CMU's Gym-Anything uses coding agents to auto-build 200 robot RL environments; best model scores 0%
AkariAsai · x · 2026-10-31
CMU researchers (Pranjal Aggarwal, Sean Welleck) unveiled Gym-Anything for Robotics at CUA Week: a framework where coding agents automatically turn physical tasks into realistic robot simulations with physics, tasks, and verification built as code.
Key facts:
- 25 simulations and 200 RL environments (25 test / 175 training) spanning manufacturing, construction, food service, households, labs, logistics—covering liquids, cloth, molten metal, granular soil, soft tissue, even acoustics
- Tasks derived from 1,016 occupations and 465 everyday activities, decomposed into 31,961 atomic actions and selected by US wage/time value
- Frontier models fail: Claude Opus 5.5 solves 0 of 25 test tasks (avg reward 0.06), though ICL methods can significantly improve VLAs like π0.5
- Core insight: simulations are code, so they improve as coding agents do. Paper, code, and leaderboard available.
More from Embodied
- New MolmoSpaces-CommonSense benchmark: no robot policy tops 30% on manipulation tasks — RanjayKrishna · 2026-10-31
- Neuralink Unveils Next-Gen R1 Surgical Robot That Implants Brain Chips — elonmusk · 2026-10-31
- Google demos Sprout, a semi-autonomous rover that hunts sunlight to care for office plants — AI_Andrew · 2026-10-31
- NVIDIA's ScheduleGen generates collision-aware multi-arm robot training data via MILP scheduling — AjayMandlekar · 2026-10-31
- Waymo's Toyota deal tackles its biggest weakness: getting enough cars — binarybits · 2026-10-31
- OpenAI shows GPT-6 Astra remotely controlling a robot arm in builder community — OpenAIDevs · 2026-10-31