Artificial intelligence
Motion prior
Definition
A motion prior is a learned or designed model of which body motions are plausible before a specific control task is considered. In robot learning, it can guide a policy towards coordinated motions represented in human or robot demonstration data.
Also known as: Motion priors
Updated
Prior knowledge about movement
A motion prior assigns greater preference or probability to some motion sequences than to others. It can be learned from motion capture, robot trajectories, or other demonstrations. The prior describes patterns in the data, such as coordinated stepping or recovery motions, without necessarily specifying the task the robot must complete.
Adversarial Motion Priors train a discriminator on unstructured reference clips and turn its output into a style reward for reinforcement learning. This separates a task reward, such as moving towards a target, from a learned preference for motions that resemble the reference data. Other motion priors can be probabilistic generative models, latent representations, or explicit dynamical models.
Use in humanoid control
A humanoid policy can use a motion prior to reduce the search over poorly coordinated joint movements. It may help retain humanlike coordination while a task controller chooses where to walk, reach, or make contact. The 2026 iGPC study reports extending a pretrained generative motion prior with interaction experts, then distilling those skills into a student driven by onboard observations.
This differs from motion retargeting. Retargeting converts a particular source motion to another body, while a motion prior represents regularities across motions and can influence newly generated behaviour. It also differs from exact reference tracking: a prior may allow several plausible motions for the same task.
The data constrains the prior
A motion prior is not a physical guarantee. Reference clips may contain movements that a robot cannot execute because of different proportions, actuator limits, contacts, or balance constraints. A prior trained on a narrow dataset can also penalise useful motions that are absent from that dataset.
Physical feasibility still depends on the robot model, contacts, controller, and hardware. Results reported for simulated characters or a particular humanoid robot demonstrate that implementation and test setting, not every system that uses a motion prior.
Sources
Related terms
Imitation learning
Imitation learning learns behavior from examples supplied by a demonstrator. In robotics, demonstrations can teach a policy how to perform a task without requiring every action or objective to be programmed by hand.
Reinforcement learning
Reinforcement learning trains an agent to choose actions that maximize expected cumulative reward through experience with an environment. In robotics, the learned policy can select movements or higher-level behaviors from observations.
Motion retargeting
Motion retargeting maps a motion recorded or designed for one body onto another body with different geometry or joints. In robotics it produces a compatible pose or trajectory reference, which still needs a controller to execute it physically.
Humanoid robot
A humanoid robot is a robot with a body arranged to resemble the human form, usually with a torso, arms, and legs. The term describes its physical form and does not by itself establish human-level intelligence or general autonomy.