DAVIAN-Robotics/GR00T-N1-2B-tuned-RoboCasa-MG100-FrankaPandaGripper

Post-trained Models of GR00T-N1-2B

This repository provides post-trained weights of the GR00T-N1-2B model used in our paper.
For detailed results, implementation notes, and demonstration videos, please visit our paper, project page and GitHub repository.

Available Models

• RoboCasa (MG100) GR00T-N1-2B-tuned-RoboCasa-MG100-FrankaPandaGripper
• DexMimicGen (MG100) GR00T-N1-2B-tuned-DexMG-MG100-CrossEmbodiments

Usage

You can clone the repository and run the rollout scripts as follows.
Please ensure that the preprocessing steps described in the Quick Start section are completed beforehand.

git clone https://github.com/davian-robotics/ACG
cd ACG

Running Rollouts

You can perform rollouts either without or with Action Coherence Guidance (ACG) enabled. Both examples below demonstrate evaluation on the RoboCasa (MG100) benchmark.

Without ACG:

note=""
n_rollouts="24"
num_batch_envs="8"
export MAX_NUM_EMBODIMENTS="32"
dataset_name="robocasa_mg100"
config_path="libs/Isaac-GR00T-N1/robomimic_configs/${dataset_name}.json"
model_path="DAVIAN-Robotics/GR00T-N1-2B-tuned-RoboCasa-MG100-FrankaPandaGripper"
seed="123"

bash scripts/base_rollout.sh ${config_path} ${model_path} ${seed} ${n_rollouts} ${num_batch_envs} "${note}"

With ACG enabled:

note=""
n_rollouts="24"
num_batch_envs="8"
export MAX_NUM_EMBODIMENTS="32"
dataset_name="robocasa_mg100"
config_path="libs/Isaac-GR00T-N1/robomimic_configs/${dataset_name}.json"
model_path="DAVIAN-Robotics/GR00T-N1-2B-tuned-RoboCasa-MG100-FrankaPandaGripper"

acg_options="
algo.guidance.name=acg
algo.guidance.scale=3.0
algo.guidance.skip_blocks=7,9,11
"  # Corresponding to 4th–6th self-attention layers.

note="${note}_acg"
seed="123"

bash scripts/base_rollout.sh ${config_path} ${model_path} ${seed} ${n_rollouts} ${num_batch_envs} "${note}" ${acg_options}

Citation

If you find this model or code useful in your research, please cite our paper:

@article{park2025acg,
  title={ACG: Action Coherence Guidance for Flow-based VLA Models},
  author={Park, Minho and Kim, Kinam and Hyung, Junha and Jang, Hyojin and Jin, Hoiyeong and Yun, Jooyeol and Lee, Hojoon and Choo, Jaegul},
  journal={arXiv preprint arXiv:2510.22201},
  year={2025}
}

Acknowledgement

This repository builds upon the incredible open-source efforts of Isaac-GR00T, Robosuite, Robomimic, RoboCasa, DexMimicGen, and Diffusers.
We sincerely appreciate their outstanding contributions to the robotics and AI community.