Models & learning
Inverse dynamics
Inverse dynamics is the problem of determining the forces, torques, or actions associated with a specified motion or state transition. In model-based robotics it commonly computes joint forces or torques from position, velocity, and acceleration under a dynamics model; learned inverse-dynamics models instead infer actions from observed states or outcomes and may not expose physical parameters.
Also known as: inverse dynamics model, IDM
Updated
Two uses share one direction of inference
Classical robot dynamics relates a robot's state and acceleration to the generalized forces or torques that produce that motion. Inverse dynamics works from the desired motion toward the required effort. Modern Robotics Chapter 8 derives this relationship for open-chain robots, while MuJoCo's computation documentation explains how its inverse-dynamics pipeline evaluates forces for a known state and acceleration.
Robot learning often uses the same name for a learned mapping. An inverse-dynamics model can observe two states, a current state plus a future, or a representation of an outcome, then infer an action that connects them. It may predict Cartesian pose commands, joint targets, motor commands, or an action distribution rather than physical joint torque.
The common idea is direction: outcome or motion first, action or effort second. The inputs, outputs, and physical meaning must still be stated.
Inverse dynamics is not forward dynamics
Forward dynamics begins with the robot state and applied forces or torques, then predicts acceleration or a later state. It answers “what happens if this action is applied?” Inverse dynamics asks “what action or effort is associated with this requested change?”
A learned world model and a learned inverse-dynamics model can therefore be paired. The world model predicts a future outcome; the inverse model maps that outcome back to an executable action. ForceWAM uses this structure for contact-rich manipulation: its inverse model learns actions from recorded visual-wrench futures in both task demonstrations and exploratory play.
A learned inverse model does not guarantee feasibility
The inferred action is conditional on the data and representation used to train the model. A visually similar future can require different forces, and an action valid for one robot may violate another robot's joint, torque, contact, or timing constraints.
Dataset documentation should identify the action space, coordinate frame, control mode, prediction horizon, sampling rate, current-state inputs, future or goal representation, and whether the target action was commanded, measured, optimized, or inferred. Evaluation should separate reconstruction of logged actions from closed-loop task execution on the target system.
Sources
Related terms
Hardware & control
Robot dynamics
Robot dynamics describes the relationship between forces and torques acting on a robot and the motion they produce, accounting for mass, inertia, gravity, velocity-dependent effects and external contact. Forward dynamics predicts acceleration from applied forces; inverse dynamics computes forces or torques for a specified motion.
Hardware & control
Forward kinematics
Forward kinematics computes the pose of robot links or end effectors from a specified robot configuration and kinematic model. It propagates known joint transformations through the kinematic chain; it does not determine the forces, torques, or motion that produced that configuration.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.
Data & collection
Trajectory
A trajectory is a time-ordered sequence of states or observations, actions and, where applicable, rewards generated as an agent or robot evolves. A complete episode or policy rollout often yields a trajectory, but the terms are not universally identical: trajectories may be partial, while episodes have dataset- or environment-defined boundaries.
Models & learning
World-action model
A world-action model is an emerging, non-standardised name for a robot model that couples prediction of future world states with generation of executable actions, often by adapting a video or world-model backbone as a policy. It differs from an action-conditioned world model that only predicts consequences of supplied actions, and from Barrett’s WAM, the Whole Arm Manipulator.