TwinBridge preprint reports 98% robot success across five laboratory tasks

Researchers report that TwinBridge achieved a final mean autonomous success rate of 98% across five real-world laboratory tasks, using at most 25 minutes of real-world interaction per task. The arXiv preprint describes a method for transferring training and progress feedback from simulated environments to physical robots. The framework combines reconstructed three-dimensional scenes with physics simulation to initialize robot behavior and train a visual model that scores progress. During physical operation, those scores and execution history guide learning with human involvement. Experiments found that detailed rewards improved interaction efficiency over sparse rewards, while simulation-based initialization further accelerated learning.