Contact-rich manipulation—like inserting a peg into a hole—requires a robot to feel resistance and adjust mid-motion. Most learned tactile policies need specialized pretraining and per-task optimization. In a new arXiv preprint, researchers propose GPT-as-Tactile-Policy, a prompting approach that lets the general-purpose multimodal model GPT-6-Astra interpret touch without any robotics-specific training.
GPT-6-Astra receives expert demonstrations summarized into task guidance and tactile references, then combines that context with multimodal observations and interaction history to generate control commands. Across five UniVTAC benchmark tasks, the method achieved a 93.8% mean success rate, comparable to state-of-the-art specialized policies and higher on selected tasks. The largest gain came on Insert Hole, where tactile feedback improved success from 31% to 99%.
Execution analysis showed the model performing contact-guided probing and correction, suggesting it uses explicit expert experience and online feedback to make progress. Since this is a single preprint, the results have not yet been independently replicated, but the findings point toward general-purpose models as viable tactile policies without per-task engineering.