RAX drives robots with a frozen coding agent; DeepSeek version beats π0.5 at 1/20 cost
du_yilun · x · 2026-10-22
RAX (Robot Agent eXploration) is an agentic robotics system that learns by scaling online exploration: a frozen LLM coding agent connects directly to a robot's perception and control APIs, fixes its own mistakes from feedback, and compiles what works into reusable skill programs — zero teleoperation, zero demonstrations, no policy training.
New update:
- Supports DeepSeek Harness and Kimi Code as agent harnesses
- Direct support for self-hosted DeepSeek-V4.1-Flash for local inference
Benchmarks (DexJoCo): RAX + DeepSeek-V4.1-Flash hits 66% success on all tasks with just 3 demonstrations, beating π0.5 (54.1%) and Fast-WAM (52.9%) by over 10%; RAX + Claude Opus 5.5 reaches 96.7%. At API pricing the DeepSeek setup costs roughly 1/20 of Claude Opus 5.5 ($0.60–1.20 vs $20.00 per 1M output tokens).
Real-robot feats include insertion with 0.1mm clearance, air-fryer manipulation, deformable objects and long-horizon mobile whole-body control; compiled programs automate a 72-minute agent session down to 18 seconds (241×). Paper, code, skill library and full session logs are open-sourced; the author list includes Pieter Abbeel, Antonio Torralba, Dawn Song and Yilun Du.
Related event: RAX: Frozen-LLM Robot Agent Beats π0.5 at 1/20 the Cost(2 posts)→
More from coding & agent
- Datadog team has been generating TLA+ via skills, eyeing a meta-skill for full suites — tianyin_xu · 2026-10-22
- Agent planned a whole trip solo: perfect logistics, zero vibe — tech__unicorn · 2026-10-22
- "AI swarms that patch the world's vulnerable software for free" — who's building this? — voooooogel · 2026-10-22
- Recursive launches Flint, a laptop coding model largely built by AI — TechCrunch AI · 2026-10-22
- Inside T800: Contrarian Reviews Handled by a Filter Agent Before Dispatch — lucasmeijer · 2026-10-22
- Local storage vs cloud: yaroslavvb argues security is the only real case as agents get cloud Macs — yaroslavvb · 2026-10-22