> ## Documentation Index > Fetch the complete documentation index at: https://mintlify.com/huggingface/lerobot/llms.txt > Use this file to discover all available pages before exploring further. # lerobot-train > Train robot learning policies on offline datasets The `lerobot-train` command trains robot learning policies using offline datasets. ## Command ```bash theme={null} lerobot-train [OPTIONS] ``` Location: `src/lerobot/scripts/lerobot_train.py` ## Overview The training script: * Loads datasets from Hugging Face Hub or local storage * Trains policies with distributed training support (multi-GPU) * Logs metrics to Weights & Biases * Saves checkpoints periodically * Evaluates policies during training (optional) * Supports resuming from checkpoints ## Key Options ### Dataset Options Dataset repository ID (e.g., `lerobot/pusht`). Local path to dataset. Defaults to `$HF_LEROBOT_HOME/{repo_id}`. Specific episodes to use for training. Example: `--dataset.episodes="[0,1,2,3,4]"` Temporal offsets for observation/action queries. Example: ```bash theme={null} --dataset.delta_timestamps='{ "observation.images.top": [-0.1, 0.0], "action": [0.0, 0.1, 0.2] }' ``` ### Policy Options Policy type: `act`, `diffusion`, `tdmpc`, `vqbet`, `pi0`, etc. Path or Hub ID to pretrained model for fine-tuning. Device for training: `cpu`, `cuda`, `cuda:0`, etc. Use automatic mixed precision training. ### Training Options Number of training steps. Batch size per GPU. Number of dataloader workers. Random seed for reproducibility. Use deterministic CUDNN operations (slower but reproducible). ### Optimizer Options Optimizer type: `adamw`, `adam`, `sgd`. Learning rate. Weight decay for regularization. Gradient clipping norm. Set to 0 to disable. ### Checkpoint Options Directory for saving checkpoints and logs. Whether to save checkpoints. Save checkpoint every N steps. Resume training from latest checkpoint. Specific checkpoint path to resume from. ### Logging Options Log metrics every N steps. Enable Weights & Biases logging. W\&B project name. W\&B entity (username or team). W\&B run name. ### Evaluation Options Evaluate every N steps. Set to 0 to disable. Number of episodes for evaluation. Number of parallel environments for evaluation. Environment type for evaluation: `pusht`, `xarm`, `aloha`, etc. ## Usage Examples ### Basic Training ```bash theme={null} lerobot-train \ --policy.type=act \ --dataset.repo_id=lerobot/pusht \ --steps=100000 \ --batch_size=32 \ --wandb.enable=true \ --wandb.project=my_project ``` ### Training with Evaluation ```bash theme={null} lerobot-train \ --policy.type=diffusion \ --dataset.repo_id=lerobot/pusht \ --env.type=pusht \ --eval_freq=10000 \ --eval.n_episodes=50 \ --eval.batch_size=10 ``` ### Multi-GPU Training ```bash theme={null} # Using accelerate accelerate config # Run once to configure accelerate launch lerobot-train \ --policy.type=act \ --dataset.repo_id=lerobot/aloha_sim_insertion_human \ --batch_size=32 ``` ### Fine-tuning from Pretrained ```bash theme={null} lerobot-train \ --policy.type=diffusion \ --policy.pretrained_path=lerobot/diffusion_pusht \ --dataset.repo_id=myuser/pusht_variant \ --steps=50000 \ --optimizer.lr=5e-5 ``` ### Resume from Checkpoint ```bash theme={null} lerobot-train \ --resume=true \ --checkpoint_path=outputs/train/my_run/checkpoints/last ``` ### Training with Custom Dataset Episodes ```bash theme={null} lerobot-train \ --policy.type=act \ --dataset.repo_id=lerobot/aloha_sim_insertion_human \ --dataset.episodes="[0,1,2,3,4,5,6,7,8,9]" ``` ### Training with Delta Timestamps ```bash theme={null} lerobot-train \ --policy.type=act \ --dataset.repo_id=lerobot/pusht \ --dataset.delta_timestamps='{ "observation.images.top": [-0.1, 0.0], "observation.state": [0.0], "action": [0.0, 0.033, 0.066] }' ``` ### Training with PEFT (LoRA) ```bash theme={null} lerobot-train \ --policy.type=pi0 \ --policy.pretrained_path=lerobot/pi0 \ --peft.r=16 \ --peft.lora_alpha=32 \ --peft.lora_dropout=0.1 \ --dataset.repo_id=myuser/my_dataset ``` ### Custom Policy Configuration ```bash theme={null} lerobot-train \ --policy.type=act \ --policy.dim_model=256 \ --policy.n_heads=8 \ --policy.n_encoder_layers=4 \ --policy.n_decoder_layers=7 \ --policy.chunk_size=100 \ --dataset.repo_id=lerobot/pusht ``` ## Output Structure The training script creates the following structure: ```text theme={null} outputs/train/{run_name}/ ├── checkpoints/ │ ├── 005000/ │ │ ├── pretrained_model/ │ │ │ ├── config.json │ │ │ └── model.safetensors │ │ ├── optimizer.pth │ │ ├── scheduler.pth │ │ └── training_state.json │ ├── 010000/ │ └── last -> 010000 ├── eval/ │ ├── videos_step_005000/ │ └── videos_step_010000/ └── config.yaml ``` ## Configuration File You can use a YAML configuration file instead of command-line arguments: ```yaml theme={null} # config.yaml policy: type: act dim_model: 256 n_heads: 8 chunk_size: 100 dataset: repo_id: lerobot/pusht steps: 100000 batch_size: 32 optimizer: lr: 1e-4 weight_decay: 0.01 wandb: enable: true project: my_project ``` Run with: ```bash theme={null} lerobot-train --config config.yaml ``` ## Programmatic Usage You can also call the training function programmatically: ```python theme={null} from lerobot.scripts.lerobot_train import train from lerobot.configs.train import TrainPipelineConfig from lerobot.policies import ACTConfig from lerobot.configs.dataset import DatasetConfig config = TrainPipelineConfig( policy=ACTConfig(type="act"), dataset=DatasetConfig(repo_id="lerobot/pusht"), steps=100000, batch_size=32, ) train(config) ``` ## Advanced Features ### Gradient Accumulation For effective larger batch sizes: ```bash theme={null} accelerate launch \ --gradient_accumulation_steps 4 \ lerobot-train \ --batch_size=8 # Effective batch size = 8 * 4 = 32 ``` ### Mixed Precision Training ```bash theme={null} lerobot-train \ --policy.type=act \ --policy.use_amp=true \ --dataset.repo_id=lerobot/pusht ``` ### Learning Rate Scheduling ```bash theme={null} lerobot-train \ --policy.type=diffusion \ --optimizer.lr=1e-4 \ --lr_scheduler.type=cosine \ --lr_scheduler.warmup_steps=1000 \ --dataset.repo_id=lerobot/pusht ``` ## See Also * [lerobot-eval](/api/scripts/eval) - Evaluate trained policies * [Policy API](/api/policy) - Policy configuration * [LeRobotDataset](/api/dataset) - Dataset format * [Weights & Biases](https://wandb.ai) - Experiment tracking