two configuration coordinates; a freely moving planar rigid body has three, including orientation. A free rigid body in three-dimensional space has six degrees of freedom. A serial manipulator with seven independently movable joints generally has a seven-dimensional configuration space before additional constraints are imposed.\n\nConfiguration spaces need not be ordinary Euclidean spaces. A revolute joint whose angle wraps around has the topology of a circle, and three-dimensional rotations belong to a curved space rather than a single globally nonsingular three-coordinate chart. Distance calculations and interpolation must respect these structures. For example, interpolating directly between numerical angles near \\(180^\\circ\\) and \\(-180^\\circ\\) can produce an unnecessarily long rotation unless angular periodicity is handled correctly.\n\nObstacles in the physical workspace induce forbidden regions in configuration space. Each forbidden point represents an arrangement in which some part of the robot intersects an obstacle or another part of itself. For a translating rigid body, this construction can be expressed using a Minkowski sum: reflecting the body about its reference point and adding it to a workspace obstacle produces the corresponding configuration-space obstacle. This transforms a finite-size-body problem into a point-motion problem in an expanded obstacle field. Rotating bodies and articulated systems produce more complicated forbidden regions that are rarely constructed explicitly.\n\nPractical planners therefore rely heavily on collision detection. Broad-phase tests use bounding boxes, spheres, spatial trees, or other inexpensive approximations to reject pairs that cannot intersect. Narrow-phase tests examine the remaining geometry more precisely. Robot links may be represented by convex shapes, triangle meshes, or simplified capsules. A motion between two collision-free endpoints is not necessarily collision-free: an arm can sweep through an obstacle between them. Planners must either check intermediate configurations at an adequate resolution or use continuous collision-detection methods that bound or compute contact along a motion.\n\nAdditional geometric constraints can restrict feasible configurations to lower-dimensional subsets. A robot carrying a full container may need to keep it upright; two arms holding the same rigid object must maintain mutually consistent poses. The relevant relationships follow from robot kinematics, while end-effector targets often admit multiple joint configurations through inverse kinematics. These alternatives matter because one solution may be blocked or near a singularity while another allows an unobstructed approach. Planning on such constrained sets requires more than sampling arbitrary joint coordinates, since randomly drawn configurations will generally fail exact equality constraints."
    },
    {
      "heading": "Graph search, exact methods, and computational complexity",
      "text": "Early robotic systems helped establish the connection between physical movement and symbolic search. Shakey, developed at SRI from the late 1960s into the early 1970s, combined perception, planning, and mobile action in structured indoor environments. A major general-purpose search method, A*, was published by Peter Hart, Nils Nilsson, and Bertram Raphael in 1968. Later work placed motion planning on a more explicit geometric foundation: Tomás Lozano-Pérez’s 1983 treatment of configuration space became particularly influential.\n\nGraph-based planning replaces continuous movement with a network of candidate states and transitions. A grid planner may assign one vertex to each traversable cell, while a manipulator planner may discretize joint angles. Edges represent allowable local motions and carry costs such as distance or travel time. Dijkstra’s algorithm finds minimum-cost paths when edge costs are nonnegative. A* prioritizes vertices using the cost already incurred plus a heuristic estimate of the remaining cost. With suitable admissibility conditions and correct graph-search handling, it retains shortest-path guarantees while often exploring fewer states.\n\nThe graph’s construction determines what the guarantee means. A shortest path on an eight-connected occupancy grid need not be the shortest continuous path through the environment. A coarse grid can eliminate a narrow passage entirely, while a fine grid can require excessive memory. With \\(k\\) samples along each of \\(d\\) coordinates, a full discretization contains \\(k^d\\) states. This exponential growth illustrates the dimensionality problem that makes straightforward grids unattractive for highly articulated robots.\n\nSome geometric problems permit specialized exact constructions. For a point moving among polygonal obstacles in a plane, a visibility graph connects mutually visible obstacle vertices and the start and goal; an appropriate shortest path can be found on this graph. Cell-decomposition methods divide free space into connected regions and search their adjacency structure. Exact algebraic methods can address much broader classes of geometrically specified systems, but their computational costs limit routine use. John Canny’s 1988 work established a polynomial-space upper bound for a broad semialgebraic formulation of the general motion-planning problem. Separate hardness results show why no uniformly efficient solution should be expected for unrestricted formulations.\n\nSearch is also useful when physical constraints are embedded in the transitions. State lattices connect discretized vehicle states using precomputed feasible motion primitives, such as short forward arcs, reverse maneuvers, and steering transitions. Hybrid A* combines continuous vehicle-state propagation with discretized search bookkeeping. These approaches preserve more vehicle structure than a simple position grid, although their guarantees depend on their specific representation and implementation. Incremental search methods can reuse previous work after map changes, making them useful when a moving robot discovers obstacles rather than planning once in a fully known environment."
    },
    {
      "heading": "Sampling-based planning",
      "text": "Sampling-based motion planning avoids explicitly representing the entire free space. Instead, it generates configurations, tests them for validity, and connects selected configurations with local motions. This approach became especially important for robot arms and other systems with many degrees of freedom, where exact geometric decomposition and dense grids become impractical. Its effectiveness rests on replacing global geometric construction with many relatively simple validity queries.\n\nThe probabilistic roadmap method, prominently developed in work published in 1996 by Lydia Kavraki, Petr Švestka, Jean-Claude Latombe, and Mark Overmars, constructs a graph of sampled collision-free configurations. Nearby samples are connected when a local planner can produce a valid motion between them. Once the roadmap is built, start and goal configurations can be attached and graph search can answer queries. This makes roadmaps particularly suitable for repeated planning in a largely unchanged environment, such as a robot arm serving several stations in one workcell.\n\nThe rapidly exploring random tree, introduced by Steven LaValle in 1998, grows outward from an initial state. At each iteration, a sampled target identifies a nearby tree vertex, and the planner extends the tree toward that target. Large unexplored regions tend to attract more extensions because their nearby vertices have larger regions of influence. Bidirectional variants grow trees from both start and goal; RRT-Connect, introduced by James Kuffner and LaValle in 2000, repeatedly extends one tree toward the other and is widely used for single-query geometric problems.\n\nUnder appropriate assumptions, major roadmap and tree methods are probabilistically complete. Their failure to find a path within a time limit, however, does not establish that no path exists. A narrow passage may occupy a very small fraction of configuration space, so uniform random sampling can miss it for a long time. Biased samplers, obstacle-boundary sampling, bridge tests, and task-specific sampling distributions attempt to address such difficulties. The choice of distance metric and nearest-neighbor method also strongly affects performance, especially when translation and rotation have different units or joints have different effects on the robot’s swept volume.\n\nBasic sampling methods do not automatically produce high-quality paths. RRT* and PRM*, developed by Sertac Karaman and Emilio Frazzoli in work published in 2011, provide asymptotic optimality under stated conditions: their best solution cost converges toward the optimum as the number of samples grows. RRT* achieves improvement partly by choosing better parent connections and rewiring nearby tree vertices. These results do not specify that convergence will be fast enough for a particular application. Practical systems often use a fast planner to obtain an initial solution and then apply shortcutting, smoothing, or trajectory optimization, while rechecking collision and physical constraints after every modification."
    },
    {
      "heading": "Trajectory optimization and optimal control",
      "text": "Where graph search and sampling methods explore alternative routes, trajectory optimization directly adjusts a parameterized motion to improve an objective while satisfying constraints. The connection to optimal control is close: both seek state and control histories governed by dynamical equations. Motion-planning applications place particular emphasis on obstacle avoidance, contact, and finding a useful solution from a difficult initial guess.\n\nA typical discrete formulation chooses states \\(x_0,\\ldots,x_N\\) and controls \\(u_0,\\ldots,u_{N-1}\\), then solves\n\n\\[\\min \\sum_{k=0}^{N-1}\\ell(x_k,u_k)+\\phi(x_N)\\]\n\nsubject to\n\n\\[x_{k+1}=F(x_k,u_k),\\qquad g(x_k,u_k)\\leq 0.\\]\n\nThe function \\(F\\) approximates the system’s time evolution, while \\(g\\) encodes limits such as actuator bounds, joint ranges, or obstacle separation. Additional equations can impose contacts, endpoint conditions, or cooperative manipulation constraints. The time intervals may be fixed or included as decision variables.\n\nDifferent numerical formulations trade off problem size and structure. Direct shooting optimizes controls and generates states by simulation, but sensitivities can become poorly conditioned over long horizons. Multiple shooting divides the horizon into segments with continuity constraints. Direct collocation optimizes states and controls at selected time points while imposing discretized dynamics between them. Sequential quadratic programming and related nonlinear-programming methods exploit derivatives and sparse structure to solve these problems efficiently. Gradient information can come from analytic models, automatic differentiation, or numerical approximation.\n\nObstacle avoidance usually makes the problem nonconvex. A path passing to the left of an obstacle and a path passing to the right may occupy separate basins of attraction for a local solver. Artificial potential fields, influential in robotics following Oussama Khatib’s 1986 work, combine attraction toward a goal with repulsion from obstacles, but can become trapped in local minima. Modern trajectory optimizers use more elaborate objectives and constraints, yet they still generally depend on initialization. A sampling-based planner can supply a route, after which optimization improves smoothness, timing, or clearance without having to discover the route from scratch.\n\nRepresentative methods include CHOMP, introduced in 2009, STOMP in 2011, and TrajOpt in 2013. They differ in how they represent trajectories, estimate improvements, and treat obstacle costs or constraints. None removes the need to validate the resulting motion between discretization points. A trajectory may satisfy collision constraints at all optimized knots yet collide between them. Similarly, a small numerical constraint residual does not necessarily imply acceptable physical behavior. Mesh refinement, continuous checking, appropriate tolerances, and a final execution-oriented validation stage are therefore integral parts of a usable optimization pipeline."
    },
    {
      "heading": "Differential constraints, timing, and contact",
      "text": "A geometrically clear path can still be physically impossible. Systems with nonholonomic constraints cannot realize arbitrary instantaneous motions even when their configurations appear unconstrained. A simplified car, for example, moves in the direction of its wheels and cannot translate directly sideways. Its motion can be modeled as\n\n\\[\\dot{x}=v\\cos\\theta,\\qquad \\dot{y}=v\\sin\\theta,\\qquad \\dot{\\theta}=\\frac{v}{L}\\tan\\delta,\\]\n\nwhere \\(L\\) is wheelbase and \\(\\delta\\) is steering angle. Steering limits bound curvature, while steering-rate limits constrain how quickly curvature can change.\n\nClassic solutions illustrate how these restrictions alter shortest paths. Lester Dubins showed in 1957 that the shortest forward-only planar paths for a bounded-curvature vehicle have particular combinations of straight segments and maximum-curvature circular arcs. James Reeds and Lawrence Shepp extended the analysis in 1990 to a vehicle that can reverse. These models explain basic parking maneuvers, but they omit effects such as tire-force limits and steering transients. At higher speeds, acceleration and road friction must enter the planning model.\n\nKinodynamic planning treats kinematic and dynamic restrictions together. The state includes quantities such as velocity, and tree expansion may simulate candidate controls rather than interpolate directly between configurations. Connecting two arbitrary states is then difficult: a straight line in state space will generally violate the dynamics. Some algorithms require a two-point boundary-value solver, while others use forward propagation and carefully designed sampling or selection strategies. The suitability of a method depends strongly on controllability, available local steering procedures, and the required endpoint precision.\n\nAn alternative is to plan a geometric path first and then determine its timing. If the path is \\(q(s)\\), joint velocities and accelerations satisfy\n\n\\[\\dot q=q'(s)\\dot s,\\qquad \\ddot q=q''(s)\\dot s^2+q'(s)\\ddot s.\\]\n\nThese relations allow joint-speed, acceleration, and some torque constraints to be translated into restrictions on progress along the path. Time-optimal path parameterization seeks the fastest admissible traversal. It cannot repair every geometric deficiency, however: a path with unsuitable curvature, insufficient clearance, or incompatible contact geometry may remain infeasible at any useful speed.\n\nLegged locomotion and manipulation introduce changing contact modes. A walking robot alternates support feet, while a robot pushing an object may switch between sticking, sliding, and separation. Planning must account for force balance, friction, impact, and unilateral contact, in which a surface can push but not pull. Full rigid-body contact models can lead to difficult nonsmooth or complementarity-constrained optimization problems. Practical planners often use simplified models or hierarchical decompositions, then verify the resulting motion against more detailed dynamics before execution."
    },
    {
      "heading": "Uncertainty, changing environments, and multiple agents",
      "text": "Real robots plan from estimates rather than exact states. Wheel slip, imperfect calibration, delayed observations, and incomplete maps introduce uncertainty about both the robot and its surroundings. Sensor fusion combines observations from devices such as cameras, lidar, and inertial sensors to estimate the current situation, but it does not eliminate uncertainty. Planning must therefore distinguish the nominal predicted motion from the range of motions that could actually occur.\n\nA simple robust strategy enlarges obstacles or reserves tracking-error margins. More elaborate methods propagate sets of possible states and require every state in the set to remain safe. Chance-constrained planning instead limits a modeled probability of violation, for example\n\n\\[\\Pr\\!\\left(x_k\\in\\mathcal{X}_{\\mathrm{safe}}\\right)\\geq 1-\\varepsilon.\\]\n\nSuch a constraint is meaningful only relative to the assumed probability model. A per-step probability bound is not automatically the same as a bound on the probability of any collision over a complete journey, especially when errors across time are correlated.\n\nBelief-space planning represents uncertainty explicitly as a probability distribution over possible states. The planner may favor actions that improve future observations, such as moving where recognizable landmarks become visible. This couples motion with information gathering. A partially observable Markov decision process provides a general framework, but exact solutions are usually intractable for realistic continuous robots. Practical methods use approximations, limited horizons, restricted policy classes, or compact uncertainty representations.\n\nMoving obstacles add a temporal dimension. Passing through a location may be safe at one time and unsafe a moment later, so waiting can become a necessary planning action. Predicted human movements are especially uncertain because people respond to the robot and to one another. Replanning addresses changing predictions, while model predictive control repeatedly optimizes a finite-horizon trajectory and executes only an initial portion. This architecture improves responsiveness but requires safeguards against computation overruns, abrupt plan changes, and horizons too short to recognize an approaching dead end.\n\nCoordinating several robots introduces additional coupling. A centralized formulation can combine every robot’s state into one large planning problem, but its dimension grows with the fleet. Multi-agent path finding often uses a discrete graph and explicitly prohibits vertex conflicts and incompatible edge traversals. Conflict-based search, introduced by Guni Sharon, Roni Stern, Ariel Felner, and Nathan Sturtevant in 2012, separates high-level conflict resolution from lower-level individual path planning. Prioritized planning is often faster but can fail because an early robot’s chosen route blocks later robots. Continuous multi-robot systems must additionally handle physical size, timing tolerances, communication delays, and the risk of reciprocal avoidance producing deadlock."
    },
    {
      "heading": "Applications, software, and evaluation",
      "text": "Industrial manipulation is a major application of motion planning. An industrial robot may need to reach into a bin, carry a part around fixtures, and insert it into an assembly without colliding with tooling or its own links. The challenge is not merely reaching a final pose: the chosen grasp determines which transport motions are possible, and the placement can determine whether the gripper can withdraw. Task-and-motion planning therefore reasons jointly about discrete choices and their geometric feasibility. Welding and inspection add process constraints, such as maintaining tool orientation, speed, or sensor distance along a surface.\n\nMobile robots use planning to navigate warehouses, hospitals, and outdoor terrain. Road vehicles combine route-level decisions with lane selection, interaction prediction, and dynamically feasible local trajectories. A robotaxi must distinguish a legally or geometrically available maneuver from one that is safe under the current observations and predictions. Aerial vehicles add three-dimensional obstacles and strong dynamic coupling, while spacecraft planning may emphasize fuel consumption, restricted thrust, and rendezvous conditions. Similar mathematical techniques appear in animation, where collision-free movement must also satisfy stylistic or biomechanical requirements.\n\nSoftware libraries separate reusable planning machinery from application-specific models. The Open Motion Planning Library, introduced in a 2012 journal publication, provides numerous sampling-based algorithms and interfaces for state spaces and validity checking. MoveIt integrates planning with robot descriptions, collision checking, kinematics, and execution components. These tools reduce implementation effort but do not supply an automatically correct model. Units, joint conventions, attached-object geometry, controller limits, and coordinate transforms must agree throughout the system.\n\nEvaluation requires more than reporting one successful path. Relevant measures include success rate under a stated time limit, time to first solution, final objective value, minimum clearance, collision-checking cost, memory use, and execution success. Randomized planners should be tested over repeated runs, and benchmark sets should contain qualitatively different challenges, including narrow passages and constrained motions. Optimality comparisons require identical objectives and comparable computational budgets. A simulated solution is also distinct from a validated executable trajectory: discretization error, actuator saturation, latency, and tracking error can invalidate an otherwise convincing result.\n\nMachine learning is increasingly used to propose promising samples, predict useful search regions, estimate costs, and initialize trajectories. Such methods can accelerate familiar classes of problems, but distribution shifts may make their proposals unreliable. A common architecture treats learned components as proposal mechanisms and retains explicit collision and dynamics checks before accepting a solution. Links to formal verification and functional safety concern what can be guaranteed about the complete system, rather than the planner alone. Current research includes planning for deformable objects, long-horizon contact-rich manipulation, interactive behavior around people, and methods that combine learned models with verifiable physical constraints."
    }
  ],
  "links": [
    "robotics",
    "mathematical-optimization",
    "automated-planning",
    "control-theory",
    "configuration-space",
    "minkowski-sum",
    "a-star-search",
    "optimal-control",
    "sensor-fusion",
    "lidar",
    "partially-observable-markov-decision-process",
    "model-predictive-control",
    "industrial-robot",
    "robotaxi",
    "machine-learning",
    "formal-verification"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}