Recent advances in multimodal models have allowed GPT-6 Astra to perform robotic manipulation through in-context learning, according to a new arXiv paper. The paper's title describes GPT-6 Astra as a "native visuo-tactile policy" for contact-rich manipulation.
Contact-rich control and tactile perception are often considered strengths of specialist policies rather than generalist multimodal models. The abstract acknowledges this view, then appears to argue that GPT-6 Astra can handle these tasks natively.
The abstract is brief, but the central claim is that GPT-6 Astra is not just a generalist that happens to work for manipulation, but a policy that integrates visual and tactile information directly. Further details would be needed to assess how this is achieved.