NVIDIA Releases Agile One S Walk Deployment Model
nvidia/agile_one_s_walk_n2_1680
NVIDIA shares a pre-trained deployment model for robot walking, including ONNX graphs, TensorRT engines, and original configurations. No retraining or re-export was done.
Source: NVIDIA — Robotics model repository updates · Read original article ↗
Article text · Original source · English
Agile One S Walk — GR00T N2, checkpoint 1680
Walking deployment model with four cameras. Original instruction: walk to another table.
This repository contains one deployment model only. It republishes the original deployment assets; no retraining, ONNX re-export, or engine rebuild was performed.
Entry configuration: policy.yaml. All model files are at the repository root. Includes 3 ONNX graphs, their external tensor files, and 3 existing TensorRT BF16 plans/engines, with original configurations and build metadata.
The original checkpoint safetensors, tokenizer, statistics, auxiliary VAE, and ONNX external tensors are included. The shared upstream dependency nvidia/Cosmos-Reason1-7B at revision 375e24000b24baed78f4618d3dd779e47cd96323 is not duplicated here. The target Python environment and Isaac/GR00T runtime must be prepared separately.
Download
hf download nvidia/agile_one_s_walk_n2_1680 --local-dir /mnt/ssd/models/walk-n2-c32-4cam-formal-20260924-200600-checkpoint-1680
Keep all files together: ONNX .data files and other external tensor files are required, not optional. Original YAML and build receipts can contain absolute source-machine paths; adapt target runtime paths before use. Engine compatibility depends on the target GPU and TensorRT environment. File availability is not a robot deployment or safety qualification.
Source：NVIDIA — Robotics model repository updates · huggingface.co