pi0.5 finetune — pnp_left_500 (step 60000)
Openpi pi0.5 finetune on a 500-episode left-arm pick-and-place LeRobot dataset. This is the step 60000 checkpoint of a 100000-step run.
Training
| Item | Value |
|---|---|
| Base model | pi0.5 (pi05_base) |
| Framework | openpi (PyTorch trainer) |
| Dataset | 502 episodes / 154,059 frames / 30 fps |
| Batch size (global) | 16 |
| Optimizer | AdamW, peak_lr=5e-5, cosine decay, warmup=500 |
| FSDP devices | 4 |
| action_dim | 54 |
| action_horizon | 50 |
| max_token_len | 256 |
| freeze_filter | LLM frozen except llm_*_1* and LoRA paths |
| Total steps | 100000 |
| This checkpoint | step 60000 |
Why step 60000
Loss trajectory (5k-step bucket means):
| step range | loss | grad_norm |
|---|---|---|
| 15k–20k | 0.0236 | 0.281 |
| 20k–25k | 0.0204 | 0.260 |
| 40k–45k | 0.0149 | 0.213 |
| 60k–65k | 0.0133 | 0.193 |
| 80k–85k | 0.0115 | 0.178 |
| 95k–100k | 0.0108 | 0.170 |
Step 60000 sits in the flattening region — most of the improvement over 20k is captured (~35% lower loss), and further training only shaves ~20% more.
Contents
model.safetensors— model weights only (7.0 GB)metadata.pt— training metadataassets/pnp-left-500-lerobot/norm_stats.json— dataset normalization statistics
optimizer.pt is not included (would add 13 GB and is only needed for warm resume).
Provenance
- W&B run:
better_guidance/pickplace-pi/8affq7ty(pi05_pnp_left_500_20260622_154344)