Reins preprint reports 88% success across five real-robot tasks

Researchers report that GPT-6.1 Sol completed 88% of real episodes across five robot tasks using Reins, a system that lets an off-the-shelf model operate a dual-arm robot without task-specific code. The findings appear in an arXiv preprint. Reins combines 3D perception with actions that report contact and grip. It chains actions subject to conditions and stops when one fails, while a second model instance watches cameras and reports changes. Without the system, the model failed every shell game and took 8 times as long. The researchers say code will be released upon publication.