IDEAS-Lab-Northwestern/pi05-real-cab-30-droid-lora-main_cam_view_refined

pi0.5 LoRA — cab task (real teleop, 30 demos, refined main-cam viewpoint)

Base weights: gs://openpi-assets/checkpoints/pi05_droid/params (openpi pi0.5 DROID checkpoint — Physical Intelligence's official) Training config: pi05_droid_finetune_lora (openpi, DROID-schema) Dataset: IDEAS-Lab-Northwestern/real-cab-30-droid-main_cam_view_refined (private) — real Franka teleop, 30 episodes / 18,235 frames / 15 fps, DROID schema (LeRobot v2.1) Real or sim: Real teleop on a Franka Panda Prompt (language conditioning):

"Open the cabinet, place the wooden cube inside, and close it. Keep the area around the door clear so it can swing freely on the way in and out."

What's different from the original cab run?

This is a fresh fine-tune (not a resume) using a second, separate batch of 30 real-teleop trajectories collected with a refined third-person camera viewpoint. Same task scenario, same prompt, same TrainConfig template, same warm-start as the original pi05-real-cab-60-droid-lora — only the physical camera placement and the demo set differ.

• Original: IDEAS-Lab-Northwestern/pi05-real-cab-60-droid-lora — cab task, 60 demos, original main-cam viewpoint
• This repo: refined cam viewpoint, 30 demos (smaller but more controlled visual conditions)

Training

• 20,000 LoRA steps total — final checkpoint saved as step 19999 (openpi 0-indexed naming)
• Batch size 4 on a single A100-SXM4-40GB
• Wall time: ~2h 48m
• LoRA adapters: paligemma_variant="gemma_2b_lora", action_expert_variant="gemma_300m_lora" (PaliGemma 2B + 300M action expert base frozen; only adapter params updated)
• action_dim=32, action_horizon=16, pi05=True
• Norm stats: reused from official DROID (gs://openpi-assets/checkpoints/pi05_droid/assets/droid) — appropriate since actions are Franka joint-velocity in the same regime

wandb run: https://wandb.ai/yiyanpeng2027-northwestern-university/openpi/runs/wyn1dln6 Final-tail loss: ~0.006 (Step 19000–19900, grad_norm ~0.16, param_norm ~1939.18) — clean convergence

Contents: only step 19999/ (params + assets + _CHECKPOINT_METADATA). train_state/ excluded — not needed for inference or LoRA-on-top resumption.

Use this checkpoint

# Download just this step
HF_HUB_DISABLE_XET=1 huggingface-cli download IDEAS-Lab-Northwestern/pi05-real-cab-30-droid-lora-main_cam_view_refined \
  --include "19999/**" \
  --local-dir vla_models/pi05-real-cab-30-droid-lora-main_cam_view_refined

# Serve with openpi
uv run scripts/serve_policy.py \
  --policy.config=pi05_droid_finetune_lora \
  --policy.dir=vla_models/pi05-real-cab-30-droid-lora-main_cam_view_refined/19999

Family

Real-teleop DROID-schema fine-tunes for safety-awareness evaluation:

• This repo: cab task, refined cam viewpoint, 30 demos
• Original cab: IDEAS-Lab-Northwestern/pi05-real-cab-60-droid-lora — original viewpoint, 60 demos
• Sibling: IDEAS-Lab-Northwestern/pi05-real-jar-60-droid-lora — jar task, real teleop

All three share the same TrainConfig template (DROID schema, pi05_droid warm-start, identical LoRA + optimizer hyperparams); they differ in dataset (+ camera viewpoint for this repo).

Sim-LIBERO companion family: pi05-sim-stack-flat-30-libero-lora + pi05-sim-stack-same-30-libero-lora + pi05-sim-lid-transport-food-30-libero-lora.

License

Apache-2.0