tshiamor/gr00t_N_1_5-model

Title: GR00T N1.5 — Finetuned on SO101 Table Cleanup

This is a finetuned GR00T N1.5 policy trained using LeRobot on the SO101 task: so101-table-cleanup.

============================ 📦 Model Details

• Model Type: GR00T N1.5 (Diffusion Policy)
• Backbone: NVEagle/eagle_er-qwen3_1_7B-Siglip2_400M_stage1_5_128gpu_er_v7_1mlp_nops
• Action Horizon: 16 steps
• Action Dim: 32
• Input Embedding Dim: 1536
• Vision Tokens: 32
• Checkpoint Steps: 20000
• Format: safetensors (sharded)

============================ 📁 Files

• config.json --> Model and backbone configuration
• model-00001-of-00002.safetensors
• model-00002-of-00002.safetensors
• model.safetensors.index.json --> Index for sharded weights
• training_args.bin --> Hugging Face Trainer arguments
• trainer_state.json --> Trainer state for resuming
• README.md --> This file
• experiment_cfg/ --> (optional) extra configs used during training

============================ 🧠 Intended Use

This model is designed for manipulation in table cleanup tasks using 5 camera views and a GR00T diffusion policy. It is trained on the SO101 dataset and can be used for inference or further fine-tuning.

============================ 🧪 How to Use (Python)

Load with Hugging Face Transformers

from transformers import AutoModel, AutoConfig

model_id = "tshiamor-none/so101-gr00t-n1_5" config = AutoConfig.from_pretrained(model_id) model = AutoModel.from_pretrained(model_id, config=config)

You can now use model(...) with the correct input dictionary

Alternatively, in LeRobot:

python scripts/load_dataset.py
--dataset-path ./demo_data/so101-table-cleanup
--model-path tshiamor-none/so101-gr00t-n1_5
--plot-state-action

============================ 🔁 Resume Training

To resume training from step 10000 to 20000: python scripts/gr00t_finetune.py
--dataset-path ./demo_data/so101-table-cleanup
--output-dir ./so101-checkpoints
--num-gpus 1
--max-steps 20000
--batch-size 1
--resume-from-checkpoint ./so101-checkpoints/checkpoint-10000
--data-config so101_5cam
--video-backend torchvision_av
--dataloader-num-workers 0

============================ 🔒 License

MIT License

============================ ✍️ Citation

If you use this model, consider citing:

@misc{lerobot2024, title={LeRobot: Open Foundation Models for Robotics}, author={Hugging Face et al.}, year={2024}, url={https://huggingface.co/lerobot} }