BiGym 2.0: Claude-written policies hit 51.9% on humanoid household tasks, nearing VLA
teortaxesTex · x · 2026-11-03
BiGym 2.0 benchmarks AI agents controlling a humanoid robot: each agent watches one demonstration and writes the program driving a Unitree G1, scored on hidden episodes. On the 9 core household tasks, Claude Opus 5.5 (via Claude Code) leads at 51.9%, followed by GPT-6 Astra at 46.5%; Kimi K3.1 Pro (18.4%) and DeepSeek V4.1 Flash (10.6%) trail clearly. On the full 20-task board, learned policies still lead (π0.5 VLA 75.0%, ACT 68.0%) but GPT-6 Astra's code-written policy reaches 64.5%, beating most RL/imitation baselines.
teortaxesTex highlights a broader pattern: both OpenAI and Anthropic now show later-released lower-tier models matching or exceeding earlier flagship performance on nontrivial long-tail capabilities. The quoted comment notes an LLM reaching this level just by writing rules is another showcase of scaling laws.
More from Embodied
- NVIDIA shows frontier AI agents driving Omniverse simulation workflows — CurieuxExplorer · 2026-11-03
- 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
- Huawei's TinyAstral Distills Self-Evolving Harnesses into 4B Robot Agents, Wins Across Six Benchmarks — Kehong Gong · 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