Kavin60606/piper-finetune-sim-validation-20260331

Piper Fine-Tuning + Sim Validation (Public Experiment)

Summary

This repo contains a full experiment log for fine-tuning a pi0_fast_piper_lora policy on a Piper pick-and-place dataset, then validating the policy on held-out episodes with side-by-side MuJoCo playback:

• Left video: model-predicted actions
• Right video: ground-truth actions

The core result is that retargeting works end-to-end and the policy tracks task intent on unseen episodes, with visible residual frame-level jitter in some segments.

Dataset

• Source: intuitioncore/piper_pick_and_place_corrected (LeRobot dataset format)
• Episode count: 157
• Frame count: 33874
• Train split used in this run: episodes 0-148
• Holdout split used for validation: episodes 149-156 (excluded from training)

Model + Training

• Config: pi0_fast_piper_lora
• Run: run3_b32_scratch_holdout5_1000
• Batch size: 32
• Checkpoint used for evaluation: step 1000

Embodiment Used in Simulation

• Asset repo: agilexrobotics/piper_isaac_sim
• MuJoCo model: piper_description/mujoco_model/piper_description.xml
• Joint mapping used for replay:
  • Dataset dims 0..5 -> joint1..joint6
  • Dataset dim 6 -> gripper opening (joint7), with mirrored finger (joint8 = -joint7)

This is the base piper_description embodiment (not piper_h_description, piper_l_description, or piper_x_description).

What We Proved

• End-to-end pipeline works: dataset -> fine-tune -> checkpoint -> held-out inference -> embodiment replay -> side-by-side video
• Unseen-episode demos are qualitatively strong for task-level behavior.
• Quantitative held-out metrics (149-156) are included in metrics/.

Key Metrics (held-out 149-156)

• Mean MAE (first action): 0.0411
• RMSE (first action): 0.0911
• Mean MAE (full horizon): 0.0490

Similarity vs Difference (Predicted vs Ground Truth)

• Similar:
  • Task intent and general movement direction are usually aligned.
  • Motion remains physically plausible in the matched embodiment.
• Different:
  • Predicted trajectories can be smoother and less reactive frame-to-frame.
  • Ground-truth contains sharper micro-corrections not always reproduced.
  • Some frames produce token decode warnings, which can add local jitter.

Repository Layout

• plots/
  • loss curves
  • benchmark summary figure
• videos/
  • side-by-side validation videos for episodes 149-153
• metrics/
  • held-out evaluation JSON
  • render manifest JSON
  • training log
• docs/
  • embodiment comparison table
• checkpoints/
  • step-1000 checkpoint payload (if fully uploaded)

Notes

• The held-out episodes shown in videos were not used during training for this run.
• Residual jitter appears attributable to limited data coverage + occasional decode instability + per-frame replay sensitivity.