Amazon grasps 66 objects with one policy, no teleop
Amazon FAR trained a hand to grasp AirPods, dumbbells and pipe wrenches with a single policy and zero teleoperation data. Then someone asked how many of the 66 objects the robot had never seen before, and Jitendra Malik answered.
The big one
Jim Fan's CoRL reveal is tomorrow, and he is promising an uncut shot
Jim Fan narrowed his CoRL keynote tease to four words the day before: real robot, real time, uncut shot, autonomous. The talk is Monday November 9 at 1:45pm at the Hilton Austin, and when asked earlier he added "no speed up". Nobody outside has seen it yet.
Real robot. Real time. Uncut shot. Autonomous.
Shown this week
Amazon's Pick Anything grasps 66 objects with one policy, no teleop
Jitendra Malik showed Pick Anything, a single policy from Amazon FAR's Dex-Hands team that grasps 66 objects ranging from AirPods to dumbbells to pipe wrenches. It uses a stereo camera head with no wrist cameras and Sharpa Wave hands, and Malik says it was trained on video and RL (learning by trial and error) with zero teleoperation data.
Objects like these were used in training (indeed lot more than these 66).
Social Humanoid makes a Unitree G1 react to a person without mocap
Vongani Maluleke released Social Humanoid, a framework for human-humanoid interaction deployed on a Unitree G1. It pairs high-level social reasoning with low-level control so the robot responds to a human partner, with no mocap (motion capture suits and markers). Angjoo Kanazawa flagged it as the thing robotics has not solved, saying current robots are not socially intelligent at all.
Don’t worry, robotics isn't solved yet bc they’re not socially intelligent at all.