Robot Motion Transition with Human Motion Prior

I. INTRODUCTION

• Independent clips lack a defined relative position, heading, and transition time, and arbitrary arrangements can demand excessive locomotion, turning, or unintended actions from a motion prior.Boundary placement and timing determine the body movement required between the motions.Source: §I · p. 1
• Pose interpolation ignores incoming and outgoing motion and foot-contact changes, while a support-foot change may require a step or body-weight transfer.Returning to a predefined pose instead interrupts motion flow; human motion data offer examples of body movements and contact changes for inbetweening.Source: §I · p. 1
• STAMP connects the end of the preceding motion to the start of the following motion while preserving both input clips, unlike approaches that may truncate or alter constituent motions.It uses a Robot-to-Human-to-Robot framework to reuse human motion knowledge across robot embodiments.Source: §I · p. 1
• A learned alignment model ranks candidate positions, headings, and transition durations before a human motion prior generates the transition.The model is learned from continuous human motions and counterfactual boundary arrangements; the completed sequence is retargeted and used to train a robot-specific tracking policy.Source: §I · pp. 1–2
• STAMP evaluates tracking success on Unitree G1 and ROBOTIS K1 against interpolation and generative motion baselines, and demonstrates long-horizon compositions on ROBOTIS K1.The evaluation also includes a ranking ablation.Source: §I · p. 2

A. Humanoid Motion Tracking and Transition

• Prior humanoid transition methods use compatible-state selection or policy blending, whereas STAMP selects spatial and temporal boundary conditions and generates the connection with a human motion prior.The cited approaches include skill graphs, motion matching, and blending expert policies.Source: A. Humanoid Motion Tracking and… · p. 2

B. Human Motion Inbetweening

• Human motion inbetweening methods generate missing intervals from boundary contexts, but continuous-motion boundaries already have defined placement and timing, unlike independently acquired motions.The approaches described include recurrent networks, Transformers, phase representations, diffusion models, and temporally composed text-conditioned motion.Source: B. Human Motion Inbetweening · p. 2

C. Human and Robot Motion Retargeting

• Retargeting transfers motion across bodies with different joint structures and proportions; STAMP adapts CoRe for bidirectional transfer between robot motions and human motion space.Prior approaches include geometric retargeting, learned shared or cross-embodiment representations, and normalized limb-direction methods.Source: C. Human and Robot Motion Retargeting · p. 2

III. PROBLEM FORMULATION

• The problem is to determine the spatial and temporal relationship between independently specified motions and generate a connecting transition that preserves their content and order.The transition should provide continuous body movement without a return to a neutral pose or an unintended action, and the completed sequence serves as a tracking reference.Source: §III · p. 2

IV. PROPOSED METHOD

• STAMP transfers robot motions to a human rig, selects a low-cost alignment, generates a human-space transition, and retargets the sequence as a robot-tracking reference.The framework performs alignment and generation in human motion space to reuse human motion knowledge across robot embodiments.Source: §IV · p. 2, Figure 1 · p. 3

A. Spatial and Temporal Motion Alignment

• Alignment candidates specify the following motion’s planar position and heading relative to the preceding motion’s final root frame, plus transition duration in seconds.The candidate applies the same rigid transformation to the complete following motion.Source: A. Spatial and Temporal Motion Alignment · pp. 2–3
• The alignment model uses preceding and transformed following boundary contexts, including motion features and relative foot configuration, to predict a scalar transition cost.Separate temporal Transformer encoders allow the model to consider incoming and outgoing dynamics beyond endpoint poses.Source: A. Spatial and Temporal Motion Alignment · p. 3
• Kinematic costs use a complete boundary-inclusive sequence, with foot-dynamics terms accounting for foot and toe-base velocities and soft contact weights.Candidate durations correspond to 15, 30, 45, or 60 frames at 30 fps; the observed transition is used for supervision, not as an alignment-model input.Source: A. Spatial and Temporal Motion Alignment · pp. 3–4
• The target transition cost is a kinematic proxy rather than a direct estimate of generated-transition cost, and its normalization scales and weights were chosen heuristically.They were not tuned using motion-inbetweening outputs, policy rewards, or robot-tracking results.Source: A. Spatial and Temporal Motion Alignment · p. 4

B. Particle-Based Inference

• The fixed alignment-particle set combines predefined planar positions, headings, and transition durations relative to the preceding motion’s final root frame.Candidate values are specified in Sec. V-A.Source: B. Particle-Based Inference · p. 4
• The alignment model evaluates all particles without generating their corresponding transitions.Source: B. Particle-Based Inference · p. 4
• STAMP supplies the selected placement, heading, and duration to the human motion inbetweening model.Source: B. Particle-Based Inference · p. 4

C. Human Motion Inbetweening

• During denoising, one Transformer predicts root position and heading, and a second predicts articulated body motion conditioned on the root representation and boundary contexts.The generated intermediate frames form the human transition.Source: C. Human Motion Inbetweening · p. 4
• Training windows use the first and last K frames as observed conditions and the intermediate frames as generation targets, with reconstruction losses applied only to unobserved transition frames.Transitions are resampled to Nτ = round(fsτ) frames, with fs = 30 fps.Source: C. Human Motion Inbetweening · pp. 4–5
• The inbetweening and alignment models are trained separately, allowing another compatible inbetweening model to use the selected alignment without retraining the alignment network.Source: C. Human Motion Inbetweening · p. 5

D. Robot-to-Human-to-Robot Motion Retargeting

• STAMP implements robot-to-human and human-to-robot retargeting by adapting CoRe.Retargeting connects the robot-specific representation with the human motion space used for alignment and generation.Source: D. Robot-to-Human-to-Robot Motion… · p. 5
• Input retargeting adapts root displacement and foot placement to human proportions while retaining motion content, heading, and support-foot sequence.Human joints without robot correspondences remain in temporally consistent configurations.Source: D. Robot-to-Human-to-Robot Motion… · p. 5
• Human-to-robot retargeting accounts for target proportions, joint structure and limits, and foot contacts, transferring the connected motion as one sequence for boundary consistency.The resulting robot motion is the reference for motion-tracking policy training and execution.Source: D. Robot-to-Human-to-Robot Motion… · p. 5

• STAMP is compared with interpolation and inbetweening baselines on Unitree G1 and ROBOTIS K1 under the same tracking-policy training protocol.Source: §V · p. 5
• The experiments evaluate alignment selection in addition to the baseline comparisons.Source: §V · p. 5
• A physical ROBOTIS K1 demonstration composes motions over a long horizon.Source: §V · p. 5

A. Experimental Setup

• The benchmark contains 18 motion pairs across boxing, industrial maintenance, and security and traffic control, with varied locomotion, heading, stance, foot placement, and upper-body changes.Source: A. Experimental Setup · p. 5
• Each constituent motion is generated independently with Kimodo from a prompt describing one motion, then retargeted to each robot; all methods use the same input pairs.Source: A. Experimental Setup · p. 5
• Root/heading interpolation matches the next motion’s initial root position and heading to the preceding motion’s final root frame, then quintically interpolates boundary states.Source: A. Experimental Setup · p. 5
• Kimodo† uses Kimodo’s representation and denoising architecture for inbetweening, while retaining the original boundary arrangement; MotionBricks is evaluated only on Unitree G1.Source: A. Experimental Setup · p. 5
• STAMP selects the following motion’s position, heading, and duration before generation, whereas interpolation and Kimodo† use a fixed 1.0 s duration.Source: A. Experimental Setup · p. 5
• STAMP ranks 388 candidate alignments per pair, formed from planar offsets, headings, and durations plus identity placements; the ablation uses the rank-100 candidate.Source: A. Experimental Setup · p. 5
• Each reference is tracked by an independently trained BeyondMimic policy [3], with identical policy architecture, observations, rewards, termination conditions, and randomization schedules across methods.Training uses one seed per reference and evaluates the final checkpoint.Source: A. Experimental Setup · pp. 5–6

B. Evaluation Metrics

• A rollout succeeds only if it reaches the final reference frame without tracking termination and stays within the stated root-position, root-orientation, and end-effector error thresholds throughout.Success rate is the percentage of successful rollouts.Source: B. Evaluation Metrics · p. 6
• Mean body velocity error is the Euclidean difference between simulated and reference world-frame linear velocities across 30 tracked body points, including the root.Evaluation averages over valid timesteps and includes successful and failed episodes across the 18 motion pairs.Source: B. Evaluation Metrics · p. 6
• Joint acceleration measures mean absolute acceleration of executed actuated joints; lower values indicate smaller velocity changes, but do not alone establish accurate tracking.The floating root is excluded, and the metric is interpreted with success rate and velocity error.Source: B. Evaluation Metrics · p. 6
• Velocity error and joint acceleration are evaluated over the transition clip, including short preceding and following boundary contexts.Source: B. Evaluation Metrics · p. 6

C. Pairwise Motion Transition

• Kimodo† comparison holds the motion prior fixed to assess boundary selection, whereas MotionBricks compares complete systems; tracking scores measure trackability, not shared-ground-truth similarity.Source: C. Pairwise Motion Transition · p. 6
• 80.69% on Unitree G1 and 87.52% on ROBOTIS K1 were STAMP’s highest transition success rates; it also had the lowest body velocity error and mean executed joint acceleration.The results cover 18 motion pairs; both robots use the same human inbetweening and alignment models, with robot-specific retargeting and tracking policies.Source: Table II · p. 6, Table III · p. 6, C. Pairwise Motion Transition · p. 6

D. Alignment Ranking Ablation

• Rank 1 exceeded rank 100 in success rate by 15.62 percentage points on ROBOTIS K1 and 9.36 percentage points on Unitree G1, with lower body velocity error and executed joint acceleration on both.The comparison keeps the candidate set, generator, retargeting, and tracking protocols fixed; only selected position, heading, and duration differ.Source: D. Alignment Ranking Ablation · p. 6, Table IV · p. 7

E. Physical Robot Evaluation

• For running-to-punching on ROBOTIS K1, STAMP places the punch farther forward and generates a connecting step for continued forward movement; interpolation showed foot sliding as the robot slowed.Source: E. Physical Robot Evaluation · pp. 6–7, Figure 2 · p. 7
• For the look-around-to-stop transition, STAMP had lower mean RMS motor current and generally smaller execution variation late in the transition, but current levels differed at onset.Both methods used 1.5-second transitions; responses summarize five executions across 23 motors in driver-reported units.Source: E. Physical Robot Evaluation · p. 7, Figure 3 · p. 7
• Three physical ROBOTIS K1 compositions of the same four traffic-control motions demonstrate pairwise alignment and transition generation in longer sequences.At each connection, STAMP selects the candidate with the lowest predicted transition cost and generates the connecting motion.Source: E. Physical Robot Evaluation · p. 7

F. Limitations

• The learned alignment cost is a kinematic guide, while downstream trackability also depends on retargeting quality and robot dynamics.Source: F. Limitations · p. 7
• STAMP currently requires a tracking policy to be trained for each composed sequence; integrating alignment and inbetweening into a shared reinforcement-learning framework is future work.Source: F. Limitations · p. 7

VI. CONCLUSION

• STAMP selects relative position, heading, and transition duration in human motion space before a human motion prior generates the transition and the completed sequence is retargeted for robot-specific tracking.The alignment model and motion prior are reused across robot embodiments.Source: §VI · p. 7
• Experiments on Unitree G1 and ROBOTIS K1 show higher tracking success and lower body velocity errors than evaluated baselines; ROBOTIS K1 demonstrations compose the same four traffic-control motions in multiple sequences.The conclusion also reports that a traffic-control case study has lower motor-current RMS than interpolation in the latter part of the transition.Source: §VI · p. 7, Figure 4 · p. 8