My robot succeeds around 63% of the time but consistently fails during the initial grasp — it pinches a single layer instead of catching the fold edge. I uploaded 1,240 episodes including the failures. Any ideas?
RT-1 Robot Action Dataset
Mobile Robot · 130K episodes · RGB Video + Actions + Robot State
Atlas
Humanoid · 148K episodes · 42 tasks
Look at the frames 200 ms before contact. If the wrist is already closing, your action chunk is committing too early — shorten it to 8 and let the policy re-observe.
Depth on folded fabric is unreliable at that range. Check whether your camera reports valid depth on the fold edge at all — ours returns holes on white cotton.
We had the identical failure. Adding 400 episodes where a human deliberately mis-grasps and then recovers fixed most of it.
Both suggestions land. Re-recording tonight with chunk 8 and a recovery subset, will post the new curve.