2026-06-06

NVIDIA and Unitree launch the Isaac GR00T reference humanoid (H2 Plus) for research

— views

NVIDIA and Unitree unveiled the Isaac GR00T Reference Humanoid Robot — the Unitree H2 Plus: a ~6-ft, 150-lb, 75-DoF open platform with onboard Jetson Thor compute and the open GR00T VLA software stack, aimed at academic robotics research. Ships from Unitree in late 2026.

At GTC Taipei (late May / June 1, 2026), NVIDIA and Unitree announced the Isaac GR00T Reference Humanoid Robot — the Unitree H2 Plus — a full-stack, open humanoid platform built specifically for academic research. After a year of one-off humanoid demos, the pitch is different: a single standardized, reproducible baseline — body, onboard brain, and software — that any robotics lab can buy and build on.

What it is

The reference design pairs a commercial chassis with NVIDIA’s newest edge compute and an end-to-end open software stack:

Component	Detail
Chassis	Unitree H2 Plus — ~1.8 m (6 ft), ~68 kg (150 lb), 31 body DoF
Hands	dual Sharpa Wave tactile five-finger — 22 DoF → 75 DoF total
Compute	Jetson AGX Thor T5000 — Blackwell GPU 2,070 FP4 TFLOPS, 14-core Arm, 128 GB unified, 40–130 W
Sensing	head stereo camera (140° × 102° FOV), wrist cameras, IMU
Actuation	arm torque 120 N·m, leg torque 360 N·m; arm payload 7 kg (peak 15 kg)
Software (open)	Isaac Teleop, GR00T VLA foundation models, Isaac Sim + Isaac Lab, Isaac ROS
Availability	from Unitree, late 2026

Why it matters

A common baseline for embodied AI. Hardware + Blackwell edge compute + the GR00T VLA stack, open-sourced end-to-end (Teleop → Sim/Lab → ROS), so labs can reproduce each other’s work instead of each hacking a bespoke rig.
Thor puts the brain on the robot. 2,070 FP4 TFLOPS and 128 GB unified memory mean large VLA (vision-language-action) models can run onboard, untethered — the compute that makes real-time humanoid VLA inference plausible.
Heavyweight first adopters. Ai2, ETH Zurich, the Stanford Robotics Center, and UC San Diego’s ARCLab are named launch partners — NVIDIA wants its stack to be the default substrate for humanoid research.
A US-compute + China-hardware axis. NVIDIA brain, Unitree body — announced as Unitree reportedly eyes an IPO.

Practitioner note

It’s a research platform, not a deployable product. Availability is late 2026 and every number here is a vendor spec, not an independent lab result — read the DoF / payload / TFLOPS as spec-sheet figures.
The unlock is standardization + an open stack. If labs converge on H2 Plus + GR00T, datasets and trained policies become portable across groups — the thing humanoid research has most lacked.
Watch the cheaper on-ramp. A GR00T reference workflow for the smaller Unitree G1 is slated for GitHub and Hugging Face soon — that’s the accessible path for labs without an H2 Plus budget.
75 DoF, incl. 22-DoF tactile hands, is the dexterity bet. Manipulation, not locomotion, is where VLA policies still have to prove out.

The under-considered angle

The quiet move is NVIDIA standardizing the substrate. Whoever supplies the reference hardware + compute + software that research converges on captures the ecosystem — datasets, benchmarks, trained policies, and the PhDs who graduate on it. By bundling Unitree’s body, Jetson Thor, and the GR00T stack into one buyable baseline, NVIDIA is doing for humanoid research what CUDA did for GPU computing: not winning a single robot, but becoming the default layer every robot is built on. The chassis vendor can change; the substrate is the moat.

Specs are vendor-published from the 2026-05-31/06-01 announcement and not yet independently lab-tested; the platform ships late 2026.

NVIDIA and Unitree launch the Isaac GR00T reference humanoid (H2 Plus) for research

What it is

Why it matters

Practitioner note

The under-considered angle

Sources