RAX hits 78.3% on RoboDojo vs prior SOTA 24.9%, scores 100% on pouring where baselines get 0%
du_yilun · x · 2026-10-17
Du Yilun's team shows RAX's exploration scaling producing large gains: 78.3% on RoboDojo vs the prior SOTA's 24.9% (+53.4 points), and 46.5% vs 8.2% on BEHAVIOR-1K long-horizon household tasks (+38.3 points).
Task-level numbers: plugging in a charger improves 26%→94% with 4→8 exploration episodes; block swapping 34%→100% with 4→16 episodes; language-conditioned pouring reaches 100% where every prior policy baseline scores 0%; car loading hits 100% vs previous best of 20%, combining perception, navigation, manipulation, and progress tracking.
Context explains how one agent controls different bodies: it works across joint targets, end-effector poses, torso–arm coordination, or whole-body references, with low-level controllers handling high-rate tracking while the agent builds perception/planning/control logic above them.
More from Embodied
- Google Gemma hardware hackathon at AGI House offers $45k prizes, $15k for LiteRT track — agihouse_org · 2026-10-17
- iPhone Duo pre-orders sell out instantly, shipping slips past January 2027 — mitchdeg · 2026-10-17
- P-BFM: unsupervised RL lets Unitree G1 humanoids adapt to rough terrain from flat-ground references — ChongZzZhang · 2026-10-17
- Researcher builds generative machine vision pipeline reconstructing scenes from monocular video — RexDouglass · 2026-10-17
- Starkey Launches Omega AI+ Hearing Aid With Quad DNN Architecture and Sleep Tracking — BrandonSawalich · 2026-10-17
- RAX open-sources everything: zero demos, all experiment sessions included — du_yilun · 2026-10-17