Digit ran mobile manipulation for 12 hours at IROS

Chris Paxton says the end-to-end whole body mobile manipulation demo on an Agility Robotics Digit ran for about 12 hours at IROS this week. He posted video and said it is 1x speed, autonomous, and in a previously unseen location that was not in the training data.

That is the part that is new. Paxton earlier said the system worked on the first try. A first try is a good sign about generalization. A conference floor shift is a different test, and it is the one most policies fail.

Why hours matter more than clips

Most humanoid manipulation footage you see is a run of a few minutes, often trimmed, sometimes sped up. Paxton is explicitly ruling out the speedup here, which means the pace in the video is the pace in the room. End-to-end means a single learned policy drives the whole thing, whole body means the legs and arms are being commanded together rather than the robot parking itself and then reaching.

A conference demo is also a hostile environment for a learned policy. People crowd in, the lighting is wrong, the floor is not the floor from the dataset, and objects get moved between runs by anyone walking past. Paxton says the location was not in the training data at all, so the policy had to transfer rather than recall.

What the post does not say

Twelve hours is the length of the demo session, not a claim of twelve hours without a reset or an intervention. Paxton does not say how often the robot was reset, how many attempts succeeded, what task it was doing, or how the policy was trained. Nobody outside the team has those numbers yet.

So treat this as a strong signal about one system holding up in public, not as a published reliability figure. The video is one clip out of a long day, and the researcher who ran it is the one telling you how the rest of the day went.

Chris Paxton

@chris_j_paxton

X

video is 1x speed, autonomous, in a previously-unseen location (not in training data).