NVIDIA pushes deeper into open-source robotics with a GR00T humanoid reference platform and physical-AI toolchain

NVIDIA is betting that robotics progress will accelerate when developers share the same data, simulation tools, and evaluation methods. On June 1, the company announced the NVIDIA Isaac GR00T Reference Humanoid Robot, an open humanoid reference design aimed at academic research teams that have struggled to assemble hardware plus software across the full workflow from robot bring-up to real-world validation. The announcement is the newest announcement of how NVIDIA is positioning its physical-AI efforts, and it was highlighted in market coverage on June 8.

At the core is a packaged approach NVIDIA describes as “body” plus “brain.” The reference robot combines a Unitree H2 Plus humanoid chassis and Sharpa Wave five-finger hands for dexterous manipulation with NVIDIA Jetson Thor onboard compute for advanced reasoning and control, then connects those components to Isaac GR00T open software and models. NVIDIA said the chassis is nearly 6 feet tall, weighs about 150 pounds, and includes 31 degrees of freedom, and it emphasized a modular design so teams can use the full platform or integrate selected components into their own pipelines. Jensen Huang, NVIDIA’s founder and CEO, said the reference design offers researchers a single open platform to work toward general-purpose physical intelligence.

The company frames Isaac GR00T as more than a robot. NVIDIA calls it an open reference platform for general-purpose humanoid robots that includes open data and data pipelines, an open robot foundation model, simulation frameworks, and middleware, plus CUDA-X accelerated runtime libraries and Jetson Thor for on-robot inference and control. In the same announcement, NVIDIA also quoted Steve Cousins of the Stanford Robotics Center saying open platforms let researchers share code and compare robot behaviors on physical hardware, a theme that NVIDIA has repeatedly returned to as robotics teams try to standardize evaluation and reduce reinvention.

NVIDIA’s broader Isaac ecosystem also includes open simulation and benchmarking components. NVIDIA’s Isaac developer portal describes Newton, an open-source, GPU-accelerated physics engine co-developed by Google DeepMind and Disney Research and managed by the Linux Foundation. NVIDIA also pointed developers to Isaac Lab and Isaac Lab-Arena, where the latter focuses on generalist robot policy evaluation across many GPU-accelerated simulation environments. NVIDIA says Isaac Lab-Arena integration with Hugging Face’s LeRobot Environment Hub can reduce evaluation time from days to under an hour by running standardized benchmarks through parallel simulation.

Another building block in NVIDIA’s pitch is OSMO, which the company describes as a cloud-native workflow orchestration platform for scaling physical-AI pipelines across heterogeneous compute. In its GitHub documentation, OSMO is designed to coordinate training, simulation, and edge hardware testing using a simple YAML configuration, avoiding what NVIDIA calls the need for custom scripts and infrastructure expertise for each environment. NVIDIA says OSMO is production-grade and that it was originally built to support physical-AI workloads at NVIDIA, including Project GR00T and Isaac Lab, and it claims it orchestrates thousands of GPU-hours daily across cloud training clusters and edge devices.

From an investors’ perspective, the key issue is whether NVIDIA’s open stack can turn into a larger share of the robotics software development funnel and related hardware deployments. NVIDIA’s recent investor relations update on open model families described partners adopting Isaac GR00T variants for humanoid robot deployments and other companies using NVIDIA Cosmos for physical-AI training and video analytics, suggesting the ecosystem is taking shape beyond prototypes. If universities and startups use NVIDIA’s reference platforms to iterate faster, NVIDIA could benefit indirectly through increased demand for its robotics compute stack, including Jetson systems and NVIDIA-accelerated runtimes that physical-AI developers need to run and evaluate these models.

Still, NVIDIA did not provide commercial details in the GR00T reference-humanoid announcement. The company did not disclose pricing, expected sales volume for the reference robot, or any quantified revenue impact from the open tooling. It also did not specify how quickly partner benchmarks will translate into real-world performance at scale, an open question for most humanoid robot efforts. What to watch next is whether NVIDIA follows up with clearer adoption metrics, more documentation on evaluation outcomes across common benchmarks, and any deployment commitments from robotics builders that go beyond simulation and lab settings.