NVIDIA announced GR00T 2 yesterday — the second generation of its foundation model for humanoid robotics. Unlike GR00T 1, GR00T 2 ships with open weights under the NVIDIA Open Model License, runs on the consumer-priced $999 Jetson AGX Thor dev kit, and supports any humanoid robot platform with at least 18 actuators.
In demos, GR00T 2 controls Boston Dynamics Atlas, Figure 02, Apptronik Apollo, and Unitree H1 from the same model weights — a level of cross-platform generalization that previously didn't exist in robotics.
What GR00T 2 does that GR00T 1 didn't
Three breakthroughs:
- **Cross-embodiment**: train once, deploy on any humanoid with similar topology
- **Language conditioning**: tell the robot what to do in natural language, no programming required for most tasks
- **Few-shot learning**: 5-10 demonstrations of a new task is enough to add it to the robot's repertoire
The architecture is a 9-billion parameter transformer that takes camera input + language prompt + proprioception data and outputs joint torques at 240 Hz.
Performance benchmarks
On the standard ROBOTHOR-Eval suite (NVIDIA's robotics benchmark):
- Task completion rate: 78% (GR00T 1: 41%; prior best: 52%)
- New task adaptation: 71% from 5 demonstrations (GR00T 1: 38%)
- Cross-platform transfer: 84% performance retention from Atlas to Figure (was: ~30% for prior models)
These are the kind of numbers that move robotics from "expensive lab demo" to "deployable product."
The pricing reality
NVIDIA structured the offering across three tiers:
- **Open weights**: free download for research, evaluation, and non-commercial use
- **Commercial license**: $20,000/year flat for any deployment up to 10,000 robots
- **NVIDIA Cloud**: $0.40/robot/hour for managed inference on H200 cluster
The Jetson AGX Thor at $999 is a real changer. Most academic robotics labs and robotics startups now have their compute affordable enough to put on the actual robot, not in a server room.
Who benefits
For robotics startups: the foundation model problem is solved. You build the chassis and sensors, NVIDIA provides the brain.
For Tesla / Apptronik / Figure: this is mixed news. They lose the moat of having to build the brain themselves. They keep the moat of integrated hardware + chip stack + production scale.
For research labs: the most consequential release of the year. Reproducible foundation models for robotics didn't exist a week ago.
Why this matters more than language models
Language models change knowledge work. Foundation robotics models change physical work. The historical precedent for "physical work" is much larger than knowledge work — agriculture, manufacturing, construction, logistics, services.
If GR00T 2 generalizes the way the demos suggest, the cost of "I want a robot that does X" drops by an order of magnitude. The bottleneck shifts from "can the robot do this?" to "can I build the chassis, sensors, and supply chain to deploy it?"
Sources
- NVIDIA GTC Spring (April 27, 2026): GR00T 2 release
- The Verge (April 28, 2026): NVIDIA's robot brain just got open-sourced
- Reuters (April 28, 2026): GR00T 2 controls four different humanoid platforms