CoolFace
Modelpublic

giakhuyendihoc/pimamba-l12-depth4-libero-nolora-fullvlm

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes
Model Card

PI-Mamba L12 Depth4 LIBERO no-LoRA fullVLM

30k-step LIBERO finetuned checkpoint for PI-Mamba L12 depth4 no-LoRA fullVLM.

Checkpoint

  • —OpenPI config: pi05_mamba_tiny_early_ctxd12_gated_fullvlm_libero_5090
  • —Experiment: icra_nolora_fullvlm_pimamba_l12_depth4_cosine5e5_5e6_b16_acc1_seed42
  • —Step: 30000
  • —Initializer: pi05_base_pytorch
  • —Adaptation: full executed VLM / full vision / full PI-Mamba action path, no LoRA
  • —Context: early VLM exit at layer 12, ctxD K=64, pooled gated conditioning
  • —Mamba action mixer: depth 4, width 512
  • —Training: batch 16, grad accumulation 1, seed 42
  • —LR schedule: cosine from 5e-5 to 5e-6 over 30k steps
  • —LIBERO eval: seed 7, 50 trials per task, 10 denoising steps

Files included:

  • —model.safetensors
  • —metadata.pt
  • —assets/physical-intelligence/libero/norm_stats.json

LIBERO Results

SpatialObjectGoalLIBERO-10Average
98.099.898.094.097.5

Efficiency

MetricValue
Action expert core29.74M
Train time10.81 h
Best train step time1.04 s/step
GFLOPs/query3199.13
Prefix + ctxD41.15 ms
Denoise56.06 ms
Denoise per step5.61 ms
Median model latency107.54 ms
Query Hz9.30
Action Hz46.50

Use

Use with the matching PI-Mamba/OpenPI code and config:

bash
--config-name pi05_mamba_tiny_early_ctxd12_gated_fullvlm_libero_5090
--checkpoint-dir /path/to/checkpoint

The checkpoint includes the LIBERO norm stats under assets/physical-intelligence/libero.