Overview
GR00T 2 is research published by NVIDIA's GEAR lab on November 10, 2026, positioned as a robotics foundation model. Its predecessor GR00T 1 was a vision-language-action model: look at the frame, read the instruction, emit motor commands. This generation is a world-action model built on a video-action diffusion transformer. It produces the next action together with a short generated video of what comes next, then derives control commands from that prediction.
The significance is in how training data is structured. Teleoperation is the expensive part of robot data: a person drives the machine repeatedly to record motions, and each recording only suits one embodiment. GR00T 2 pretrains on a mix of three sources: 402,730 hours of egocentric human video from head-mounted cameras, 10,184 hours of handheld UMI gripper demonstrations covering 2,897 tasks, and 23,500 hours of teleoperation. That teleoperation data is spread thinly across platforms, with 15,146 hours on YAM, 3,591 on ABC-130k, 2,890 on Sharpa hands, 1,256 on Wuji hands mounted on Tianji arms, 350 on DROID's Franka and 275 on the Unitree G1.
The team measured the scaling effect on 72 Dexter tasks: going from zero to full human-video share moved task progress from 64 to 77 percent in video rollouts and from 50 to 70 percent on real hardware.
On capability, a robot with five-fingered hands can light a match, slice a banana and toss packages onto a conveyor without having been teleoperated for those tasks, with success rates between 10 and 80 percent depending on the task. Few-shot adaptation is cheap: across 50 tasks where the model already did reasonably well, under three minutes of conventional data lifted success from 40 to 95 percent, and the team cites 30-minute LoRA fine-tuning on a single RTX 5090.
Long assembly is a second track that pairs GR00T 2 with a higher-level language model which splits the task into steps, marks grasp points in the frame and launches recovery routines when a part falls. Assembling a toy car from loose parts scattered on a table succeeded in at least 97 percent of trials with that pairing, against 23 percent without the corrective model and extra fine-tuning. Humanoids get a slimmer variant, GR00T 2 Turbo, at 2 billion parameters, running at 38 milliseconds on an RTX 5090 and 124 milliseconds on the Jetson Thor inside a Unitree G1, with whole-body motion driven by the SONIC controller that plans a 1.6-second motion token roughly once a second and executes at 50 Hz.
The report is equally clear about the current stage. The model does not yet ingest direct tactile sensor data and can miss fine-grained state such as a half-turned screw, and the higher-level language model introduces waiting: during trash sorting the robot sat idle for about a third of the time, averaging more than seven seconds for the planner's next computation.
The release boundary deserves its own paragraph. Model weights, licence and SDK are not published. What is public is the technical whitepaper, the Human-1M-100K dataset on Hugging Face with a dataset explorer, the open hardware designs for GEAR UMI and DexUMI, and the SimFoundry simulation tooling. The model is offered through a WAM API so a higher-level agent can call it as a tool, but no date or licence for weights has been promised.
Key Features
- World-action model: Predicts the next action and a short future video together, using diffusion-generated frames to judge whether a motion is sound before deriving control commands, replacing direct motor regression.
- Cross-embodiment transfer: Video prediction lets one motion sequence be expressed across parallel grippers, dexterous hands and humanoids, with per-robot heads emitting commands for Dexter, YAM and G1 among others.
- Human video as the base: 402,730 hours of egocentric human video anchored by 10,184 hours of UMI gripper demos and 23,500 hours of teleoperation, which demotes teleoperation from staple to supplement.
- Few-shot adaptation: Five demonstrations and under three minutes of robot data per task took success from 40 to 95 percent, with the team citing 30-minute LoRA fine-tuning on one RTX 5090.
- GR00T 2 Turbo: A 2-billion-parameter variant that trades image detail for speed, at 38 milliseconds on an RTX 5090 and 124 milliseconds on the Jetson Thor inside a Unitree G1, running without external compute.
- Long assembly with a planner: A higher-level language model splits steps, marks grasp points and triggers recovery when a part drops, lifting toy-car assembly from 23 percent to at least 97.
- SONIC whole-body control: Plans a 1.6-second motion token about once per second for bipedal robots while the underlying controller executes at 50 Hz.
- WAM API: Exposes the model as a tool that a higher-level agent can call, so robot motion can be wired into existing task orchestration.
- Open dataset and hardware: Human-1M-100K ships on Hugging Face with an explorer, the GEAR UMI and DexUMI hardware designs are open, and SimFoundry covers simulation.
Use Cases
- Research groups testing cross-embodiment transfer on their own arms while cutting teleoperation collection.
- Factory or warehouse assembly with long task sequences that need a planner above and grasp control below.
- Humanoid makers evaluating onboard inference, where Jetson Thor latency and power matter.
- Data teams reusing public egocentric video to reduce in-house capture cost.
- Simulation and synthetic-data work that needs an action representation aligned with real robots.
Pros
- Human video replaces most teleoperation, which changes the cost structure of data collection.
- Cross-embodiment transfer has a measured control, with real-robot progress up 20 points at full human-video share.
- Few-shot adaptation is cheap, with half an hour of fine-tuning on a single GPU.
- The Turbo variant runs onboard, which covers robots that cannot carry external compute.
- Dataset and hardware designs are open, so reproduction does not depend on internal resources.
Pricing
NVIDIA has not published model weights, a licence, an SDK or a price list. What is publicly available is the technical whitepaper, the Human-1M-100K dataset of 100,000 hours on Hugging Face, the open hardware designs for GEAR UMI and DexUMI, and SimFoundry simulation tooling. The model is offered through a WAM API that higher-level agents can call as a tool. No timeline for weights or commercial licensing has been committed, so confirm terms with NVIDIA before planning a deployment.
Summary
GR00T 2 changes the technical route. Instead of translating camera frames and language commands straight into motor commands, it predicts the next action and a short stretch of future video at the same time, using the generated video to judge whether the action makes sense. What that buys is cross-embodiment transfer: a person wearing a head-mounted camera does a task once, and the robot has a chance of learning it, which turns teleoperation from staple into supplement. The evidence is that going from no human video to all of it lifts real-robot task progress from 50 to 70 percent. What needs stating clearly is the open boundary: weights, licence and SDK are not released, and what is public is a 100,000-hour dataset and an open hardware design.
Version History
- NVIDIA releases GR00T 2 world-action model (2026-11-10): NVIDIA's GEAR lab released GR00T 2, a robotics foundation model that moves from a vision-language-action design to a world-action model built on a video-action diffusion transformer, generating a short future video alongside each action. Pretraining mixes 402,730 hours of egocentric human video, 10,184 hours of UMI gripper demonstrations and 23,500 hours of teleoperation, computed on more than a thousand GB300 chips. Raising the human-video share from zero to full lifted real-robot task progress from 50 to 70 percent. A 2-billion-parameter GR00T 2 Turbo runs at 38 milliseconds on an RTX 5090 and 124 milliseconds on the Jetson Thor in a Unitree G1. Weights and licence are not published; the Human-1M-100K dataset and UMI, DexUMI open hardware designs are.