Paxton says end-to-end mobile manipulation ran first try on Digit v4
Two clips from the conference floor, no success rate and no task list.
Chris Paxton ran end-to-end mobile manipulation on an Agility Robotics Digit v4 at IROS 2026 and says it worked immediately. He posted two clips from the session, labelled Pt 2 for the second one, with a short line attached.
Worked out of the box, first time. E2E mobile manipulation on @agilityrobotics Digit v4 at #IROS2026
"Worked out of the box, first time" is the entire claim. End-to-end here means one learned system driving the whole thing, moving the base and the arms together toward a task, rather than a stack of hand-tuned modules handing off to each other. Mobile manipulation is the part of humanoid work that tends to break, because the robot has to walk somewhere, stop, and then do something useful with its hands at a location it did not have a perfect setup for.
What the post does not say
No success rate. No number of attempts. No list of tasks. No word on how much of the run was autonomous versus set up in advance, and no comparison against anything else. Paxton did not say what the policy was or who trained it. That matters because "first try" describes one run on a conference floor, which is the demo condition most likely to flatter a system and also the one most likely to embarrass it. Nobody has a measured reliability figure for this yet.
The framing is worth keeping straight. This is a researcher's own account of a session he was part of, posted with video, not a company announcement and not a benchmark result. Read it as a signal that the thing ran, not as evidence of how often it runs.
The Digit context
The clips are on Digit v4, the previous generation of Agility's humanoid. Agility has since shown Digit 5, which swapped the backward-bent bird legs for forward-facing ones. So this is research work on hardware that is already a version behind, which is normal for academic labs and conference demos, where the robot you can actually get your hands on beats the one in the launch video.
What would make this concrete is the boring stuff. How many runs, how many succeeded, which tasks, and whether it holds up on a floor nobody cleared for it first. Until somebody publishes that, two clips and a sentence is what exists.