LATENT wins IROS 2026 award: humanoid robots rally at human level from imperfect motion data
chris_j_paxton · x · 2026-10-01
Galbot's LATENT system won the IROS 2026 Best Entertainment and Amusement Paper Award, enabling a humanoid robot to sustain human-level tennis rallies.
The key insight: instead of needing perfect, complete human tennis motion captures, LATENT only uses motion fragments capturing primitive skills (like backswings), which dramatically eases data collection. It then:
- Learns a latent action space from this imperfect data
- Trains a high-level policy in simulation that composes these primitives to execute tennis skills on a real robot
Authors Haofei Lu and Yunrui Lian discussed the method on RoboPapers Ep#80. Project page, arXiv paper, and code are all public.
More from Embodied
- Microsoft teases 'something new is coming from Windows' ahead of Oct 7 Surface event — tomwarren · 2026-10-02
- Asimov unveils Asimov 1, a humanoid robot built as a platform for custom solutions — victormustar · 2026-10-02
- Autonomous's open-source desk robot Lamp ships this month at $499 — dee_hw · 2026-10-02
- Boston Dynamics shows off new hands for its Atlas humanoid robot — Recoil42 · 2026-10-02
- Huawei scales Tau chip roll-out as Mate XT 2 nears 1 million sales — pstAsiatech · 2026-10-02
- Roboticist Argues Real-World Robot Performance Is Never Dominated by Numerics — ZeYanjie · 2026-10-01