RAX scales robot exploration, not demos: zero teleop, +37 points on benchmarks
du_yilun · x · 2026-10-17
RAX is an agentic robotics system that learns by scaling online exploration rather than demonstration data or model size. A coding agent connects directly to the robot's perception and control APIs, learns from feedback, and builds programs that perceive, plan, and act — with the model kept frozen. As exploration scales, the agent improves its programs and accumulates reusable skills. Results: zero teleoperation or action demonstrations on real robots, from insertion tasks with 0.1mm clearance and air-fryer manipulation to deformable objects and long-horizon mobile whole-body control. Across five simulation benchmarks (RoboDojo, BEHAVIOR-1K, HumanoidArena, DexJoCo bimanual, LIBERO-Pro), it outperforms leading learned policies by an average of 37 percentage points.
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