PI0.5 L18 LIBERO no-LoRA fullVLM baseline
30k-step LIBERO finetuned checkpoint for the PI0.5 L18 no-LoRA fullVLM baseline used for PI-Mamba comparison.
Checkpoint
- OpenPI config:
pi05_libero_full_5090 - Experiment:
icra_nolora_fullvlm_pi05_original_l18_cosine5e5_5e6_b16_acc1_seed42 - Step:
30000 - Initializer:
pi05_base_pytorch - Adaptation: full VLM / full vision / full PI0.5 action expert, no LoRA
- Training: batch 16, grad accumulation 1, seed 42
- LR schedule: cosine from
5e-5to5e-6over 30k steps - LIBERO eval: seed 7, 50 trials per task, 10 denoising steps
Files included:
model.safetensorsmetadata.ptassets/physical-intelligence/libero/norm_stats.json
LIBERO Results
| Spatial | Object | Goal | LIBERO-10 | Average |
|---|---|---|---|---|
| 97.6 | 98.8 | 96.6 | 93.2 | 96.5 |
Efficiency
| Metric | Value |
|---|---|
| Action expert core | 311.46M |
| Train time | 14.20 h |
| Best train step time | 1.70 s/step |
| GFLOPs/query | 4690.08 |
| Prefix/context | 40.76 ms |
| Denoise | 160.26 ms |
| Denoise per step | 16.03 ms |
| Median model latency | 228.59 ms |
| Query Hz | 4.37 |
| Action Hz | 21.87 |
Use
Use with the matching OpenPI code and config:
--config-name pi05_libero_full_5090
--checkpoint-dir /path/to/checkpoint
The checkpoint includes the LIBERO norm stats under assets/physical-intelligence/libero.