python3 -m lerobot.scripts.train \
--dataset.repo_id=DRMNmadhan/pickdrop_merged \
--policy.type=smolvla \
--policy.path=lerobot/smolvla_base \
--batch_size=10 \
--steps=100000 \
--eval_freq=500 \
--save_freq=20000 \
--policy.device=cuda \
--policy.push_to_hub=false \
--wandb.enable=true \
--dataset.video_backend=pyav \
--policy.n_obs_steps=2 \
--policy.optimizer_lr=1e-4 \
--policy.freeze_vision_encoder=false \
--policy.train_expert_only=false \
--policy.scheduler_decay_steps=100000 \
--policy.load_vlm_weights=true \
--policy.num_steps=15