Event

GPT-6-Astra tactile policy achieves 93.8% success rate on five UniVTAC tasks

1 reports · 1 sources

Ziyong Wang and other researchers proposed the GPT-as-Tactile-Policy method, which summarizes expert demonstrations into task guidance and tactile references, enabling GPT-6-Astra to generate control commands by combining multimodal observations and interaction history and to adjust actions based on contact feedback, without robot-specific pretraining or per-task policy fine-tuning. On five contact-rich manipulation tasks in UniVTAC, the method achieved an average success rate of 93.8%. In the Insert Hole task, adding tactile feedback raised the success rate from 31% to 99%. The researchers state that overall performance is comparable to current state-of-the-art policies, with higher success rates on some tasks.