Summary: At the Conference on Robot Learning in Austin on November 9, 2026, NVIDIA introduced GR00T 2, an open world action model for dexterous manipulation, alongside Isaac Lab 3.0, the Newton physics engine integration, and two new open benchmarks. The shift is from teaching robots to walk to teaching them to align gears and plug in cables. For builders, this stacks into an open pipeline: pretrain on shared models, rehearse at scale in simulation, then score generalization before touching hardware.
What GR00T 2 actually does
GR00T 2 is an open frontier world action model built on NVIDIA Cosmos, trained on diverse robot and human data, including 100,000 hours of NVIDIA’s open source human egocentric data. The interesting claims are about transfer rather than raw capability. According to NVIDIA, the model completed 40 of 50 evaluated tasks at least once without any task-specific demonstrations, and hit 71% average success across 50 tasks with just one demonstration per task, rising to 95% with five. It also learns from human demonstrations and carries skills between different robot bodies without robot demonstrations.
That last point is the mechanism worth understanding. Instead of training one policy per task per robot, a world action model learns a predictive model of how actions change an environment, so a skill demonstrated by a human on one platform can be adapted to another. AI agents can then call GR00T 2’s API to chain skills in sequence and recover from errors — a step toward the multistep autonomy that single-task policies struggle with.
NVIDIA reports the model ranks in the top three across all vision language action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks. Early adoption is real but early-stage: Nexuni is testing it for laundry handling, Foxconn is applying it to server factory automation, and Siemens is evaluating industrial use cases. Testing and evaluating — not deployed at scale.
The data side matters too, if you’ve ever tried to build a robot dataset. NVIDIA’s open Video to Data workflow converts human video into simulated robot demonstrations, and the EgoHand model, which reconstructs 3D hand motion using data from Noitom Robotics, crossed 40,000 Hugging Face downloads in its first two weeks. Open robotics models are getting real distribution — Hugging Face’s SmolVLA pointed in this direction earlier, and GR00T 2 is the frontier-scale version of the same bet.
Where the training and scaling happens
A pretrained model gives robots core skills; simulation is where you refine and evaluate them cheaply. Isaac Lab 3.0, NVIDIA’s open source robot-learning framework, now lets agents set up training environments with simulation-ready assets, switch between the PhysX and Newton physics solvers, pick renderers for perception-based training, and apply a reinforcement post-training library to foundation models like GR00T 2.
The Newton integration is the technically meaningful change. Accurate physics determines whether skills learned in simulation transfer to a real robot, and Newton can now model deformable objects like cables and cloth, plus articulated systems like dexterous hands — exactly the manipulation problems that rigid-body-only simulation couldn’t represent. NVIDIA reports that in a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. Skild AI extended this further with STEEL, a throughput-optimized rigid-body solver built on Newton.
Scale shows up in the results. AlphaNIST, built in Isaac Lab, trains a single policy on all 20 tasks of the NIST Assembly Task Board — millimeter-scale parts — using only a yes-or-no success signal. It reached more than 98% training success versus 81% for separate per-task policies, with roughly 39x less GPU time. Underneath that efficiency is Success-Guided Sampling, a method from the University of Washington and NVIDIA that concentrates training on tasks where the policy succeeds about half the time — the band where it’s actually learning. With SGS, training can scale beyond a million parallel environments.
Benchmarks that score generalization, not demos
Performing a skill once is not the same as performing it under changing conditions, and NVIDIA’s own framing admits this gap. Isaac Lab-Arena 0.4, in developer preview, composes repeatable evaluations and runs policies across thousands of variations in parallel, then reports which conditions cause failures.
Two open benchmarks build on it. DexBench, from RLWRLD with NVIDIA, covers 15 tasks across four scenes spanning retail, logistics, manufacturing and hospitality, testing generalization across object, spatial and physics variations; Genesis AI and mimic robotics are evaluating it through early access, and teams can run it on Bifrost’s Manifold platform with free initial credits. VABAR, from UC Berkeley and NVIDIA, evaluates six industrial tasks with standardized variations, difficulty tiers and success metrics. For anyone who has shipped a demo that broke in production, having failure-locating evaluation infrastructure as an open tool is the quietly important part of this release.
Agents writing robot programs — and what’s still unproven
Multistep reliability remains the hard problem, and NVIDIA’s agentic robotics pitch addresses it. UC Berkeley’s Graph-as-Policy framework uses agents to connect skills into robot programs, tests them in NVIDIA Isaac simulation, and refines them from failures. Because the output is interpretable code, engineers can review and verify programs before deployment. Ambi Robotics, Bosch Research and Siemens are exploring the approach. Ken Goldberg, the UC Berkeley professor leading VABAR and GaP, said he considers agentic robotics the most exciting technological development since the internet in 1995.
It’s also worth noting that two of the six Best Paper Award finalists at CoRL, out of thousands of submissions, were NVIDIA papers: GRAIL, on generating humanoid loco-manipulation from world models, and RoboTTT, which scales visuomotor context three orders of magnitude without growing inference latency.
Balance that against what these announcements don’t establish. The benchmark numbers are vendor-reported, not independently peer-reviewed. GR00T 2’s industrial partners are testing, not shipping. And no source provides quantitative evidence that agentic workflows reduce cost or risk versus conventional pipelines yet.
What to check next
If you’re building in physical AI, the practical move is to run before you train. Evaluate whether DexBench’s four scenes or VABAR’s six tasks match your workflow, use Manifold’s free credits to score an existing open policy against them, and only then invest in hardware or custom data collection. The open questions that matter for your case: how well GR00T 2 transfers human demonstrations to a custom robot body with minimal task-specific data, and whether GaP-style agent-generated programs pass an engineer’s review for your safety requirements. The infrastructure to answer both cheaply just became open source.