NVIDIA GR00T 1.7 and Agile ONE SN Launch at GTC Berlin

NVIDIA Expands Isaac GR00T Platform with GR00T 1.7 VLA Model

NVIDIA announced GR00T 1.7 on October 21, 2026, during GTC Berlin. The 3-billion-parameter vision-language-action model uses a Cosmos backend and carries a commercial license. It is available on Hugging Face and LeRobot for immediate download and fine-tuning. Inference runs at 25 Hz on Jetson Thor hardware after TensorRT NVFP4 quantization. This release follows earlier GR00T iterations and adds relative end-effector action space plus expanded pretraining on roughly 32,000 hours of real human egocentric data and 8,000 hours of simulated rollouts.

Developers gain a cross-embodiment foundation that accepts camera images, language instructions, proprioception, and an embodiment token. The model outputs chunks of continuous actions for manipulation and whole-body control. Benchmarks reported in NVIDIA technical materials show gains over the prior release on tasks such as DROID-F6 and SimplerEnv Bridge. The architecture supports ONNX export and integrates with frontier models like GPT-6 Astra for high-level task decomposition before GR00T handles low-level execution.

Agile ONE SN Reference Humanoid Integrates GR00T and Jetson Thor

Agile Robots introduced the Agile ONE SN as a complete reference platform built around the new GR00T stack. The robot stands 1.4 meters tall, weighs 40 kilograms, and features 31 degrees of freedom in the body plus additional DOF from two five-finger Wuji hands for a system total of 71 DOF. Each arm carries a 3-kilogram payload. Multimodal sensing includes head and wrist RGB-D cameras, lidar, and inertial measurement units. The platform ships preconfigured with GR00T software for demonstration capture, training, evaluation, and deployment.

Availability is scheduled for early 2027. The design targets research and validation teams that previously spent weeks integrating separate hardware, simulation, and control layers. NVIDIA states the unified stack reduces setup time from weeks to days. Planned integration of the NVIDIA Halos safety stack adds infrastructure-level protections spanning operating system, middleware, and applications. Partners including ABB Robotics, KUKA, and Universal Robots are adopting elements of the broader Isaac platform announced alongside the reference humanoid.

Isaac Lab 3.0, Isaac OS Runtime, and Open Workflows

NVIDIA paired the model and hardware announcement with Isaac Lab 3.0 updates for robot learning in simulation. A new open-source Isaac OS runtime coordinates policy execution, sensor fusion, and control loops on physical hardware. Three reference workflows for DGX Station cover reinforcement-learning refinement of GR00T 1.7 policies, dexterous manipulation training inside Isaac Lab, and Cosmos-based synthetic data generation. These workflows incorporate ChatGPT Astra for goal decomposition.

The runtime differs from the foundation model itself by managing real-time coordination of components once a policy leaves simulation. Open datasets and LeRobot-compatible formats further lower barriers for teams using consumer-grade or research hardware. European ecosystem partners listed in the announcement include Flexion, Hexagon Robotics, NEURA Robotics, and Sereact, each building autonomous systems on the shared stack.

Technical Architecture and Performance Details

GR00T 1.7 employs a Cosmos-Reason2-2B vision-language backbone based on Qwen3-VL architecture. A diffusion transformer action head performs flow-matching denoising over 16-step action horizons. Pretraining emphasizes relative end-effector deltas rather than absolute targets to improve cross-embodiment transfer. On-device execution benefits from NVFP4 quantization that sustains 25 Hz closed-loop control on Jetson Thor without external GPUs.

The Agile ONE SN body uses the 31-DOF Agile ONE S platform as its base. Head-mounted and wrist cameras feed the VLA at native resolutions supported by the flexible vision encoder. Lidar and IMU data supplement visual input for navigation and balance. Whole-body control policies coordinate locomotion with upper-body manipulation through the GR00T workflow.

Limitations and Open Questions

GR00T 1.7 remains a foundation policy that requires embodiment-specific post-training and safety certification for production use. The 3-kilogram arm payload and 40-kilogram total mass constrain heavy industrial tasks compared with larger humanoids. Availability of the full Agile ONE SN in early 2027 leaves a gap for teams seeking immediate hardware validation. Questions persist around long-horizon task reliability after fine-tuning and the compute cost of scaling synthetic data pipelines on DGX clusters.

Integration of Halos safety features is still in planning stages rather than shipping code. Real-world generalization across unseen environments depends on the volume and diversity of additional demonstration data collected by adopters. NVIDIA has not published detailed latency breakdowns beyond the 25 Hz figure or power consumption numbers for Jetson Thor deployments.

Ecosystem Adoption and Second-Order Implications

ABB, KUKA, and Hugging Face have publicly aligned with the expanded Isaac stack. Research institutions and smaller humanoid developers gain access to pre-trained weights and reference runtimes that previously required substantial internal engineering. Serviceability improves because a common runtime and model format reduce custom middleware across fleets. Cost implications include lower per-robot integration expenses once the data-to-deployment pipeline stabilizes, though DGX-based workflow training still demands significant capital for organizations without cloud credits.

Counter-arguments note that open models invite rapid community forks that could fragment compatibility. Hardware reference platforms like Agile ONE SN compete with in-house designs from companies already shipping units. The 25 Hz control rate supports many manipulation tasks yet may prove marginal for high