humanoidsdata.com

Search

Search datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Simulation & transfer

Motion retargeting

Motion retargeting is the adaptation of a recorded or generated motion from one body to another with different proportions, joints or limits. For humanoid robots, it maps source poses or trajectories into robot configurations while preserving task-relevant relationships such as contacts and end-effector paths and satisfying kinematic, balance, collision and actuator constraints.

Also known as: motion retargetting

Updated

From source motion to target body

Directly copying joint angles rarely works when the source human, character or robot has different limb lengths and joint structure. Motion retargeting instead preserves selected properties of the motion while solving for a feasible target trajectory. The foundational Gleicher method treated important spatial and temporal features as constraints when adapting motion between articulated characters.

The source may be motion-capture markers, estimated human poses, animation or another robot's trajectory. The output is usually a sequence of target poses or joint configurations, sometimes accompanied by contact labels or reference velocities.

Humanoid constraints

For a humanoid, visual similarity is insufficient. Feet must make and break contact at plausible times, hands must reach task-relevant locations, the body must avoid self-collision, and joint position, velocity and torque limits must be respected. Balance and contact forces can make a geometrically close pose physically impossible.

An AIST optimisation formulation jointly considered spatial relationships, dynamic balance and torque limits. Online approaches have also enforced joint, self-collision and balance constraints while mapping tracked human motion to a humanoid, as demonstrated in markerless retargeting to ASIMO.

Role in robot learning

Retargeted trajectories can become reference motions for imitation learning, reinforcement learning or whole-body controllers. They are generated data, not evidence that the physical robot can execute the motion unaided.

A controller still has to track the reference under real dynamics and disturbances. Good datasets therefore preserve the source motion, target robot model, coordinate conventions, contacts, optimisation settings and feasibility checks rather than publishing only the final joint sequence.

Sources