Unitree and NVIDIA Release H2 Plus Humanoid Reference Robot

Per PR Newswire and NVIDIA, Unitree Robotics and NVIDIA unveiled the H2 Plus reference humanoid robot built on the Isaac GR00T development platform (PR Newswire; NVIDIA). The reference design pairs Unitree's H2 humanoid body with Sharpa Wave five-finger tactile hands and onboard Jetson Thor compute, and ships with NVIDIA's Isaac GR00T software and models for data capture, simulation, training, evaluation and deployment (PR Newswire; NVIDIA). NVIDIA named Ai2, ETH Zurich, Stanford Robotics Center and UC San Diego as early institutional users of the reference design (NVIDIA). The launch materials include direct quotes from Unitree founder Xingxing Wang and NVIDIA CEO Jensen Huang (PR Newswire; NVIDIA). CNBC reports sales to research institutions are slated to begin later this year (CNBC).

The public materials describe the package as an integrated development stack: the Unitree H2 Plus body, Sharpa Wave five-finger hands, and onboard Jetson Thor hardware running Isaac GR00T models and simulation workflows (PR Newswire; NVIDIA). Reporting by South China Morning Post cites the Jetson AGX Thor T5000 configuration with 128 GB memory and up to 2,070 FP4 teraflops of AI compute as the on-device accelerator powering the platform (SCMP). CNBC and other outlets reference NVIDIA's Blackwell-generation GPU family as the underlying architecture for the Thor module (CNBC; SCMP).

Editorial analysis - technical context: Researchers building humanoid skills typically spend months on hardware integration, sensor calibration, simulation fidelity and co-simulation of control and perception. Reference designs that combine a validated kinematic body, dexterous end effectors, and on-device, high-throughput AI accelerators shorten that integration path. For practitioners, pre-integrated tactile hands and a packaged software stack for data capture and policy training reduce immediate engineering overhead and enable faster iteration on control, imitation learning, and sim-to-real transfer experiments.

Editorial analysis: Public reporting frames this announcement as part of NVIDIA's broader push into what the company calls "physical AI," where on-device large-model inference, simulation tools, and reference hardware lower barriers to robotics research. For the academic and lab market specifically, an open reference design with named institutional adopters can accelerate reproducibility and benchmarking across research groups. From an industry-coverage perspective, the combination of a China-based hardware vendor, a Singaporean manipulator vendor, and a U.S. chip/software provider underscores a cross-border supply and collaboration pattern already visible in robotics R&D coverage (PR Newswire; NVIDIA; SCMP; CNBC).

adoption and tooling. Observers should follow availability and pricing details for research accounts and institutions, reported integration notes from early adopters at Ai2, ETH Zurich, Stanford and UC San Diego, and third-party evaluations of sim-to-real performance using the bundled Isaac GR00T models (NVIDIA; PR Newswire). What to watch: interoperability and extensibility. Researchers will test how easily alternative hands, sensors, or control middleware can be integrated with the reference stack. What to watch: compute and thermal trade-offs. Published benchmarks for on-device inference latency, power draw, and real-world policy throughput using the Jetson Thor configuration will determine practical limits for long-duration tasks (SCMP; CNBC).

Editorial analysis: For practitioners, the immediate value of H2 Plus is reducing engineering friction when spinning up humanoid experiments. Over time, repeated use of a common reference design across labs could improve result comparability and accelerate incremental progress on dexterous manipulation, whole-body control, and policy generalization, provided the research community can access the hardware and software at scale (NVIDIA; PR Newswire).