GR00T 2 drops, and Jim Fan's uncut demo lands today
NVIDIA says GR00T 2 learns skills from people wearing wristbands and runs them on robots that never practiced, trained on 400,000 hours of human video. Meanwhile 11,398 real trials say nobody has a general robotics model yet, OpenAI included.
The big one
Jim Fan promised a real, autonomous, uncut robot demo for today
Jim Fan spent the week building to a CoRL keynote at the Hilton Austin on Monday Nov 9 at 1:45pm, and he kept narrowing the promise: real robot, real time, uncut shot, autonomous. Asked directly whether it was sim or real hardware, he replied "Real. Autonomous, no speedup." Nobody outside the room has seen it yet.
Real robot. Real time. Uncut shot. Autonomous.
Shown this week
Amazon's Pick Anything grasps 66 objects with one policy and zero teleop
Jitendra Malik showed Pick Anything, a single policy from Amazon FAR's Dex-Hands team that picks 66 objects from AirPods to dumbbells to pipe wrenches, trained on video and RL with zero teleoperation data. Asked how many of those objects were unseen, Malik said objects like them were in training, far more than 66, and that generalization across lighting, position and clutter comes free.
the policy learned human-like dexterity from absolutely no human-robot interactions, no teleop or handheld tool data
General AI agents edge out robot VLAs in 901 head-to-head trials
RoboColosseum's full numbers landed and they complicate the week's loudest claim: across 901 head-to-head comparisons, general AI agents made more progress in 394 runs, VLA models (vision-language-action models built for robots) led in 265, and 242 tied. Jiafei Duan's topline from 41 days, 124 sites and 11,398 real-world evaluations was that OpenAI's GPT-6 Astra is seriously strong.
There remains no true general purpose robotics model -- not even from OpenAI
Out in the world
NVIDIA ships GR00T 2, which learns skills robots never practiced
Yuke Zhu announced GR00T 2, one model that runs on parallel grippers, dexterous hands and humanoids, with a feature called X-Adapt that transfers skills shown only by people wearing in-house devices to robots that never practiced them. The training mix: 400k+ hours of egocentric human video, 10k hours of UMI demos, 23.5k hours of robot teleoperation.
Much of what a robot learns need not come from robot data.
Teleop and other arguments
Chris Paxton waves off a paper claiming models can be tricked into stabbing
Robocurve said video tampering misled every frontier model it tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas and pouring water into a toaster. Chris Paxton's response: "I too would stab a baby if I thought I was seeing an apple. Come on guys, this is nothing." The work is being presented at CoRL Late Breaking Results.
Frequently asked questions
What did Jim Fan demonstrate at CoRL?
Jim Fan showed a real, autonomous robot demonstration in real time with no speedup or editing at a CoRL keynote at the Hilton Austin on Monday November 9. The demo was uncut and nobody outside the room had seen it before the event.
How many objects can Amazon's Pick Anything grasp?
Amazon's Pick Anything, a single policy developed by the Dex-Hands team, can grasp 66 different objects ranging from AirPods to dumbbells to pipe wrenches. The system was trained on video and reinforcement learning (using trial-and-error learning to improve performance) with no teleoperation (remote human control) data.
Who won more trials in RoboColosseum's robot comparison?
In 901 head-to-head comparisons, general AI agents made more progress in 394 runs while VLA models (vision-language-action models built for robots) led in 265 runs, with 242 ending in ties. RoboColosseum concluded there is no one-size-fits-all winner between the two approaches.
What is X-Adapt in NVIDIA's GR00T 2?
X-Adapt is a feature in NVIDIA's GR00T 2 model that transfers skills shown only by people wearing in-house devices to robots that never practiced those skills. The model runs on parallel grippers, dexterous hands and humanoids.
Can video tampering fool robot models into harmful actions?
Robocurve found that video tampering misled frontier models into harmful actions, including stabbing a baby doll at a 100% attempt rate. Chris Paxton characterized this as a perception failure rather than a safety failure, since a wrong action results from corrupted input to the robot's vision system.