RAX Uses a Frozen Coding Agent to Drive Robots: 96.7% on DexJoCo at 1/20 the Cost
du_yilun · x · 2026-10-22
Researchers behind RAX present 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, learns from feedback, and builds reusable programs and skills — no teleoperation, no policy training.
Highlights:
- Zero-demonstration real-robot tasks: insertion with 0.1 mm clearance, air-fryer manipulation, deformable objects, long-horizon mobile whole-body control
- Exploration compiles into code: 72-minute agent exploration on the air fryer becomes an 18-second automated program (241× faster)
- Now supports DeepSeek Harness and Kimi Code, plus self-hosted DeepSeek-V4.1-Flash; on DexJoCo, RAX + DeepSeek beats π0.5 and Fast-WAM by 10%+ with just 3 demonstrations, at 1/20 the API cost of Claude Opus 5.5
- All session logs (transcripts, tool calls, camera footage) are open on Hugging Face
Related event: RAX: Frozen-LLM Robot Agent Beats π0.5 at 1/20 the Cost(2 posts)→
More from coding & agent
- How to Build an AI Agent Harness from Scratch in TypeScript: freeCodeCamp's deep tutorial — adnan_hashmi · 2026-10-22
- Four cloud agents in two months: how Claude Duet, Grok Bot, Muse, and Dots actually differ — dotey · 2026-10-22
- The Ontology Pipeline: a semantic knowledge management framework for LLMs — adnan_hashmi · 2026-10-22
- Anthropic Duet preview sparks debate: is the "I built my own harness" crowd facing sunk cost? — altryne · 2026-10-22
- Coinbase's x402 was built to kill API keys; Stripe ships Machine Payments SDK for agent payments — kleffew94 · 2026-10-22
- Anthropic ships first Claude API version since 2023: 2026-10-21 — ClaudeDevs · 2026-10-22