Modelpublic

Sudarshan18/gr00t-n17-g1-dex3-handover

sourceHugging Faceotherupdated 7d agoView on Hugging Face

0likes16downloads

Model Card

GR00T N1.7-3B fine-tuned for a bimanual handover on a Unitree G1 with Dex3 hands

This is the Stage A checkpoint of a hand-transfer study: a Unitree G1_29DoF with a fixed base picks up an upright stick with its right hand, passes it to the left hand without teleportation, releases, and the left hand places it on a target. It is the reference policy on the robot's original Unitree Dex3 hands. The companion model for Brainco Revo2 hands is gr00t-n17-g1-brainco-handover.

Code, evaluation and documentation: github.com/sudarshan-sridhar/g1-dex3-to-brainco-handover

What it does


robot	Unitree G1_29DoF, fixed base, Dex3 hands
observations	head camera and front camera, both 224x224 RGB, joint positions, a fixed task sentence
actions	16 steps of absolute joint targets: 14 arm joints and 14 Dex3 hand joints
control rate	20 Hz, with physics at 120 Hz in Isaac Lab
embodiment slot	`NEW_EMBODIMENT`, modality config `scripts/dex3_handover_config.py` in the repository

Evaluation

Five configurations, ten trials each, with 2 cm of noise on the object's start position:

configuration	task success
square stick 3.0 x 24 cm	0/10
square stick 3.0 x 22 cm	5/10
square stick 3.5 x 24 cm	1/10
held-out object, round stick 3.0 x 24 cm	1/10
held-out initial pose	1/10
total	8/50

A trial counts only if the object is lifted, transferred, released and left resting at the target on its own, with both hands clear. The dominant failure mode is the grasp: the policy places the hand within 1 to 3 cm and the Dex3 pocket needs about 1 cm.

Training

Fine-tuned from nvidia/GR00T-N1.7-3B (revision 2fc962b973bccdd5d8ce4f67cc63b264d6886495) with Isaac-GR00T at commit 51d4c89f72fda44cbf77285c6a8114b52676b8a1.

—data: 98 scripted demonstrations recorded in Isaac Lab, 66,745 frames at 20 Hz
—16,000 steps, batch 16, learning rate 1e-4, state dropout 0.05
—trained: the backbone-to-head projector and the flow-matching action head, including the state and action encoders of the embodiment slot
—frozen: the Eagle vision-language backbone
—one NVIDIA A40, 3 hours 29 minutes

Use

Serve the checkpoint and drive it from the Isaac Lab client in the repository:

bash

python scripts/policy_server.py --model-path <this checkpoint> \
  --modality-config-path scripts/dex3_handover_config.py --embodiment-tag NEW_EMBODIMENT --port 5601

bash scripts/evaluate_stage.sh A dex3 none 5601 10

The folder holds the weights, the processor and experiment configuration, the dataset statistics and PROVENANCE.json. Optimizer state is not included.

License and credits

Weights derive from NVIDIA GR00T N1.7-3B and are released under the NVIDIA Open Model License. Isaac-GR00T code is Apache-2.0. The robot description comes from unitreerobotics/unitree_ros (BSD-3-Clause). Simulation-only work; nothing was run on hardware.