See2Act teaches robots where to look and how to act in one denoising loop

heghbalz · x · 2026-07-22

A robotics research thread introduces See2Act, a diffusion-based policy for learning where to look while learning how to act.

The core idea is to couple perception and action inside a single denoising loop.
The project argues that robots cannot act well on things they cannot see, so gaze selection becomes part of the policy.
The post links to the paper, project page, and a longer thread explaining the approach.

Because this is a robot method demo, it belongs primarily to robotics/hardware, with research as a secondary angle.

More from Embodied