Unitree H2 Plus

What Separates H2 Plus from H2?

The body is identical: 1820 × 456 × 218 mm, ~70 kg, 31 degrees of freedom (6 leg, 7 arm, 3 waist, 2 head), 360 N·m leg and 120 N·m arm peak torque, ~3 hours from a 15 Ah / 0.972 kWh battery. The difference sits entirely in the compute and sensing layer. H2 Plus ships with an NVIDIA Jetson T5000: a 2560-core Blackwell-architecture GPU with 5th-generation Tensor Cores, a 14-core Arm Neoverse-V3AE 64-bit CPU and 128 GB of 256-bit LPDDR5X memory. The manufacturer's stated peak is 2070 TFLOPS at FP4 precision.

What 128 GB of Memory Means

In a humanoid the bottleneck is usually not raw compute but keeping the model resident on the device. 128 GB of LPDDR5X over a 256-bit bus is enough to hold a vision-language-action (VLA) class large model without forcing lossy quantization, and to run it alongside the perception pipeline. On-device inference without a cloud dependency is an architectural choice for projects that need in-facility latency and data privacy, a cloud round trip cannot serve a 100 ms control loop.

Sensing Hardware

• Head camera: horizontal field of view 140° ±3, vertical 102° ±3.
• Wrist cameras (optional): 3 MP ultra-wide, the scene as the hand sees it at the moment of grasp.
• Array microphones and high-power speakers.
• Optional SharpaWave dexterous hands: 22 active degrees of freedom per hand, 1000 tactile pixels per fingertip.

Engineering Assessment

H2 Plus is the version that bolts a research-class compute stack onto the same mechanical body. The distinction matters if you will run a large model on-device, pursue manipulation research with tactile feedback, or need cloud-independent inference. Mission time and torque budget, however, are identical to H2: the Jetson T5000 does not extend battery life; it tightens the thermal and energy budget. H2 Plus is therefore not a robot that runs longer, but one that runs a heavier model in the same window.