Unitree Agx Module For H1-275 Tops (AGX-H1-275)

Unitree Agx Module For H1-275 Tops (AGX-H1-275)

External Back-Mount Design

The AGX-H1-275 shares the same external back-mount installation approach as the 550-TOPS variant. Unlike modules that require internal installation, the AGX module mounts externally on the robot's back. This design not only makes installation straightforward but also provides easy access for quick adjustments, maintenance, or upgrades. External mounting preserves the robot's internal structure, ensuring that existing hardware and configurations remain intact while minimizing installation time and effort.

The external mounting approach is significant for research teams in several ways. First, it enables the module to be installed on an existing H1 or H1-2 without factory re-configuration — the module is a field-installable upgrade. Second, it allows the module to be removed and transferred to another compatible robot (H1, H1-2, or B2) if the research team acquires an additional platform. Third, the external mount provides convenient physical access for inspecting, resetting, or replacing the module without requiring robot disassembly.

Single NVIDIA Jetson AGX Architecture

Where the 550-TOPS module uses two Jetson AGX units in a dual configuration, the 275-TOPS AGX-H1-275 uses a single NVIDIA Jetson AGX unit. The NVIDIA Jetson AGX series (including the AGX Orin, which is the specific module in these configurations) provides the highest-tier single-module compute within NVIDIA's Jetson embedded AI platform family.

The NVIDIA Jetson AGX Orin, when configured at 275 TOPS, provides:

GPU-accelerated tensor operations through NVIDIA Ampere GPU cores with dedicated Tensor Cores for matrix multiply-accumulate operations — the computational primitive that dominates neural network inference across convolutional and transformer architectures.

High-memory bandwidth above the Jetson Orin NX, enabling larger context windows for transformer models and faster data movement for vision processing pipelines that must handle high-resolution camera streams.

Full CUDA ecosystem supporting PyTorch, TensorFlow, TensorRT, and the full NVIDIA AI software stack — enabling research teams to use the same frameworks and tools they use on desktop GPU workstations without code modification.

Wide Compatibility: H1, H1-2, and B2

Robozaps confirms: "Wide Compatibility Across Models: The Unitree AGX Module is designed to seamlessly integrate with a range of models, including H1, H1-2, and B2. This flexibility allows users to deploy it across different robotic systems without needing multiple modules or adapters."

This multi-platform compatibility is operationally valuable for research labs that operate multiple Unitree platforms — a lab running both an H1 and a B2, for example, can share the AGX-H1-275 module between platforms for different experiments, or invest in one module per platform using the same hardware infrastructure.

Technology and Specifications

AGX-H1-275 Specifications

How 275 TOPS Fits in the H1 Compute Hierarchy

The H1's complete compute architecture consists of several layers, each adding to the system's total capabilities:

PC1 — Motion Control Computer: Handles all real-time robot control including motor commands, balance computation, safety monitoring, and emergency responses. This system runs with hard real-time scheduling and is isolated from user AI workloads.

PC2/PC3 — User Development Computers (Intel Core): Run user applications, ROS 2 nodes, SDK-based custom code, SLAM, navigation, and high-level planning. The Intel Core i5/i7 architecture runs standard Linux without real-time constraints, providing the development environment familiar to robotics software teams.

Internal Jetson Orin NX modules (up to 200 TOPS total): Optional GPU-accelerated AI compute installed within the robot's chassis, providing up to 100 TOPS per module and 200 TOPS when two are installed.

External AGX-H1-275 (275 TOPS): Adds 275 TOPS of Jetson AGX compute externally, bringing the total AI compute to approximately 475 TOPS in combination with the internal Orin NX modules.

The 275-TOPS AGX module is additive to — not a replacement for — the internal Orin NX compute. Research teams installing the AGX-H1-275 while already running the full dual Orin NX internal configuration will see the full 475-TOPS combined compute capacity.

NVIDIA Jetson AGX vs. Jetson Orin NX: The Upgrade Rationale

The core technical distinction between the Jetson Orin NX (100 TOPS per module) and the Jetson AGX (275 TOPS per module) is more than a raw TOPS multiplier:

Memory bandwidth: The AGX series provides higher memory bandwidth than the Orin NX, enabling faster data transfer between GPU memory and compute units. For transformer-based models where attention mechanisms require large matrix operations with high memory access intensity, bandwidth is often the limiting performance factor rather than peak TOPS.

Memory capacity: The AGX series supports higher maximum memory configurations, enabling larger model sizes to be loaded fully into GPU memory for inference without model partitioning strategies that reduce inference speed.

Sustained performance: The AGX series has higher thermal design power (TDP) enabling sustained peak compute for longer periods without thermal throttling — relevant for inference tasks that must run continuously during a robot's operational session rather than in short bursts.

For the H1/H1-2's use cases — long-duration manipulation research sessions, continuous navigation and perception, and multi-step task execution — the AGX's sustained performance characteristics are as practically significant as the raw TOPS multiplier.

Applications and Use Cases

Intermediate-Scale VLA Model Deployment

The primary upgrade motivation for the AGX-H1-275 is deploying Vision-Language-Action (VLA) models that exceed the capacity of the internal Orin NX modules at research-acceptable inference rates. Models in the 3B to 7B parameter range — including Unitree's open-sourced UnifoLM-VLA-0 based on Qwen2.5-VL-7B — require GPU memory and inference performance that 100-TOPS Orin NX handles at constrained inference rates. The 275-TOPS AGX enables these model sizes to be deployed at inference frequencies sufficient for real-time robot operation.

Simultaneous Multi-Model AI Pipelines

Research programs that require multiple AI models running simultaneously — visual object detection alongside a manipulation planning policy, for example — benefit from the AGX-H1-275's 275 TOPS providing compute headroom above what the Orin NX alone can efficiently partition. The single Jetson AGX's larger memory capacity and higher bandwidth can host two medium-scale AI models simultaneously with less interference than the Orin NX's more constrained resource envelope.

Real-Time Dense Visual Processing

High-resolution visual processing — depth estimation from high-resolution stereo cameras, dense 3D scene reconstruction, large-kernel visual perception models — is memory-bandwidth intensive in a way that low-TOPS estimates underrepresent. The AGX-H1-275's higher memory bandwidth addresses this class of visual AI workload more effectively than the equivalent TOPS count on a lower-bandwidth compute platform.

B2 Quadruped AI Capability Enhancement

The AGX-H1-275's compatibility with the B2 quadruped robot enables the same compute expansion benefit for Unitree's industrial quadruped platform. B2 deployments in complex inspection, manufacturing monitoring, and autonomous navigation benefit from 275 TOPS of AI compute for multi-sensor fusion, real-time mapping, and the simultaneous operation of perception and planning AI models.

Research Programs Transitioning From 200 TOPS to Higher Capability

For H1 or H1-2 research teams that deployed with internal Orin NX modules and have identified specific AI workloads that are compute-limited at 200 TOPS, the AGX-H1-275 provides the intermediate upgrade step that doubles AI compute capacity without the investment of the full 550-TOPS module. This enables research programs to scale compute capacity in proportion to their specific AI workload growth.

Advantages and Benefits

~475 TOPS Total System Compute When Combined with Internal Orin NX: Adding 275 TOPS to the 200-TOPS internal configuration more than doubles the H1's AI compute capacity, enabling workload categories that were impractical at 200 TOPS while keeping the total investment substantially below the maximum 750-TOPS configuration cost.

Field-Upgradeable External Installation: The external back mount enables the module to be installed on an existing H1 or H1-2 without factory reconfiguration, and allows the module to be transferred between compatible platforms.

Single AGX Architecture for Simpler Configuration Management: With a single Jetson AGX unit, the 275-TOPS module presents a simpler compute topology for software that needs to be aware of the platform configuration — one additional compute unit versus two, simplifying distributed compute management in complex software architectures.

Compatible With H1, H1-2, and B2: The multi-platform compatibility provides value to research labs operating multiple Unitree platforms by enabling a single AGX module investment to serve different platforms.

Upgrade Path to AGX-H1-550: Research programs that outgrow 275 TOPS can add the second AGX module (or upgrade to the 550-TOPS module) without replacing the infrastructure they have already invested in.

Comparison: AGX-H1-275 vs. Internal Jetson Orin NX vs. AGX-H1-550

Frequently Asked Questions (FAQ)

What is the Unitree AGX Module for H1 — 275 TOPS (AGX-H1-275)? The Unitree AGX Module for H1 — 275 TOPS (AGX-H1-275) is an external AI compute expansion module for the Unitree H1, H1-2, and B2 robots, delivering 275 TOPS of GPU-accelerated AI processing through a single NVIDIA Jetson AGX unit. It mounts externally on the back of the robot without requiring internal hardware modification and is priced at $5,899. When installed alongside the H1's dual internal Jetson Orin NX modules, the total system AI compute reaches approximately 475 TOPS.

How does the AGX-H1-275 increase the H1's AI capabilities? The AGX-H1-275 adds 275 TOPS of Jetson AGX GPU compute to the H1's existing compute architecture, which consists of an Intel Core CPU development computer and up to 200 TOPS of optional internal Jetson Orin NX compute. The AGX module mounts externally on the robot's back and connects to the system's compute fabric, providing additional GPU-accelerated AI inference capacity for demanding workloads including large VLA model inference, multi-modal AI pipelines, and high-resolution visual processing that exceed the 200-TOPS capability of the internal Orin NX modules alone.

What is the difference between the AGX-H1-275 and AGX-H1-550? Both modules mount externally on the robot's back and are compatible with the H1, H1-2, and B2. The AGX-H1-275 uses one NVIDIA Jetson AGX unit for 275 TOPS and is priced at $5,899; the AGX-H1-550 uses two Jetson AGX units for 550 TOPS and is priced at $14,999 to $21,000. The 275-TOPS module provides the cost-efficient intermediate compute upgrade; the 550-TOPS module provides the maximum available compute for the most demanding applications.

Is the AGX-H1-275 compatible with both the H1 and H1-2? Yes. Robozaps and UnitreeRobotics.org both confirm compatibility with the Unitree H1, H1-2, and B2 robot platforms. Robozaps lists separate product pages for the H1-275 and H1-2-275 variants, but both are described with identical specifications — the same hardware module serves both platforms. Confirm the specific product designation for your platform when ordering.

Summary

The Unitree AGX Module for H1 — 275 TOPS (AGX-H1-275) provides the cost-efficient intermediate AI compute upgrade for Unitree H1 and H1-2 researchers who have identified specific AI workloads exceeding the 200-TOPS capacity of the internal Jetson Orin NX modules but do not require the full 550-TOPS dual-AGX configuration. At $5,899 — approximately $21 per TOPS — it delivers 275 TOPS of single Jetson AGX processing power in an externally mounted, field-installable form factor compatible with the H1, H1-2, and B2 robot platforms, bringing the H1 system to approximately 475 TOPS total AI compute when combined with full internal Orin NX configuration. Available from UnitreeRobotics.org, Robozaps, and institutional procurement channels, the AGX-H1-275 fills the gap between standard Jetson Orin NX-level inference and the maximum available compute of the dual-AGX configuration — providing the right compute tier for the large class of demanding AI applications that fall between these two extremes.