Skip to main content
Back to timeline
arXivSource publication:

A 27B open-weight model learns to run robots in simulation: 18 of 30 real dual-arm trials succeed, base model none

Synopsis

The open-weight reasoning model Qwen3.8-27B barely succeeded at robot tasks before, reaching 1.6% on LIBERO. The researchers built the PEAR training engine and trained it with robot-use reinforcement learning in simulation to output end-effector poses directly, without demonstrations or perception tools. After training it reaches 74.4% on LIBERO and succeeds in 18 of 30 trials on a real dual-arm robot it never saw, against none for the base model.

Source-provided article image: Training Reasoning Models in Simulation to Natively Use Real Robots
arXiv

Interpretation

A 27B open-weight reasoning model can learn to operate a robot directly from reinforcement learning in simulation alone, without demonstrations or external tools, and carry the skill to real hardware. Strong robot-use agent performance had so far come only from proprietary frontier models such as GPT-6 Astra and Claude Fable 5.2, while open-weight models stayed far less capable on the same benchmarks and failed some simple tasks. On LIBERO, training raises the success rate from 1.6% for the base model to 74.4%, approaching GPT-6 Astra at 80.6% and GPT-6.1 Sol at 76.2% at maximum effort, and exceeding both at medium effort (49.3% and 48.1%).

The model works like a computer-use agent: it observes camera images and arm state, reasons in a chain of thought, and outputs the next end-effector pose. Inverse kinematics converts that pose into joint targets and the robot's own controller executes it, so simulation and the real robot share one interface, and switching robots only changes the parts and cameras described in the prompt. Training runs in two simulators, LIBERO and RoboDojo, and the real-robot trials use a dual-arm YAM never seen in training, with 30 trials per model; the final model succeeds 18 times, the LIBERO-only checkpoint 7 times, and the base model none.

Perspective

The demonstrated scope is a small set of tasks in two simulators, LIBERO and RoboDojo, plus 30 real-robot trials on one dual-arm YAM; the intended reader is a researcher working on robot foundation models and reinforcement-learning post-training.

Training stopped after two RoboDojo tasks because of limited GPU resources; the model reasons at length before each move, about 4.3 s per round, and a chain of thought can reach 32k tokens; the 27B base model lacks knowledge of its embodiment, so success depends heavily on the embodiment and environment description given in the prompt.