NVIDIA GR00T 2 with a reasoning agent recovers from failures: 97-98/100 toy assemblies vs 71-84 without
shoumikgoswami · reddit · 2026-11-10
NVIDIA launched GR00T 2 at CoRL on Nov 9, a robot model that jointly predicts future video and actions. The standout result: a reasoning agent helped a YAM robot recover from failed steps during long assembly runs — 98, 97 and 97 successes out of 100 trials assembling a toy plane, tractor and Hot Wheels car, versus 71-84 for the same post-trained policy without the agent, with parts dumped in cluttered layouts absent from training data. The agent watches camera images and robot state, calls GR00T 2 for manipulation, and switches to code-based skills when a part drops or progress stalls; some recovery skills were trained in simulation and transferred to hardware. Caveats: results are from NVIDIA's own experiments on three toy tasks, don't reflect base-model reliability alone, and deployment costs aren't reported. The author's takeaway: the interesting product direction is a robot that finishes despite imperfect attempts.
Related event: NVIDIA Unveils GR00T 2 World-Action Model for Robots(12 posts)→
More from Embodied
- Using open models to reverse-engineer a Surface Pro Linux kernel build, almost done — vincent_koc · 2026-11-10
- Robotics-at-Scale Pipeline Chart Sparks Debate: Almost All Bitter Lesson — KyleMorgenstein · 2026-11-10
- CoRL 2026 to host twin workshops on adapting pretrained robot policies — huihan_liu · 2026-11-10
- Robbyant releases LingBot-Touch, a tactile model giving robot hands human-like touch sensing — CyberRobooo · 2026-11-10
- Dyna Robotics shows RL vs IK head-to-head on wheeled robot, unveils workflow agent Dyna-2.1 — chris_j_paxton · 2026-11-10
- AI researcher Chris Paxton comes away positive from a Tesla Cybercab ride — chris_j_paxton · 2026-11-10