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

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

Spatial	Object	Goal	LIBERO-10	Average
98.0	99.8	98.0	94.0	97.5

Efficiency

Metric	Value
Action expert core	29.74M
Train time	10.81 h
Best train step time	1.04 s/step
GFLOPs/query	3199.13
Prefix + ctxD	41.15 ms
Denoise	56.06 ms
Denoise per step	5.61 ms
Median model latency	107.54 ms
Query Hz	9.30
Action Hz	46.50

Use

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

--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.