Paxton tries TRIMI, a robot hand you can also wear
Chris Paxton posted clips from the CoRL floor of a hand that doubles as a data collection tool.
Chris Paxton spent time with TRIMI at CoRL, a robot hand from @DennisHongRobot and team that a human can also wear, and posted clips of what it did.
Impressively dexterous and presumably gives you "perfect" data with no remapping.
pulling a credit card out of my wallet
The first clip is a credit card coming out of his wallet. Paxton picked that task on purpose, because it needs force and precision at the same time, and he says the hand handled it with very little issue. The second is harder. In-hand manipulation of a Rubik's cube, meaning the cube gets moved and rotated within the fingers rather than set down and regrasped. That one took a couple of tries.
Why wearable matters
Paxton frames TRIMI as the logical extension of an UMI system, the handheld grippers researchers strap to their own hands to record demonstrations without ever driving a robot. The catch with those setups is usually retargeting. The tool you wear and the hand on the robot are not the same shape, so every recorded motion has to be remapped onto different kinematics, and some of the skill leaks out in translation. If the thing on your hand is the robot hand, that step goes away.
"Impressively dexterous and presumably gives you 'perfect' data with no remapping."
Note the "presumably". Paxton is describing the pitch, not a result.
Shown versus claimed
Everything in the clips is a human wearing the hand and moving it. There is no policy running here, no autonomy, no success rate, and no paper or dataset in the post. What was demonstrated is that a human-worn version of the hand is dexterous enough to do two tasks that are genuinely awkward, which is the precondition for the data argument rather than proof of it. Whether a model trained on TRIMI data actually performs better than one trained on remapped UMI data is not something the post answers, and nobody has shown that yet.
It is also one researcher's hands-on impression at a conference booth, over a few minutes, with the team standing there. That is worth something, because Paxton is not shy about saying when a demo does not hold up. It is not a benchmark.
The interesting question is scale. Wearable grippers took off because they are cheap and anyone can collect data with them in any kitchen. A full dexterous hand you wear is a different cost structure, and the post says nothing about what TRIMI costs or whether it will be released.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.
