✦
Summary
FlashDexRetarget converts collections of human hand-object demonstrations into physically grounded robot-hand trajectories. One reference-conditioned reinforcement-learning policy tracks many motions, sharing training across demonstrations instead of optimizing each one independently.
Key Features
Human-to-robot motion retargeting.
Multi-reference tracking.
A shared RL policy.
Object-motion preservation.
Robot-contact constraints.
Physics-based trajectories.
Collection-level training.
Released research implementation.
About FlashDexRetarget
Robotics developers can generate dexterous manipulation data more efficiently while accounting for the target hand's contacts and dynamics. Core capabilities include Human-to-robot motion retargeting, Multi-reference tracking, A shared RL policy, and Object-motion preservation.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Robot-contact constraints, Physics-based trajectories, Collection-level training, and Released research implementation.