Which Camera to Trust

In fog, we lean on whichever senses still work. In the paper I'm writing, I ask whether a robot policy can do the same with its cameras: when one view goes bad, trust the other more.

I'm starting with pick and place, using a front camera and a wrist camera. I'll add noise to either view, measure how much task success drops, and see whether the policy can shift its weight to the clearer camera.

quality-aware attention

I'm building a quality estimator that scores each camera stream from 0 to 1 at the model's inference rate. The plan is to add the log of that score, with a small positive floor, to the cross-attention scores in the flow-matching action decoder:

score(view) = qkᵀ / √d + log(max(q_view, ε))

A clean view, with a score near 1, is left almost untouched. A corrupted one, near 0, is pushed down, so the decoder pays more attention to the view that still helps. I'll compare task success with and without this change under the same camera noise.