xuewei-huang/BehaPi-ckpts

BehaPi · π0.5 LIBERO Ablation Checkpoints

5 × 30k-step LoRA-free fine-tunes of pi05_base on LIBERO (4 task suite, 10 tasks each), trained on 8×A800 with FSDP. Companion checkpoints for the ablation study in github.com/Xuewei-Huang/BehaPi.

Results (each config: 4 suite × 50 trial × 10 task = 2000 evaluation episodes)

Config	Path	Spatial	Object	Goal	LIB10	Mean	Δ vs BL
Baseline (vanilla pi0.5)	libero_baseline_a800_train/30000/	91.8	94.6	90.8	86.2	90.85	—
T1: Per-timestamp Normalization	t1_per_ts_norm_a800_train/30000/	94.0	96.0	92.4	88.2	92.65	+1.80
T2: Correlated FM + K=8 Multi-sample	t2_corr_multi_a800_train/30000/	91.4	97.0	92.0	87.8	92.05	+1.20
T3: KV Transform	t3_kv_transform_a800_train/30000/	90.4	97.8	92.6	85.0	91.45	+0.60
Combined (T1+T2+T3)	combined_all_a800_train/30000/	93.0	96.0	92.6	87.8	92.35	+1.50

Full per-suite analysis + debug stories in the code repo's RESULTS.md.

Key findings

1. T1 (per-timestamp normalization) is the single biggest contributor (+1.80pp). Smallest code change, biggest impact.
2. Combined is sub-additive (+1.50pp) — less than T1 alone.
3. T3's KV mixing matrix is still ≈ identity at 30k steps (mean ≈ 1/18, Frobenius-from-identity ≈ 0). The trick is under-trained, not ineffective.

Download a single ckpt

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="Xuewei-Huang/BehaPi-ckpts",
    allow_patterns="t1_per_ts_norm_a800_train/*",
    local_dir="./behapi_ckpts",
)

Load with openpi

from openpi.policies import policy_config as pc
from openpi.training import config as cfg
train_cfg = cfg.get_config("t1_per_ts_norm_a800_train")
policy = pc.create_trained_policy(train_cfg, "./behapi_ckpts/t1_per_ts_norm_a800_train/30000")
# policy.infer({"observation/image": ..., "observation/wrist_image": ..., "observation/state": ..., "prompt": ...})

You'll need the matching code from the corresponding branch:

• t1_per_ts_norm_a800_train → branch dev/trick/per-ts-norm
• t2_corr_multi_a800_train → branch dev/trick/correlated-and-multi
• t3_kv_transform_a800_train → branch dev/trick/kv-transform
• combined_all_a800_train → branch dev/trick/combined-all
• libero_baseline_a800_train → branch dev/port-b1k-tricks

Provenance

• Base: gs://openpi-assets/checkpoints/pi05_base/params (Physical Intelligence)
• Fine-tuned on: LIBERO 4 suites (xuewei-huang/libero HF dataset)
• Training: 8×A800 (HKUST-GZ HPC), FSDP, 30k step, batch 256 / 128 (×K=8 for T2)
• Code: github.com/Xuewei-Huang/BehaPi

Citations

• Larchenko et al. 2025, BEHAVIOR-1K Challenge 2025 1st place solution (arXiv:2512.06951) — source of the 3 tricks
• Physical Intelligence, π0.5 — base model