The strongest models score 0 in robot simulation—and that 0 is more useful than the demo videos

Pranjal Aggarwal announced Gym-Anything for Robotics on X. It has 25 simulations and 200 RL environments, with materials covering liquids, cloth, molten metal, soil, and even sound, and tasks drawn from economically valuable domains. An RL environment can be understood as a task arena for repeated trial and error; VLA stands for vision-language-action model. Eitan Shaar's comment was direct: before putting robots into the real world, use simulated environments to find failures early and learn safely.

Existing robot simulations favor rigid bodies on tabletops because they are easy to simulate. Real physical work cannot avoid liquids, cloth, molten metal, and chemical reactions. Gym-Anything has code-writing agents build these simulations. The problem is right there: good-looking visuals do not equal accurate physics. The project sets five criteria for a simulation: physically accurate, visually realistic, solvable, fast to configure, and resistant to gaming. Creator agents submit evidence for each criterion—rendering, physics measurements, and auxiliary demonstrations all count. Four auditor agents check them, one of which specifically tries to break the simulation, over three rounds.

Each simulation has more than one RL environment. The environment must correspond to real work, state the task in natural language, be solvable, have a correct verifier, provide rewards that reflect progress, and span a range of difficulty. Six auditor agents check them. Each environment can run different robots, from single arms to humanoids, and must withstand thousands of variations in object pose, appearance, and physical parameters.

On the hardest split, frontier models also score 0%. On the easy split, the best, Claude Opus 5.5, reaches 32%. Someone online asked whether 0% means the model cannot complete any task or only the hardest tasks. The thread specifies 0% for the hardest split and 32% for the easy split. These numbers give the benchmark discriminative power and help select suitable difficulty for RL training.

Can it learn within the simulation? Yes. The thread says frontier models can learn using different methods such as ICL. After receiving the true state from the simulator, Claude Opus 5.5 improves faster over multiple attempts. It can also teach a student that cannot see that state, doubling the student's first-attempt reward. The code policies it generates can produce demonstrations that train π0.5 from 0% to 53%, running 14 times faster than Opus itself.

25 simulations and 200 environments do not look like much. The team says the method is scalable; what is missing is compute, and they have already selected 120 simulations to keep adding. The code, RL environments, benchmark, and paper are all open source. People working on robot learning can directly spin up environments to evaluate models, or use code policies to generate demonstrations for training VLAs. General readers can watch videos of robots writing GPT with two hands, spraying π, playing the xylophone, and driving a forklift. Demos are demos—the 0% on the hardest split is what deserves more attention.

Links:

• Project page: https://gym-anything.github.io/robotics/
• Code: https://github.com/gym-anything/gym-anything-robotics
• Paper: https://arxiv.org/abs/2610.40565
• Original thread: https://x.com/PranjalAggarw16/status/2116195676342088082
• Eitan Shaar's comment: https://x.com/EitanShaar/status/2116603269317624113