pi0-fast LoRA Baseline — LIBERO (1000 steps)
Experimental LoRA fine-tune of pi0-fast on the full LIBERO dataset.
Purpose
This is the baseline run (Run A) for a sample efficiency experiment comparing:
- Run A: Raw 2h teleop data → fine-tune pi0-fast (this checkpoint)
- Run B: Raw 30min teleop data → fine-tune pi0-fast
- Run C: Enriched 30min teleop data (via Fidelity-Teleop-Workbench enrichment pipeline) → fine-tune pi0-fast
Claim: Enriched 30min data can match raw 2h data performance → 4x sample efficiency.
Training Details
| Parameter | Value |
|---|---|
| Base model | pi0-fast (PaLI-Gemma 2B + FAST tokenizer) |
| Method | LoRA (rank 16) |
| Dataset | physical-intelligence/libero (1693 episodes, 7.6h, 40 tasks) |
| Steps | 1,000 |
| Batch size | 32 |
| Step time | ~3.5s |
| Total time | ~58 min |
| GPU | NVIDIA A100-SXM4-40GB |
| VRAM used | 36.8 GB |
| Framework | JAX (OpenPI) |
Loss Curve
| Step | Loss | Grad Norm |
|---|---|---|
| 0 | 12.54 | 39.40 |
| 100 | 9.69 | 30.13 |
| 200 | 4.68 | 5.80 |
| 300 | 3.25 | 3.04 |
| 400 | 2.66 | 3.37 |
| 500 | 2.37 | 3.89 |
| 600 | 2.21 | 4.25 |
| 700 | 2.09 | 4.19 |
| 800 | 1.99 | 3.91 |
| 900 | 1.92 | 4.02 |
| 1000 | 1.89 | 3.77 |
Checkpoint Structure
Usage with OpenPI
WandB
Training tracked at: wandb run
Enrichment Pipeline
The enrichment pipeline produces per-episode signals:
- Quality scores (stability + contact consistency)
- Intent classification (approach/engage/manipulate/disengage)
- Failure detection (slip, failed grasp, retry loops)
- Contact estimation (kinematic-based)
- Temporal segmentation (change-point detection)
These signals enable quality filtering and sample weighting for Run C.