Amazon's Pick Anything grasps 66 objects with one policy
Amazon FAR's Dex-Hands team says the system learned human-like grasps from video and trial and error, with no teleoperation data at all.
Jitendra Malik posted a video of "Pick Anything", a single policy that produces human-like grasps for 66 different objects, from AirPods to dumbbells to pipe wrenches. The work comes from the Dex-Hands team at Amazon FAR. Vision comes from a stereo camera head with no wrist cameras, and the hands are Sharpa Wave.
The part that got attention is the training recipe. Malik says the system was trained using video and RL (reinforcement learning, where the robot improves by trial and error), with zero teleoperation data. Most dexterous manipulation work still leans on humans puppeteering the robot to collect demonstrations, so a policy this broad with none of that is a different bet.
How they did it
@notmahi, who says he has been spending time with Malik's dexterous hands team at Amazon FAR, wrote up the idea behind it. His framing is that behavior cloning mastered learning from "robot-complete" data, and the open question is what to do with data that is not so perfect. Pick Anything, he says, learned dexterity with no human-robot interactions at all, no teleop and no handheld tool data.
He also lays out the two failure modes the method tries to bridge. Learning from humans without action labels usually collapses into a semantics problem, a fancy way to prompt the robot. RL for manipulation has the opposite issue, where you end up with technically correct solutions that look wrong. He says the paper is still to come.
Shown versus claimed
@aiquanting asked the obvious question under the video, how many of the 66 objects were never seen in training, noting the dumbbell and pipe wrench are much more convincing if they are new. Malik answered directly. Objects like these were used in training, in fact a lot more than these 66. His argument is that you have to show examples of humans grasping various objects to learn human-like grasps, and once there are enough of them, big data does the work.
Objects like these were used in training (indeed lot more than these 66).
So this is generalization across object categories, size and shape, not a cold start on a brand new object. Malik adds that generalization across lighting, position and background clutter comes for free and does not need big data.
The team is running live demos at CoRL 2026, demo 12, on November 9 and 10. @notmahi says the booth is in Griffin hall and that he will be there to answer questions. Until the paper lands, the no-teleop claim rests on the video and on what the team says about its own pipeline.
the policy learned human-like dexterity from absolutely no human-robot interactions, no teleop or handheld tool data, just letting humans be humans and robots be robots
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.