An arXiv paper proposes Agent as Policy (AGP), a method in which a general-purpose agent handles both planning and execution while directly operating a physical robot. The authors report that this requires no training tailored to a specific task or environment. The work sits in the cs.CL category and is a revised submission.
A new arXiv paper proposes TIDAL, a control scheme that alternates between diffusion-based planning and action execution to keep vision-language-action models running at high frequency. The authors argue that current VLA systems rely on a low-frequency batch-and-execute approach, and that the resulting mismatch between model inference speed and robot control rate creates gaps in which the agent cannot react. TIDAL aims to close that blind spot while retaining the semantic generalization of large VLA models.
Researchers behind the Show-Harness project propose an embodied harness that lets general-purpose vision-language models operate robots directly. The approach aims to bridge the gap between the broad world knowledge of VLMs and the practical challenge of turning that intelligence into robot control.