CMU Releases Gym-Anything: Auto-built Robot Simulations and Software Test Environments
A CMU team (Pranjal Aggarwal et al., with wellecks involved in the release) has introduced Gym-Anything, a framework that converts physical tasks and real software into scalable testing environments, spanning two directions: robotics simulation and computer use.
Confirmed
- Gym-Anything for Robotics covers 25 simulators with 200 RL environments in total, spanning modalities including liquids, cloth, molten metal, soil, and even sound — all chosen from economically valuable domains
- Task selection method: using US wage and time-use data, the team decomposed 1,016 occupations and 465 daily activities into 31,961 atomic physical actions, and picked the 25 most economically valuable simulation directions from there
- Benchmark results are brutal: frontier models score a 0% pass rate on the hardest test tasks
- A "create-audit loop" is introduced: a fleet of auditing agents is deployed, one of which specifically tries to break the simulation using legitimate robot actions, to catch environment bugs
- Software side: 200+ real applications (GIMP, Blender, OpenEMR, etc.) can be automatically converted into computer-use agent test environments, auto-installed and configured inside sandboxed VMs
- The project page and paper are out; due to limited resources the team stopped at 25 simulators and 200 RL environments, and is openly inviting collaborators to help scale up
Why it matters
This work showcases both an automated environment-generation pipeline (data-driven task selection + automated install/config + audit-agent QA) and the stark shortcomings of frontier models on realistic physical tasks, providing scalable infrastructure for robotics and agent evaluation.
2026-10-30 ~ 2026-10-31 · 7 related posts
Primary sources
- [source] Gym-Anything launches: 200 robot RL environments across 25 simulations — wellecks · 2026-10-30
- CMU Auto-Builds 25 Robot Simulations with Agents; Frontier Model Solves 0 of 25 Test Tasks — wellecks · 2026-10-30
- [source] CMU's Gym-Anything Auto-Builds 200+ Agent Environments from Any Software, Selected by GDP — wellecks · 2026-10-30
- Gym-Anything Adds Audit Agents That Try to Break Simulations With Legal Robot Actions — wellecks · 2026-10-30
- [source] Gym-Anything Team Stopped at 25 Simulations Due to Resource Limits, Seeks Collaborators — wellecks · 2026-10-30
- Gym-Anything launches 200 robot RL environments across 25 simulators; best model scores 0% — wellecks · 2026-10-31
- CMU's Gym-Anything uses coding agents to auto-build 200 robot RL environments; best model scores 0% — AkariAsai · 2026-10-31