Real LEAP Hand standard cylinder demo
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/real_standard_cylinder_success_web.mp4
A proprioceptive in-hand cylinder rotation project on the low-cost LEAP Hand. The system adapts the HORA/RMA two-stage idea into an 8D skill-prior representation: a privileged teacher first learns stable rotation in Isaac Lab, then a deployable policy recovers the same prior from proprioceptive history only. A refined second stage adds teacher-deployment action consistency and wider center-of-mass coverage, improving OOD success from 0.8750 to 0.9141 while keeping ID success at 1.0000. The same Deploy-Refined checkpoint transfers to real hardware, reaching 10/10 trials on standard and smaller cylinders, 8/10 on a slender bottle, and exposing ribbed bottles as the current contact-geometry boundary.
Simulation skill learning
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/sim_teacher_success.mp4
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/sim_deploy_base_success.mp4
OOD adaptation and boundary
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/sim_deploy_base_high_com_failure.mp4
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/sim_deploy_refined_high_com_success.mp4
Real hardware deployment
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/real_small_cylinder_success_web.mp4
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/real_slender_bottle_success_web.mp4
Real hardware limitation
Video pending
/assets/projects/leap-hand-inhand-rotation/videos/real_ribbed_bottle_failure_web.mp4