Motion retargeting transfers a time-based movement from one skeleton or body to another. Good retargeting preserves the action's timing, trajectory, pose intent, end-effector placement, and contacts while adapting the motion to the target body's hierarchy, proportions, joint axes, and constraints.
It is not a rotation-copy operation. If the source has longer arms, shorter legs, different local axes, another rest pose, or a different number of joints, identical rotations will not place the target's hands and feet in the same locations.
For a 3D character, the goal is usable animation. For a humanoid robot, the target also has embodiment-specific degrees of freedom, joint limits, link geometry, balance, and dynamics. A kinematically retargeted robot sequence is a reference — not proof that the hardware can execute it.
Retargeting starts with a source motion and ends with a target motion. The source contains a hierarchy of joints changing over time. It may also contain global root motion, joint positions, joint rotations, contact information, and metadata. The target has its own hierarchy and coordinate system. A retargeter must establish semantic correspondence, decide which features matter, and solve for a new sequence on the target body.
A practical mental model has three layers:
“Motion intent” is not a mystical representation. It is the set of constraints and objectives the retargeter chooses to preserve: foot positions during stance, hand trajectories, pelvis path, facing direction, joint orientation, timing, self-contact, or style.
A skeleton is a tree of parent-child relationships. Rotating the pelvis changes every descendant. Rotating an upper arm changes the forearm and hand. Two rigs can represent the same body with different intermediate bones, twist joints, control bones, or spine segmentation.
Names are not structure. “LeftArm,” “upper_arm_l,” and “L_shoulder” may refer to comparable joints. Two bones with identical names can still have different parents or local axes.
Joint rotations are defined relative to a rest or bind pose and a local coordinate frame. One rig may start in a T-pose, another in an A-pose, and another with bent knees or stylized shoulders. The local X, Y, and Z axes may point in different directions.
Copying a quaternion or Euler rotation without reconciling those frames can rotate a limb around the wrong axis or introduce a constant pose offset.
An arm reaching the source character's hip may reach the target character's waist, thigh, or empty space. Joint angles describe configuration; they do not guarantee end-effector location across different limb lengths.
This is the source of many visible errors. A retargeted foot can have the “correct” orientation while no longer touching the floor. A hand can follow the source arm rotations while missing the other hand.
Some skeletons include clavicles, twist bones, multiple spine joints, detailed fingers, or facial joints. Others reduce the body to a smaller set. Robots may constrain a joint to one rotational degree of freedom where a human anatomical joint is approximated with three.
The source can store global movement on a root joint, pelvis, separate trajectory channel, or engine-specific object transform. Forward direction, up axis, scale, and units can also differ. If root motion is mishandled, the character may animate in place, drift away from the intended path, rotate around the wrong pivot, or accumulate an offset over time.
Forward kinematics, or FK, starts with joint rotations and propagates them down the hierarchy to calculate every joint's world-space position and orientation. It answers: given these joint rotations, where does the hand end up?
Direct rotation transfer is largely an FK workflow. Map a source joint to a target joint, account for rest-pose and axis differences, and apply the adapted rotation. It is fast and predictable when the rigs are similar. It becomes fragile as morphology and hierarchy diverge.
Inverse kinematics, or IK, starts with a desired spatial result and solves for the joint configuration that produces it. It answers: which shoulder, elbow, and wrist rotations place the hand here?
IK is useful for preserving end-effectors and contacts. A planted foot can remain near a fixed point while the leg joints adapt to a shorter or taller target. But IK still requires choices. Which joints can move? Which target matters more when constraints conflict? How should unreachable targets be handled? How much should the target pose deviate from the source?
A control rig exposes animator-facing controls above the raw skeleton. Controls can represent hands, feet, hips, chest, gaze, pole vectors, and higher-level relationships. Retargeting through a control rig can preserve meaningful spatial behavior and produce animation that remains editable in the destination workflow.
Quality is multidimensional. A visually smooth result can still be wrong.
What it looks like: a planted foot moves across the floor during stance.
Why it happens: the retargeter preserves joint rotations or root motion without enforcing the target foot's world-space contact. Different leg lengths change where the foot lands.
What to inspect: track the heel and toe in world space during the contact interval. Separate true source movement from target drift.
What it looks like: feet hover above the floor or pass through it.
Why it happens: the target's leg length, root height, foot geometry, or ground-plane convention differs from the source.
What it looks like: clasped hands separate, crossed arms open, or a hand misses the torso.
Why it happens: rotation transfer preserves local joint orientation but not the world-space relationship between two body parts with different proportions.
What it looks like: a limb snaps, twists, or rotates through an unnatural path.
Why it happens: local-axis mismatch, Euler-angle discontinuity, rest-pose error, ambiguous IK, or a solver switching between equivalent solutions.
Jitter comes from noisy source data, independent frame solves, unstable correspondence, or conflicting constraints. Root drift appears when root translation is omitted, rescaled incorrectly, integrated inconsistently, or transformed into another coordinate convention. Mesh intersection occurs because skeleton-only retargeting does not automatically reason about the target mesh or collision volumes — a valid joint configuration can still produce invalid geometry.
Different methods optimize different parts of the problem. This is a taxonomy, not a winner chart.
| Method | Mechanism | Strength | Tradeoff |
|---|---|---|---|
| Direct bone mapping | Map corresponding joints and adapt rotations between rest poses and axes | Fast, understandable, effective for similar rigs | Fragile under hierarchy, proportion, and contact differences |
| IK or control-rig retargeting | Solve target joints to place end-effectors or controls while staying near the source pose | Better spatial preservation across different bodies | Requires objectives, weights, limits, and unreachable-target behavior |
| Optimization or contact-aware | Optimize a sequence against pose, contact, smoothness, collision, or other objectives | Can preserve task-critical constraints across time | More complex; outcome depends on the objective and target model |
| Learned retargeting | Train a model to transform motion across skeletons, morphologies, or representations | Can learn nonlinear corrections and broad mappings | Generalization depends on training data, representation, and target coverage |
| Mesh-aware retargeting | Optimize relationships defined on or around the target surface | Can address body-shape and self-contact errors not visible in the skeleton alone | Requires mesh correspondence or geometry-aware machinery |
| Robot-specific retargeting | Solve human motion into target robot joints under embodiment constraints | Produces a robot-specific reference for simulation or tracking | Kinematic validity does not guarantee dynamic execution |
Kinetix publicly describes a mesh-aware method that places key reference points on a character's surface rather than reasoning only about bones. That is one valid approach to morphology and contact.
In robotics, General Motion Retargeting uses a multi-objective IK formulation and evaluates how retargeting artifacts affect downstream humanoid tracking. The paper's central practical point is broader than its method: bad reference motion creates a harder control-learning problem.
Use a repeatable test set. Do not rely on a single walk cycle or a highlight reel.
| Dimension | What to inspect | Useful measurement |
|---|---|---|
| Semantic fidelity | Does the target still perform the same action and style? | Expert rating with defined rubric; event alignment |
| Ground contact | Sliding, floating, penetration, stance stability | Heel/toe drift during labeled contact; minimum ground distance |
| Self-contact | Hands, arms, torso, and other body relationships | Distance between defined contact points during contact intervals |
| Root trajectory | Path, facing, speed, displacement, turns | Position/orientation error after scale and coordinate alignment |
| Joint plausibility | Flips, extreme bends, discontinuities, limit violations | Limit violations; angular velocity and acceleration outliers |
| Temporal quality | Jitter, pops, frame-to-frame instability | Smoothness metrics plus visual review at full and slow speed |
| Morphology generalization | Quality across realistic and stylized proportions | Same test suite on short, tall, wide, exaggerated, asymmetric bodies |
| Automation | Setup, mapping, manual corrections, failures | Operator minutes, reruns, rejected assets, throughput |
| Production usability | Editability, export, target-engine behavior | Import success, curve/keyframe access, root-motion behavior |
| Robot feasibility | Limits, collision, contact, balance, tracking | Simulator validation and downstream policy/control metrics |
Include at least: straight locomotion with stops and turns; a jump with takeoff and landing; crossed arms or clasped hands; a hand reaching a known target; a floor or seated transition; fast athletic movement; a long clip that can expose drift; several target morphologies; one skeleton with a different hierarchy and naming convention; and one humanoid robot target if robotics is in scope.
Both problems transfer movement across bodies. Their acceptance criteria differ.
| Character target | Humanoid robot target |
|---|---|
| Skeleton and mesh designed for visual performance | Mechanical links, joints, actuators, sensors, collision geometry |
| Quality judged by action fidelity, deformation, contacts, editability | Quality judged by reference fidelity plus feasibility and downstream tracking |
| Joint ranges may be artistic or permissive | Joint limits and degrees of freedom are explicit hardware constraints |
| Mesh penetration may be fixed by animation or rigging | Self-collision can invalidate the trajectory or damage hardware |
| Root motion can be consumed by an engine controller | Floating-base motion and balance interact with physics and control |
| Produces editable character animation, currently downloaded as FBX or GLB | Produces a target-specific robot representation; the public Unitree G1 path returns CSV, while other robot outputs are scoped by engagement |
For robots, retargeting should document what is constrained. Does the solve enforce joint limits? Does it model self-collision? Are contacts preserved? Is the center of mass considered? Are velocities or accelerations bounded? Has the output been validated in a simulator? Has it been executed on hardware?
If the answer is “not included,” that is not necessarily a defect. It defines the boundary between a kinematic motion reference and a dynamically validated trajectory.
Uthana applies a generated or existing motion to a supported bipedal character without requiring the source and target rigs to share identical bone names or proportions.
The public workflow is:
Robot retargeting changes the target. The output must match a robot's joint structure and downstream execution convention rather than a character-animation rig.
Uthana's public Unitree G1 path converts a motion into a CSV robot trajectory. That export uses raw authenticated HTTP and is not yet included in the official Python or JavaScript clients. Other robot-description formats, outputs, constraints, and simulator packaging are defined through scoped engagements.
A retargeted robot trajectory is not automatically a dynamically feasible policy or safe hardware command. Controller integration and hardware validation remain separate unless included in the engagement.
Motion retargeting transfers an animation or time-based body movement from one skeleton to another. It adapts the source motion to the target's hierarchy, rest pose, coordinate frames, proportions, joints, and constraints while preserving selected properties such as timing, trajectory, end-effector paths, and contacts.
Joint rotations are defined relative to a skeleton's hierarchy, rest pose, and local axes. Different limb lengths also change where hands and feet land. Direct copying can work for nearly identical rigs but often causes pose offsets, missed contacts, sliding, and axis errors across different bodies.
IK retargeting uses inverse kinematics to solve for target-joint configurations that place hands, feet, or other controls at desired spatial locations. It can preserve end-effector placement across proportion differences more effectively than rotation transfer alone.
Foot sliding occurs when the target foot moves during a source contact interval. Common causes include different leg lengths, incorrect root motion, missing contact constraints, ground-plane mismatch, and a solver that prioritizes joint rotation over world-space foot position.
Uthana's character retargeting targets bipedal skeletons and does not require manual bone mapping, chain setup, or matching names. Non-bipedal characters, unusual appendages, incomplete hierarchies, mesh or skeleton errors, and workflow-specific requirements should be tested on your own assets before committing a pipeline to them.
Yes. Uthana exports FBX and GLB animation with editable keyframes. Confirm the exact format, skeleton, curve, and root-motion behavior required by the destination DCC or engine.
No. Robot retargeting produces a motion reference in the target embodiment. A controller or learned policy must still track that reference while satisfying dynamics, contact, balance, actuator, and hardware constraints.
Yes. Motions from different source skeletons can be mapped to one target representation before training or evaluation. The transformation must be documented because retargeting can change joint distributions, contacts, trajectories, and motion quality.
Apply motion to production characters without rebuilding bone maps by hand. For robot embodiments, work with Uthana on a target-specific representation and delivery specification.
Proprietary IK retargeting that adapts motion to any bipedal skeleton automatically.
Explore character retargetingTransfer human motion into joint representations built for a target robot embodiment.
Explore robot retargetingTurn an unrigged bipedal mesh into a motion-ready body, typically in 30–60 seconds.
Explore auto-riggingStart generating motion today, or work with us on a custom dataset, embodiment system, or simulation program.
