NVIDIA and Sharpa, a Singapore-based embodied AI startup, unveiled a joint humanoid reference platform at GTC Taipei. Named the NVIDIA Isaac GR00T Reference Humanoid Robot, the system integrates a Unitree H2 Plus chassis, Sharpa Wave tactile hands, and a Jetson AGX Thor T5000 compute module into a single, research-ready environment. General availability is set for October 2026.
Key takeaways
- Total degrees of freedom: 75 DOF — 31 from the Unitree H2 Plus chassis and 22 per Sharpa Wave hand
- Jetson AGX Thor T5000 module delivers 2,070 FP4 TFLOPS (Blackwell architecture) and 128 GB of unified memory
- Sharpa Wave hands: 1,000+ tactile pixels per fingertip, 0.02 N pressure sensitivity
- Platform ships preloaded with GR00T 1.7 model, integrated with Isaac Lab and Isaac Teleop
- The announcement coincides with Unitree Robotics' STAR Market IPO hearing in Shanghai
Three pillars: chassis, hands, and compute
NVIDIA Isaac GR00T Reference Humanoid Robot targets a long-standing bottleneck in humanoid research. Labs have been forced to choose between off-the-shelf platforms with limited customization and building their own stacks from scratch — at the cost of months of engineering work. This platform aims to close that gap.
The foundation is the Unitree H2 Plus chassis — a full-sized humanoid with 31 degrees of freedom. Mounted to its arms are Sharpa Wave hands: five-fingered, 22 DOF each, with a built-in array of over 1,000 tactile pixels per fingertip. With 0.02 N pressure sensitivity, the hands detect contact before the camera sees it — critical for tasks where the grasp occludes the field of view.
The compute layer is the NVIDIA Jetson AGX Thor T5000. Built on Blackwell GPU architecture, the module delivers 2,070 FP4 TFLOPS, a 14-core Arm CPU, and 128 GB of unified memory. That's enough to run policy inference in real time without offloading to an external server. The platform totals 75 degrees of freedom across the full system.
Seamless simulation-to-hardware workflow
The platform ships preloaded with the GR00T 1.7 model. Direct integration with NVIDIA Isaac Lab and Isaac Teleop lets teams collect data via teleoperation, train policies in simulation, and deploy them back to the physical hardware through Jetson Thor — with minimal sim-to-real degradation.
Rev Lebaredian, NVIDIA's Vice President of Physical AI Simulation, explained the rationale at a GTC Taipei Q&A session:
The problem we've had so far is that research labs spend a lot of time and energy on just making the basics of the robot work. Having a platform that's being developed at scale with spare parts and everything coming out of the box integrated together — they can just use it out of the box and go.
— Rev Lebaredian, VP of Physical AI Simulation, NVIDIA
The collaboration has prior history. Sharpa has been training tactile policies in Isaac Lab for months — including the widely covered bimanual apple peeling milestone earlier this year. The new reference stack formalizes that integration and opens it to external research teams.
IPO timing and geopolitical crosswinds
The June 1 announcement date is deliberate. The same day, Unitree Robotics faced the listing committee of the Shanghai Stock Exchange for its STAR Market IPO hearing. The company seeks to raise 4.2 billion yuan ($621 million) to fund manufacturing capacity for bipedal humanoids.
Unitree's financials show a company deep in investment mode. Q1 2026 revenue rose 68% year-on-year to 422.8 million yuan. Adjusted net profit dropped 52%, weighed down by surging R&D and sales costs. An official NVIDIA endorsement at GTC provides a strong signal to investors evaluating Unitree's transition from quadruped platforms to full-sized humanoids.
The geopolitical backdrop complicates the picture. US lawmakers have introduced the bipartisan American Security Robotics Act, which would impose a federal procurement ban on Chinese-made unmanned ground vehicles. Unitree's G1 and H1 models are standard equipment in Western academic labs today — a change in law could cut these platforms from publicly funded research.
The three-way partnership — NVIDIA from Silicon Valley, Sharpa from Singapore, Unitree from Hangzhou — may be a calculated response to that scenario. The platform is positioned as an open-frontier reference design backed by an American software stack. Whether that framing can insulate it from trade restrictions remains an open question.
First commercial pilots
Sharpa announced that a major, publicly visible pilot using its Wave tactile hands will launch in summer 2026, targeting food and beverage use cases. The company cited Grab in Singapore as a named enterprise partner. Commercial pricing for the reference platform has not yet been disclosed.
Why it matters
Humanoid robotics research has long suffered from deep fragmentation. Every lab builds its own stack from scratch: different chassis, different end-effectors, different software layers. The result is high entry costs, slow research cycles, and results that are nearly impossible to compare across teams.
NVIDIA Isaac GR00T Reference Humanoid Robot is aiming for the role the IBM PC played in personal computing — an open, standardized platform everyone builds on. A unified hardware and software stack shortens the path from research idea to working policy on a physical robot. NVIDIA's positioning is deliberate: GPUs, simulators, foundation models — and now reference hardware too.
For Sharpa, this is a chance to make tactile hands a default component rather than an optional add-on. For Unitree, it is an opportunity to exit the geopolitical grey zone under a Western technology partner's umbrella. Whether the platform delivers on its out-of-the-box promise will become clear after the October 2026 launch.
What's next?
- October 2026: planned release of the NVIDIA Isaac GR00T Reference Humanoid Robot per the GTC Taipei announcement
- Summer 2026: Sharpa launches commercial pilot with Wave hands in food and beverage — named partner is Grab in Singapore
- STAR Market's ruling on Unitree Robotics' IPO will determine the company's capacity to scale humanoid production
