Gym-Anything builds RoboWorld, 25 coding-agent robotics simulations
Original title: Gym-Anything for Robotics: Scaling physical simulations with digital agents
Frontier models score an average reward of 0.06 and fail every RoboWorld test task; several teaching methods raise performance.
Gym-Anything for Robotics treats simulation authoring as a coding-agent task and uses an adversarial creation-audit loop. Simulations are chosen using U.S. GDP data to cover economically valuable physical work. The resulting RoboWorld set has 25 simulations, 25 test environments and 175 training RL environments. In-context retries, a privileged teacher, self-assisted learning and code policies for vision-language-action models such as π0.5 each improve results. The pipeline, environments and benchmark are released.
- RoboWorld has 25 simulations, 25 test and 175 training RL environments.
- The best frontier models fail every test task, with average reward 0.06.
- A privileged teacher reads simulator state and teaches a student in context.
- Code policies generate demonstrations to train VLAs such as π0.5.
- Code and data are at gym-anything.github.io/robotics.
- Countries
- United States
- AI models
- π0.5
From the original article
1 min read at arxiv.orgRealistic, fast robotics simulations could enable large-scale training of embodied AI models that transfer to the real world. However, despite decades of progress, available simulations cover only a narrow range of real-world tasks and physical phenomena. The core bottleneck is that creating better simulations requires significant expert effort and therefore does not scale.
By Pranjal Aggarwal · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.