CMU Auto-Builds 25 Robot Simulations with Agents; Frontier Model Solves 0 of 25 Test Tasks
wellecks · x · 2026-10-30
CMU released Gym-Anything for Robotics, which automatically converts physical tasks into robot simulations and RL environments. Grounded in U.S. wage and time-spent data, the team decomposed 1,016 occupations and 465 everyday activities into 31,961 atomic physical actions, then built 25 high-value simulations (manufacturing, construction, households, healthcare, logistics) with physics spanning liquids, cloth, molten metal, soft tissue, and acoustics.
- 25 test + 175 training RL environments, supporting swappable single-arm, bimanual, and humanoid robots
- Core insight: scenes, physics, tasks, and rewards are all code—building a simulation is a coding-agent task, so better coding agents yield better sims
- A creation-audit loop deploys audit agents that try to break environments with legal robot actions
- Brutal eval: frontier Claude Opus 5.5 solves 0 of 25 test tasks with an average reward of just 0.06
- The team developed three methods leveraging privileged simulator information, including training Physical Intelligence's Pi 0.5
More from Embodied
- NVIDIA's MotionBricks robotics motion-generation model trends on Hugging Face — nvidia · 2026-10-31
- Agility Robotics to raise $465M ahead of Nasdaq debut as AGLT on Nov 5, Foxconn leads with $201M — xmercury_one · 2026-10-31
- Tesla to open Optimus humanoid robot reservations in early December, per source — Scobleizer · 2026-10-31
- 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