RAX drives robots with a frozen LLM via pure exploration at 1/20th Opus cost
du_yilun · x · 2026-10-22
RAX (Robot-Agent-Exploration) lets a frozen coding agent connect to a robot's perception and control APIs, learning from feedback and compiling exploration into reusable programs and skills — no teleoperation or demonstrations needed. Highlights: zero-demo real-robot tasks including 0.1mm-clearance insertion and deformable objects; exploration-to-code speeds up tasks 241x; now supports DeepSeek Harness, Kimi Code, and self-hosted DeepSeek-V4.1-Flash. On DexJoCo, RAX + DeepSeek-V4.1-Flash beats π0.5 and Fast-WAM by 10%+ with just 3 demos, at roughly 1/20th the API cost of Claude Opus 5.5. Paper, code, skills and full session logs are open-sourced.
Related event: RAX Drives Robots via Frozen LLM Exploration at 1/20 the Cost(3 posts)→
More from Embodied
- Tesla FSD hits 1.74M subscriptions, Robotaxi unsupervised miles grow 8x in a quarter — XFreeze · 2026-10-22
- CMU paper defines solver-sufficient symbols to cut hybrid robot planning cost — Yiyuan Pan · 2026-10-22
- XDOF Benchmark: Top LLM Agents Score Just 9-19% on Physical Tasks — mihdalal · 2026-10-22
- Samsung Next, NVIDIA and CoreWeave to host robotics founder night at CoRL 2026 in Austin — wandb · 2026-10-22
- Generalist AI's GEN-1.5 Robot Packs 12 Hockey Pucks with Self-Correction — mihdalal · 2026-10-22
- NVIDIA releases Cosmos Simulation: real-time interactive world models for Physical AI — CSProfKGD · 2026-10-22