Two new papers push robots past the fingertips

Two research teams posted work on the same day that attacks the same blind spot in robot manipulation, which is everything that happens outside the fingertips.

Manipulating with the chest and forearms

Conor Mc Gartoll introduced Whole-Body Dexterity, or WBD, which treats the entire humanoid body as one large dexterous manipulator. The pitch starts from how people actually carry things. As Mc Gartoll put it, humans brace boxes against their chests, tuck objects under their arms, and roll balls along their forearms, while humanoids today mostly manipulate with their hands.

Conor Mc Gartoll

@cmcgartoll

X

we brace boxes against our chests, tuck objects under our arms, and roll balls along our forearms

Nima Fazeli, who shared the work, framed it as moving skills normally associated with multi-finger hands onto the full body. He listed rolling, sliding and handoffs as the behaviours in question, and credited @cmcgartoll, @dangan1693, @pthangeda_, @arjun__gupta, @carohiguerarias, @mukadammh and @ManikantanNambi.

Seeing how hard the gripper squeezes

Separately, Haonan Chen posted VisualForce, which infers how hard a gripper is squeezing from how much the gripper deforms, instead of reading a sensor. The system learns force-aware manipulation and, Chen says, needs no force or tactile sensors at deployment. The tasks shown range from handling berries, where too much grip ruins the object, to inserting plugs, where too little grip fails the task.

Haonan Chen

@HaonanChen_

X

Can a robot see how hard it is squeezing without force sensors?

That combination is the interesting part. Soft or compliant grippers are forgiving but hard to control precisely, and adding tactile sensing adds cost, wiring and failure points to every single finger. Using the visible bend of the gripper as the signal sidesteps that, if it holds up.

Shown versus claimed

Both of these are announcements from the teams themselves. Neither thread posts success rates, and nobody outside either lab has replicated the results yet. The clips are demonstrations chosen by the authors, which is normal for a paper release and is also the reason to wait before treating either as solved.

What they share is a direction. Most manipulation research puts the intelligence in the hand and the policy, then assumes the rest of the robot is a mount for it. WBD says the torso and arms are contact surfaces too. VisualForce says the hardware's own deformation is already a sensor if you look at it. Both are cheaper paths to touch than bolting on more sensors, and both still have to prove they work on tasks the authors did not pick.

Nima Fazeli

@NimaFazeli7

X

Dexterity we usually expect from a multi-finger hand (rolling, sliding, handoffs), now across the whole humanoid body.