Huawei's TinyAstral Distills Self-Evolving Harnesses into 4B Robot Agents, Wins Across Six Benchmarks
Kehong Gong · hf · 2026-11-03
Huawei researchers present TinyAstral, which transfers recursively self-improving (RSI) agent harnesses to tiny generalist manipulation agents. The harness is iterated via propose-edit paired A/B loops with a frozen 180B teacher, then frozen and distilled into a 4B student that keeps the performance gains at far lower compute.
On RoboTwin-50, the teacher (180B) climbs from 33.4 by +30.2 points to 63.6/62.8; distilled TinyAstral-4B gains +21.6 points over its untrained base (14.6 → 53.2 before distill). Across six benchmarks (RoboTwin-50, LIBERO-PRO, GuaVa, VLABench, SimplerEnv WidowX/Grab VM), TinyAstral-4B and Guava-4B post 41.2–93.6, beating GPT-5.x/GPT-6-driven baselines and the best VLAs.
More from Embodied
- Tsinghua & Berkeley open-source uni-robot-agent: Qwen beats GPT-6 zero-shot — Jiarui Yuan · 2026-11-03
- Stanford CoRL paper adds safety filters to humanoid RL policies with no retraining needed — drmapavone · 2026-11-03
- Reviewer: The $2000+ Duo Is Fun but Hard to Recommend as Your Only Phone — film_girl · 2026-11-03
- General-purpose LLM with robot arm beats specialized robot models: Kohane — zakkohane · 2026-11-03
- Zhiyui runs 6 robots for 24 hours straight in a convenience store; Google douses space datacenter hype — 创业邦 · 2026-11-03
- Apps still not optimized for iPhone Duo's dual screen weeks after launch — evielync · 2026-11-03