Training Reasoning Models in Simulation to Natively Use Real Robots
An open-weight 27B model trained entirely in simulation succeeded in 60% of real dual-arm robot trials, up from zero before training.
The researchers used reinforcement learning without demonstrations, letting the model observe, reason and generate the robot’s next end-effector pose. Training raised its success rate on the LIBERO simulation benchmark from 1.6% to 74.4%. On LIBERO-Pro tasks requiring a new task in a familiar scene, it achieved 57.4% success, compared with 0.75% for the demonstration-trained π₀.₅ model. The authors report no measurable degradation across nine reasoning, coding and vision benchmarks.
ArXiv · AI/CL/LG's note
The researchers used reinforcement learning without demonstrations, letting the model observe, reason and generate the robot’s next end-effector pose. Training raised its success rate on the LIBERO simulation benchmark from 1.6% to 74.4%. On LIBERO-Pro tasks requiring a new task in a familiar scene, it achieved 57.4% success, compared with 0.75% for the demonstration-trained π₀.₅ model. The authors report no measurable degradation across nine reasoning, coding and vision benchmarks.
ArXiv · AI/CL/LG's note
score 5