AI Glossary / Reinforcement learning and robotics / Intermediate
Motion planning
Definition
Motion planning is the process of finding a feasible movement from a robot's starting configuration to a goal while respecting relevant constraints. These commonly include collision avoidance and joint limits, and may include velocity, acceleration, or dynamic requirements. A motion plan must account for the robot's whole configuration, not just the destination of one point.
What it means
Knowing where a robot should finish does not tell it how to get there. A mobile base may need to pass through a narrow opening, while a robot arm must position every link so the tool reaches its target without hitting surrounding objects.
The planning problem depends on what movement is permitted and what the environment model contains. A feasible solution need not be shortest or fastest. Optimization adds a criterion for choosing among feasible motions, and execution still requires a controller and appropriate feedback.
How it works
A robot configuration specifies its geometric arrangement, such as joint angles for an arm or position and heading for a mobile base. The set of possible configurations is the configuration space. Configurations that cause collisions or violate constraints are excluded.
A geometric planner seeks a continuous path through the permitted configurations. Search, sampling-based methods, and optimization are among the approaches used. Collision checking must examine motions between candidate configurations, not only their endpoints.
A path describes a sequence of configurations without necessarily specifying timing. A trajectory includes timing; its speed and acceleration must be feasible. Some problems incorporate motion dynamics directly during planning rather than applying timing afterward.
The result is evaluated against the modeled environment and robot. Changes in obstacles, localization error, or tracking error can invalidate assumptions made during planning and may require monitoring or replanning.
Three practical examples
- Manipulation: a robot arm moves a part around fixtures while accounting for its links, gripper, and attached object.
- Mobile navigation: a robot finds a route through obstacles that fits its footprint and respects relevant turning constraints.
- Animation and simulation: articulated motion is planned through a virtual environment, using the geometry and movement rules of the simulated system.
Think of it this way
Think of carrying a long ladder around a corner. The point you want to reach is easy to name, but the ladder's orientation and the space swept by its entire length determine whether the move works. A line drawn for one end of the ladder misses much of the problem. Robot motion planning formalizes this whole-body reasoning.
Worked example
Consider an original planar arm with two links, each one meter long. Let theta_1 be the first link's angle from the horizontal, and theta_2 the second link's angle relative to the first. The tool position is x = cos(theta_1) + cos(theta_1 + theta_2), y = sin(theta_1) + sin(theta_1 + theta_2).
At the starting configuration (0 degrees, 90 degrees), the elbow is at (1, 0) and the tool at (1, 1). At the goal configuration (90 degrees, -90 degrees), the elbow is at (0, 1), while the tool is again at (1, 1). Identical tool positions do not imply identical arm configurations.
Place a small circular obstacle centered at (square root of 2, square root of 2), approximately (1.414, 1.414), with radius 0.1 meters. Assume idealized zero-thickness links solely for this calculation. Neither endpoint configuration intersects this obstacle; the nearest endpoint of each relevant outer segment is (1, 1), approximately 0.586 meters from its center.
Now interpolate the joint angles linearly. Halfway, the configuration is (45 degrees, 0 degrees), so both links point along the same diagonal. The tool is exactly at (square root of 2, square root of 2), the obstacle's center. The interpolated motion is therefore invalid despite both endpoint configurations being collision free.
This does not prove that no valid path exists. It proves only that this candidate fails. A planner must search for and check alternatives. Real links have thickness, and a carried object changes the geometry, so a practical model requires more than the simplified line segments used here.
Illustrative case study
Fictional case: Maple Assembly plans a robot-arm transfer between two fixtures. Engineers initially verify that the pickup and placement poses are reachable, then discover that the direct joint-space interpolation sweeps the carried part through a fixture.
They add the part's attached geometry to the planning model and test alternative intermediate configurations. After finding a collision-free path in the model, they evaluate timing constraints and controlled execution with the actual payload.
The team records failures separately for reachability, collision checking, and tracking. This hypothetical case illustrates why a successful endpoint calculation is insufficient; it reports no measured cycle-time improvement or deployment guarantee.
Common misconceptions and mistakes
Inverse kinematics can find a configuration for a tool target, but does not by itself provide a collision-free path to that configuration. Checking only waypoints can miss collisions between them. A shortest geometric path may be dynamically infeasible or undesirable to execute.
A planner's failure within a time budget does not necessarily prove that no feasible motion exists. Claims of completeness or optimality depend on the algorithm and assumptions. Finally, a plan valid in an inaccurate model may fail when executed.
Business use
Motion planning supports automation where the route and body arrangement matter as much as the endpoint. Teams should evaluate planning success, computation time, clearance, execution quality, and task performance under representative conditions. The robot's tools, payloads, and environmental changes must be reflected in the model. Comparisons should separate finding any valid motion from optimizing a motion for speed, energy, or another business objective.
People also ask
Is motion planning the same as control?
No. Planning determines intended movement; control generates actions to follow it, although some systems combine these functions closely.
Is a path always a trajectory?
No. A geometric path need not contain timing information.
Does motion planning require machine learning?
No. Many established planners use geometric, search, sampling, or optimization methods without learned components.
Check your understanding
Why are collision-free start and goal configurations insufficient in the worked example?
Answer: The arm passes through a different configuration halfway along the proposed interpolation, placing the tool inside the obstacle. The movement between endpoints must also be checked.
Related terms
- Configuration space
- Path planning
- Inverse kinematics
- Trajectory planning
- Collision checking
- Robot footprint
- Kinodynamic planning
- Motion control
Learn next
Compare workspace coordinates with joint configurations using a two-link arm. Then study configuration-space obstacles and continuous collision checking before adding timing and dynamic constraints to a geometric path.
More in Reinforcement learning and robotics
- Modified policy iteration - Modified policy iteration is a dynamic-programming method that alternates greedy policy improvement with partial policy evaluation.
- Monte Carlo localization - Monte Carlo localization is a robot localization method that represents uncertainty about pose using a set of weighted samples called particles.
- Morphogenetic robotics - Morphogenetic robotics uses ideas from biological morphogenesis - the processes that generate an organism's form - to design robot structures, controllers...
- Multiarmed bandit - A multiarmed bandit is a sequential decision problem in which a learner repeatedly chooses among alternatives, called arms, and observes the reward of the...
- N-step Q(sigma) - N-step Q(sigma) is a family of multi-step action-value learning algorithms that mixes sampled action backups with expectations over actions.
- n-step return - An n-step return is a reinforcement-learning target that combines the next n observed rewards with an estimated continuation value at the end of that horizon.
Sources
Steven M. LaValle, Planning Algorithms, Chapter 4, The Configuration Space: https://lavalle.pl/planning/ch4.pdf Steven M. LaValle, Planning Algorithms, Chapter 14, Sampling-Based Planning Under Differential Constraints: https://lavalle.pl/planning/ch14.pdf IEEE ROBOTS, Robotics Glossary, Motion Planning: https://robotsguide.com/learn/robotics-glossary
Independent review checked configuration-space and differential-constraint formulations.
Previous: Motion model Next: Motion vector