Tsinghua humanoid robot plays badminton with one policy from just 30 min of human motion data
ChongZzZhang · x · 2026-09-29
Humanoid Badminton, from Tsinghua, CUHK and Hong Kong Embodied AI Lab, accepted at CoRL 2026, trains a humanoid to play badminton from only 30 minutes of human motion data.
- Method: a three-stage hierarchical RL framework. Task-randomized motion augmentation expands sparse annotated hitting events into diverse executable strokes forming a latent skill space; a high-level planner composes skills online from the shuttle state; a context-conditioned adversarial regularizer keeps motion natural.
- Results: a single policy executes forehand, backhand and highly dynamic jump returns, sustaining multi-skill human-robot rallies with natural whole-body motion.
- Project page and arXiv paper are public.
More from Embodied
- NOOP open-source app makes WHOOP band work locally without subscription — mhdfaran · 2026-09-29
- Recursive Harness Distillation lifts robot manipulation success from 37.3% to 64.0% — SeoulNatlUniv · 2026-09-29
- Bose's $99 USB-C wired ANC earbuds ditch the battery brick by drawing power from your phone — aakashgupta · 2026-09-29
- Unitree's Dex5-1 dexterous hand: 22 DoF, all backdrivable, from $6,500 — lukas_m_ziegler · 2026-09-29
- Niantic opens 100 real-world 3D scenes for robot sim; Vesoma stealth-launches humanoid Kyle — lukas_m_ziegler · 2026-09-29
- Ex-DeepMind director's startup Vesoma unveils humanoid Kyle, robot due 2027 — lukas_m_ziegler · 2026-09-29