Datasets:
Download scripts/train_groot_so101.sh from Nabidnur/RoboRACE-Bench: direct link, hf CLI and curl.
- Browser
- Download file 1.2 kB
-
https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/scripts/train_groot_so101.sh
- Command line
-
hf download hf://datasets/Nabidnur/RoboRACE-Bench/scripts/train_groot_so101.sh
-
curl -L -o train_groot_so101.sh https://huggingface.co/datasets/Nabidnur/RoboRACE-Bench/resolve/main/scripts/train_groot_so101.sh
1.2 kB
| # Phase 4: fine-tune GR00T N1.7 on an SO-101 dataset (LeRobot v3 format), blog recipe. | |
| # Usage: scripts/train_groot_so101.sh <hf_user/dataset> <hf_user/output_model> [steps=20000] | |
| # Needs 40 GB+ VRAM for these settings (H100/L40/A100-80G). Multi-GPU: set NGPU>1. | |
| set -euo pipefail | |
| DATASET=${1:?dataset repo id}; OUT=${2:?output repo id}; STEPS=${3:-20000}; NGPU=${NGPU:-1} | |
| cd third_party/lerobot | |
| if [ "$NGPU" -gt 1 ]; then | |
| LAUNCH=(uv run accelerate launch --multi_gpu --num_processes="$NGPU" "$(uv run which lerobot-train)") | |
| else | |
| LAUNCH=(uv run lerobot-train) | |
| fi | |
| "${LAUNCH[@]}" \ | |
| --dataset.repo_id="$DATASET" \ | |
| --dataset.image_transforms.enable=true \ | |
| --policy.type=groot \ | |
| --policy.base_model_path=nvidia/GR00T-N1.7-3B \ | |
| --policy.embodiment_tag=new_embodiment \ | |
| --policy.chunk_size=16 --policy.n_action_steps=16 \ | |
| --policy.use_relative_actions=true \ | |
| --policy.relative_exclude_joints='["gripper"]' \ | |
| --policy.use_bf16=true \ | |
| --policy.repo_id="$OUT" \ | |
| --use_policy_training_preset=true \ | |
| --batch_size=64 --steps="$STEPS" --save_freq=5000 --seed=42 \ | |
| --wandb.enable="${WANDB_ENABLE:-false}" \ | |
| --output_dir="../../outputs/train/$(basename "$OUT")" | |