Robots crack whips with zero learning: motion capture retargeting enables dynamic manipulation
ATTlKA · x · 2026-09-27
Krishna Suresh and Chris Atkeson show that robots can perform highly dynamic manipulation—like cracking a whip or lassoing a cleat—by simply copying a human demonstration, with no learning at all.
- Teleoperation via grippers (e.g., UMI) struggles with dynamic tasks; lightweight markerless motion capture (GVHMR, HaMeR) captures fast human motion directly
- Key finding: if the robot tracks the demonstrated hand motion well enough, it can execute complex dynamic skills open-loop without any training
- The whole pipeline can be implemented almost automatically via an agentic coding setup the authors call "GPT-6 Astra"—"we just solved the orchestration problem for humanoids"
- Existing retargeting work (e.g., PHP Parkour) mostly covers locomotion; manipulation retargeting remains an open problem, which one author notes awkwardly undercuts a thesis on learning dynamic tasks
Bonus: the whips are signal/stock whips, whose tips break the sound barrier to produce the crack—a popular performance art.
More from Embodied
- VLA-REPLICA benchmark accepted at NeurIPS 2026, adds MolmoAct 2 and GR00T N1.7 — YuXiang_IRVL · 2026-09-27
- Humanoid Robotics Debate: Claims US Firms Fake Demos While Chinese Companies Stay Legit — cixliv · 2026-09-27
- Asimov open-sources humanoid locomotion policy; non-technical user fine-tunes fall recovery on a single 4090 — freelerobot · 2026-09-27
- Rabbit OS3 Update Brings Cross-Device r1 Control, Direct Email Output Without Gmail Linking — SimonBalmain · 2026-09-27
- Tom Riddle Diary experiment runs fully local on iPad Pro M5 with Gemma 4, TTS and handwriting recognition — alexcovo_eth · 2026-09-27
- Awesome-UMI list adds 9 entries spanning CoRL papers and commercial devices — SongShuran · 2026-09-27