What is the late-2026 humanoid-robot collaboration between Singapore-based Sharpa, Nvidia and China’s Unitree, including the robot’s intended research applications, announcement at Computex, components and capabilities, Sharpa’s role and prior work, Nvidia’s accompanying AI softw

The collaboration is NVIDIA’s Isaac GR00T Reference Humanoid Robot: an open, research-focused reference design announced by Jensen Huang on June 1, 2026, alongside Computex in Taipei, with availability from Unitree planned for late 2026—not a late-2026 announcement.

• Purpose: It is intended to give universities and research institutes a common platform for physical-AI research: collecting data, testing algorithms, training manipulation and loco-manipulation policies, and validating robot behaviours in the real world. Target tasks include delicate work such as computer assembly, handling objects, administering an IV, food preparation, cleaning and ironing.

• What it combines:

  • China’s Unitree H2 Plus supplies the roughly human-size body—about 1.83 m tall and 68 kg.
  • Singapore-headquartered Sharpa supplies tactile, five-finger hands with 22 degrees of freedom, designed for human-like dexterity and touch-sensitive manipulation.
  • NVIDIA Jetson AGX Thor T5000 provides onboard AI compute: a Blackwell GPU, 14-core Arm CPU, 128 GB unified memory and up to 2,070 FP4 teraflops for real-time sensing and inference.

• Sharpa’s contribution and track record: Sharpa, founded in 2024, develops the dexterous hands that let the robot grasp, adjust and use objects while generating the physical feedback needed to teach manipulation. It has R&D in Shanghai and business operations in California, previously worked with NVIDIA on data collection for robotics foundation models, and demonstrated its hands dealing blackjack and assembling pinwheels at CES 2026.

• NVIDIA’s software and training layer: The hardware is paired with the open Isaac GR00T development platform—software, models and workflows covering simulation, training, evaluation and deployment. NVIDIA says labs retain control over robot data, training data, telemetry and logs; the modular stack can be adopted whole or integrated into existing pipelines. NVIDIA also plans a Unitree G1 reference workflow on GitHub and Hugging Face.

• The university-lab problem it addresses: Huang said NVIDIA built it for higher education because assembling such a system independently is “insanely hard.” In practice, labs have had to integrate disparate bodies, hands, compute, data collection, simulators and control software—the “Frankenstein robot” problem—before starting their actual research. The reference design standardizes the basic platform so researchers can spend more time developing and comparing skills.

• Early users and availability: NVIDIA named Ai2, ETH Zurich, Stanford Robotics Center and UC San Diego’s Advanced Robotics and Controls Laboratory as early research users. Unitree is expected to sell the reference robot in late 2026; pricing and fuller commercial details were not announced.

• How Huang framed it strategically: He presented humanoids as an avenue for bringing “physical AI” into major industries and described the resulting opportunity as multitrillion-dollar. At the same event, he tied the robot to a wider NVIDIA pitch: agentic AI will increase demand for computing infrastructure; NVIDIA would supply the stack from data centers—via its Vera CPU—to local AI systems, including AI PCs using the RTX Spark chip to run models and agents on-device. In that framing, GR00T is NVIDIA’s robotics/embodied-AI endpoint for the same hardware-and-software platform strategy.