PEAR brings open-weight robot reasoning out of simulation
A new robotics preprint reports that simulation-only reinforcement learning can teach an open 27-billion-parameter reasoning model to control a physical robot. The PEAR team reports 18 successful trials out of 30 on a real dual-arm YAM, versus none for the base model. Those are the researchers’ measurements, not independent validation. Research report
The work trains Qwen3.8-27B without demonstrations. Its actions specify an end-effector pose; conventional inverse kinematics and controllers still execute movement. “No harness” therefore does not mean software-free robot control. The team’s evaluated weights use BF16: it explicitly says NVFP4 quantization reduces accuracy, so consumer-hardware feasibility should not be confused with equal benchmark quality. Research report
The public PEAR repository exposes training, evaluation and deployment infrastructure, with an Apache-2.0 code license and a linked model collection. Its deployment documentation includes YAM and SO-101 robot stations. That makes this more useful to builders than a video alone.
Guth analysis: The interesting shift is where the capability lives. If general reasoning models can absorb robot skills during post-training, researchers can test alternatives to stacks of separately engineered skills. But 30 physical trials are a starting point, not a reliability certificate. The next meaningful evidence is replication across more tasks and robots, especially under constrained inference budgets.
Sources and citations
Revision history
Published .