Researchers led by Jiaming Tang have trained an open-weight reasoning model with 27 billion parameters to control real robots directly, using reinforcement learning entirely in simulation and without demonstrations, according to a paper on arXiv. So far, they write, only proprietary frontier models, among them GPT-6 Astra and Claude Fable 5.2, have done this well.
Like an agent that uses a computer, the model looks at the scene, reasons step by step and outputs the next position of the robot's gripper, without extra tooling. The team built a training engine called PEAR to train it at scale on simulated tasks. After training on the LIBERO and RoboDojo tasks, the model succeeded in 60% of trials on a real dual-arm robot, while the base model succeeded in none.
On LIBERO, its success rate rose from 1.6% to 74.4%, close to what GPT-6 Astra and GPT-6.1 Sol reach with maximum reasoning effort. Given a new task in a familiar scene on LIBERO-Pro, it succeeded in 57.4% of episodes, against 0.75% for the vision-language-action model pi-0.5, and it showed no measurable loss on nine reasoning, coding and vision benchmarks.
Background
More stories about GPT-6 Astra
Claude Fable 5.2
AI model
Claude Fable 5.2 is a general AI model from Anthropic with improved physical robot control skills.



