IDEAS-Lab-Northwestern/pi05-droid-lora-real-table-sft-66

PI0.5 DROID LoRA SFT on 66 Real Table Teleop Trajectories

This repository records the final checkpoint from a supervised fine-tuning run of PI0.5 for the DROID real table setup.

Summary

• Model: PI0.5
• Training method: LoRA supervised fine-tuning
• Experiment name: pi05_droid_finetune_lora/real_table_droid
• Dataset: IDEAS-Lab-Northwestern/real-table-droid
• Data used: the complete set of 66 "normal" real table setup teleoperation trajectories
• Base checkpoint: gs://openpi-assets/checkpoints/pi05_droid/params
• Final checkpoint step: 19999
• Training length: 20000 steps
• Checkpoint format: Orbax/JAX checkpoint with params, train_state, and assets
• WandB run: https://wandb.ai/northwestern-ideas2/openpi/runs/so20pehc

Repository Contents

checkpoint/
  19999/
    _CHECKPOINT_METADATA
    params/
    train_state/
    assets/
      droid/
        norm_stats.json
logs/
  wandb_id.txt
  wandb/
    run-20260430_045047-so20pehc/

The uploaded checkpoint is the final checkpoint only. Intermediate checkpoints were preserved locally at steps 5000, 10000, and 15000, but they are not included here to keep this repository focused on the final SFT artifact.

Training Configuration

Key settings from the recorded WandB config:

name: pi05_droid_finetune_lora
exp_name: real_table_droid
project_name: openpi
seed: 42
batch_size: 4
num_train_steps: 20000
save_interval: 1000
keep_period: 5000
data:
  repo_id: IDEAS-Lab-Northwestern/real-table-droid
  assets:
    asset_id: droid
    assets_dir: gs://openpi-assets/checkpoints/pi05_droid/assets
model:
  pi05: true
  paligemma_variant: gemma_2b_lora
  action_expert_variant: gemma_300m_lora
  action_dim: 32
  action_horizon: 16
  max_token_len: 200
  dtype: bfloat16
lr_schedule:
  peak_lr: 2.5e-05
  decay_lr: 2.5e-06
  warmup_steps: 1000
  decay_steps: 30000
optimizer:
  clip_gradient_norm: 1
  weight_decay: 1e-10
freeze_filter: All(llm parameters except LoRA parameters)
weight_loader:
  params_path: gs://openpi-assets/checkpoints/pi05_droid/params

Provenance

The checkpoint was produced on 2026-04-30 and finalized successfully at step 19999. The training progress reached 20000/20000 steps, then waited for the Orbax checkpoint manager to finish the final asynchronous save.

Original local paths:

/home/ubuntu/sentinel-sft/openpi/checkpoints/pi05_droid_finetune_lora/real_table_droid/19999
/home/ubuntu/sentinel-sft/openpi/wandb/run-20260430_045047-so20pehc

The final checkpoint includes the DROID normalization statistics at:

checkpoint/19999/assets/droid/norm_stats.json

Notes

This repository is intended as an experiment record for the PI0.5 SFT run on the full 66-trajectory normal real table teleoperation dataset. It is not a generic pretrained PI0.5 release.