NVIDIA shipped GR00T 2 and Isaac Lab 3.0 at CoRL 2026 in Austin this week. The headline metric is hard to ignore: five human video demonstrations are now enough to teach a robot a new manipulation skill with 95% average success across 50 evaluated tasks. That is an NVIDIA-controlled benchmark, not an independent audit, so weight it accordingly. But the direction is clear — robot development is following the LLM playbook, and the data bottleneck that made robotics expensive is starting to crack.
What GR00T 2 Actually Does #
GR00T 2 is a world action model, not a chatbot bolted onto a robot arm. It observes an environment, predicts how actions affect it, executes multi-step dexterous manipulation — gear alignment, cable connection, fine assembly — and recovers from failures via API. It is built on NVIDIA Cosmos and trained on 100,000 hours of egocentric human video plus diverse robot manipulation datasets.
The performance breakdown:
- Zero demonstrations: completed 40 of 50 evaluated tasks at least once
- One demonstration: 71% average success across 50 tasks
- Five demonstrations: 95% average success across 50 tasks
GR00T 2 currently sits in the top three across the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces benchmarks for both the VLA and world action model categories. The AlphaNIST result is worth flagging separately: a unified GR00T 2 policy hit 98% training success across all 20 NIST Assembly Task Board tasks using roughly 39 times less GPU time than per-task policy engineering.
The Data Bottleneck Is Cracking #
The most practically interesting pieces are EgoHand and the Video-to-Data pipeline — two tools shipping alongside GR00T 2 that attack the data collection problem directly.
EgoHand is an open model that reconstructs 3D hand motion from first-person video. It was trained with Noitom Robotics data and hit 40,000 downloads on Hugging Face within two weeks of its quiet release. The Video-to-Data (V2D) pipeline takes human demonstration video and converts it to simulation-ready robot training data. It is open source on GitHub at nvidia-isaac/video_to_data.
Cross-adaptation (X-Adapt) ties it together: you can now transfer skills between robot bodies without robot-specific training data. Record a human demonstrating a task, run EgoHand, run V2D, fine-tune GR00T 2 in Isaac Lab — that is the new proposed workflow. Whether it holds up outside controlled lab conditions is the real question. Organizations like Foxconn, Siemens, and Nexuni are listed as early evaluators, but public production deployment data is not available yet.
Isaac Lab 3.0: What 2.x Users Need to Know #
Isaac Lab 3.0 is a significant rewrite. If you are running 2.x in production, the migration is not optional-later. It is a fresh install with uv as the only supported installer, and several APIs changed in ways that will silently break existing code.
The most dangerous change: quaternion order flipped from (w, x, y, z) to (x, y, z, w) across all APIs, configuration values, asset data, and math utilities. If you do not catch this, your robot will move in the wrong direction with no error thrown.
Other breaking changes to address before migration:
.data.*properties no longer returntorch.Tensordirectly — append.torchor.warpAppLauncherreplaced bylaunch_simulation; Hydra config helpers removed--headlessflag removed; visualizer behavior changed- The
Imusensor renamed toPVA(reports angular velocity and linear acceleration only)
The full migration guide is at docs.nvidia.com/isaaclab/isaac-lab/migrating-to-3-0. Read it before touching anything.
Newton Physics and What It Unlocks #
Isaac Lab 3.0 ships with the Newton physics engine as an alternative to PhysX. Newton handles deformable objects — cables, cloth, multi-jointed robotic hands — that previously required custom simulation workarounds or were simply left out of training environments. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton delivered up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.
The architecture is now backend-agnostic: one task API works across PhysX, Newton, and multiple renderers. You write the task once and can switch physics engines without rewriting it. This is a meaningful quality-of-life change for teams doing sim-to-real research.
Industrial Benchmarks and Developer Access #
Two new evaluation frameworks ship with this release. DexBench covers 15 tasks across retail, logistics, manufacturing, and hospitality. VABAR (from UC Berkeley and NVIDIA) tests six industrial tasks across standardized difficulty tiers, designed for agentic robotics workflows. Both are available through the Isaac Lab-Arena evaluation platform.
Access: GR00T 2 weights are gated on Hugging Face under a non-commercial research license — request access on the model page, approval typically takes 24-48 hours. Isaac Lab 3.0 is fully open source at github.com/isaac-sim/IsaacLab. The V2D pipeline is at github.com/nvidia-isaac/video_to_data. For the migration guide if you are on 2.x: docs.nvidia.com/isaaclab/isaac-lab/migrating-to-3-0.
If you are building on Isaac Lab 2.x, the quaternion change alone justifies a careful migration plan before updating. If you are starting fresh with robot learning in 2026, GR00T 2 plus the V2D pipeline is the most accessible entry point NVIDIA has shipped to date.