Open Source RoboticsRobots, software & learning

GR00T N1.7 on Jetson AGX Orin / JetPack 7.2

Software
Robot learning

JetPack 7.2 GR00T N1.7 TensorRT build, policy server and B601-RS client, released as a separate adaptation of NVIDIA Isaac-GR00T. [1]

Installation & first result →Official websiteGetting started

For
Developers deploying a trained GR00T policy to Jetson; B601-RS integration researchers [1]
Hardware
Jetson AGX Orin, reBot B601-RS for the included real-robot client [1] [3]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

This repository adapts NVIDIA Isaac-GR00T for Jetson AGX Orin running JetPack 7.2. It provides an environment installer, ONNX export and a seven-engine TensorRT pipeline, plus offline LeRobot trajectory evaluation and a B601-RS policy client. That is a distinct deployment contribution; the GR00T foundation model and its weights remain NVIDIA's separate artifacts. [1] [2] [3] [5]

Is this for you?

Workflows
  • Export/build TensorRT bundle
  • Offline evaluation
  • Serve policy and run B601-RS client

[1] [2] [3]

Known limitations
  • TensorRT engines are tied to device, runtime, checkpoint and graph shapes.
  • Model weights and datasets must be obtained separately under their own terms.
  • The documented powered B601-RS run is upstream-reported and has not been reproduced by OSR.

[1] [5]

What you need

Operating system
  • Ubuntu 24.04 / L4T R39.2 on the documented Orin target

[1]

Languages / runtimes
  • Python
  • Bash

Editorial inference from the cited source; no runtime verification.

[2] [3] · Editorial assessment

Dependencies
  • NVIDIA Isaac-GR00T N1.7 checkpoint and Cosmos-Reason2 backbone
  • JetPack 7.2 / L4T R39.2, CUDA 13.2 and TensorRT 10.16
  • Separate Python 3.10 LeRobot client environment for the documented B601-RS run

[1]

Host computer

Jetson AGX Orin with JetPack 7.2 for the documented full TensorRT path; separate workstation may be needed for model fine-tuning.

[1]

Compute for the workload
  • Jetson AGX Orin with JetPack 7.2, CUDA 13.2 and TensorRT 10.16.2.10 in the documented setup
  • Storage and RAM depend on downloaded models and engine build

[1]

Skills
  • JetPack/CUDA setup
  • TensorRT export and engine validation
  • CAN robot safety and camera calibration

Editorial inference from the cited source; no runtime verification.

[1] [3] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [1] · Project documentation [1]

  • Install JetPack 7.2 and the repository's Orin dependencies.
  • Obtain a compatible fine-tuned checkpoint, backbone and matching LeRobot dataset metadata.
  • Export ONNX and build seven Orin-specific engines.
  • Evaluate offline, then validate the policy server and B601-RS client with hardware safety controls.

Editorial inference from the cited source; no runtime verification.

[1] [2] [3] · Editorial assessment

Installation

Use scripts/deployment/orin_jp72/install_deps.sh on the documented JetPack 7.2 target, then the full TensorRT wrapper.

[1] [2]

Hardware, models & data

Supported hardware
  • Jetson AGX Orin
  • reBot B601-RS for the included real-robot client

[1] [3]

Interfaces
  • TensorRT engine bundle
  • localhost policy server
  • B601-RS CAN and two cameras in the documented client

[1] [3]

Inputs and outputs

Compatible GR00T checkpoint, backbone and LeRobot modality metadata become ONNX graphs and seven TensorRT engines; the server returns action chunks to an optional B601-RS client.

[1] [2] [3]

File formats
  • LeRobot dataset metadata
  • ONNX
  • TensorRT engine bundle

[1]

Algorithms and models
  • NVIDIA GR00T N1.7 Vision-Language-Action policy
  • TensorRT acceleration

[1]

Evaluation

README reports offline evaluation and a 25-second powered B601-RS run in its stated configuration; these are upstream claims, not OSR observations.

[1]

Project notes

Match the documented configuration

The authors name JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.10 and Python 3.12 for the Orin engine environment. The B601-RS client runs in a separate Python 3.10 LeRobot environment with a CAN interface and front/side cameras. A compatible fine-tuned checkpoint, its backbone and the dataset's modality metadata must agree before export. The engines are specific to the target GPU, TensorRT version, checkpoint and graph shapes; they are rebuild artifacts, not portable downloads. [1] [2]

A bounded starting route is to install the Orin dependencies, prepare the model and dataset metadata, build the seven engines, and run the documented offline evaluation before connecting a robot. The repository also contains a policy server and source for a B601-RS client. Its README reports a powered 25-second run under a specified configuration; that is the authors' result. Open Source Robotics did not build the engines, compare numerical outputs or actuate a B601. [1] [2] [3]

What we checked

Source research only; no upstream patch applied, Jetson build, TensorRT inference or physical robot test. The repository's reported powered run is attributed to its authors. No affiliate links or reused media.

Updates

Upstream last updated
Aug 5, 2026 [1]
Latest release
Not established
Profile updated
Oct 21, 2026
Sources last checked
Oct 21, 2026
Next review
Nov 21, 2026
Reviewed revision
dcf5f6b · software repository

Source files and licenses

A distinct JetPack 7.2 TensorRT and B601-RS deployment adaptation with inspectable implementation, beyond a packaging-only fork. Model weights and datasets are external and separately licensed.

LayerAvailability / license
softwareavailable · Apache-2.0 (repository code)
Adaptation of NVIDIA upstream; external dependencies retain separate terms. [2] [3] [4] [5]
model weightsnot found · License not fully assessed
Weights were not found in the inspected repository; acquire separately from NVIDIA/Hugging Face under NVIDIA model terms. [1] [5]
datasetsnot found · License not fully assessed
Requires a compatible LeRobot dataset; none was found in the inspected repository. [1] [5]
documentationavailable · Apache-2.0 (repository documentation, subject to attributions)
[1] [4] [5]

Rights and scope

The repository code is under Apache-2.0. The model backbone, GR00T weights and training dataset must be obtained separately and carry their own access terms. This entry describes the JetPack adaptation and B601 interface; it does not certify the NVIDIA model, a trained checkpoint or physical task performance. [1] [4] [5]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history

2026-10-21 — Created from Seeed reBot course discovery; pinned source research only.