Overview
Humanoid motion-tracking policies can execute a wide variety of individual behaviors, but linking independently captured movements into uninterrupted sequences presents substantial operational hurdles. To address this, researchers developed Spatio-Temporal Alignment with a Motion Prior (STAMP). The framework implements a Robot-to-Human-to-Robot pipeline that leverages human motion priors to generate smooth behavioral transitions, allowing motion transition knowledge to be reused across different robot embodiments.
Research Context
Connecting discrete, independently acquired motions into continuous streams typically relies on basic interpolation or generative motion priors, both of which exhibit notable failure modes:
- Interpolation: Blends boundary poses directly, but fails to evaluate whether the relative spatial positioning, heading, and temporal intervals create feasible conditions for a natural physical transition.
- Generative Motion Priors: Capable of synthesizing richer movements, but become unstable when boundaries are misaligned, often generating unintended behaviors such as excessive locomotion, erratic turning, or other unprompted actions.
Approach
STAMP operates within human motion space to exploit pre-existing human motion priors and enable transferability across diverse humanoid robot designs. The framework processes transitions through defined steps:
- Retargeting to Human Rig: The two separate robot motions designated for connection are first retargeted onto a standardized human skeletal rig.
- Learned Spatio-Temporal Alignment: A learned alignment model evaluates and ranks candidate positions, headings, and transition durations for the subsequent motion to establish optimal boundary conditions for inbetweening.
- Transition Synthesis: Operating under the selected spatio-temporal alignment, a human motion prior synthesizes the inbetweening motion trajectory.
- Robot Retargeting and Policy Training: The completed motion sequence is retargeted back to the target robotic embodiment, where it serves as a reference trajectory to train a humanoid motion-tracking policy.
Findings
The framework was experimentally validated across multiple platforms:
- Tracking Success: Transition tracking performance was tested on the Unitree G1 and ROBOTIS K1 humanoid platforms.
- Physical Deployments: Physical robot demonstrations showed that STAMP successfully joins independently acquired motion segments into continuous, long-horizon behavioral sequences.