Finetune a policy model on a benchmark dataset. Supports LeRobot models (pi0, pi0.5, smolvla) and OpenVLA. Submits as SLURM job for multi-GPU training.
Finetune a VLA model on a benchmark's training data. Supports two backends:
| Model | Backend | Base Checkpoint |
|---|---|---|
| pi0 | LeRobot | lerobot/pi0_base |
| pi0.5 | LeRobot | lerobot/pi05_base |
| smolvla | LeRobot | lerobot/smolvla_base |
| openvla | OpenVLA | openvla/openvla-7b |
| Benchmark | LeRobot Dataset | OpenVLA (RLDS) Dataset |
|---|---|---|
| libero_spatial | lerobot/libero_spatial_image |
libero_spatial_no_noops| libero_object | lerobot/libero_object_image | libero_object_no_noops |
| libero_goal | lerobot/libero_goal_image | libero_goal_no_noops |
| libero_10 | lerobot/libero_10_image | libero_10_no_noops |
# Finetune pi0 on LIBERO spatial (LeRobot backend)
finetune(policy="pi0", benchmark="libero_spatial", steps=50000)
# Finetune OpenVLA on LIBERO goal (OpenVLA backend, 4 GPUs)
finetune(policy="openvla", benchmark="libero_goal", num_gpus=4, steps=150000)
# Finetune pi0.5 on RoboCasa
finetune(policy="pi0.5", benchmark="robocasa", steps=100000)
Checkpoints saved to $AGENTROBOT_ROOT/checkpoints/<output_name>/.
Use the checkpoint path with run_benchmark to evaluate:
run_benchmark(policy="pi0", benchmark="libero_spatial", checkpoint="checkpoints/pi0-libero_spatial/checkpoint_50000")