Robot Control Viable with 32×32 Depth Input

yacineMTB · x · 2026-07-17

In the context of robot control, the discussant suggests that depth sensors are sufficient. Input can be downscaled to 32×32 pixels and then used to train a control model, which is already enough to support RL control.

The conversation also touches on robot "blindness" as a prominent issue, necessitating more reliable external inputs for perception. However, a counterpoint is raised that depth cameras have a limited effective range, which might be insufficient if mounted directly on a robot.

Related event: T800 Robot Combat Exposes Perception Flaws(3 posts)→

Original post →

More from Embodied

Embodied channel →