pi0_libero_regression_sa_ckpts

pi0 LIBERO regression LoRA checkpoints (appearance sensitivity study)

LoRA fine-tuned π0 checkpoints on LIBERO for a visual appearance-sensitivity / invariance study: the same LIBERO trajectories are re-rendered into three appearance domains, and separate models are trained per domain plus one on the combined mixture.

Models

• Base: pi0_base (PaliGemma 2B VLM + Gemma 300M action expert)
• Action head: regression (L1 loss, single forward pass at inference, 50-step action horizon) — this is a fork feature, not upstream openpi
• LoRA: rank 16 / alpha 16 on all attn + ffn linears of all layers, applied to the VLM, the action expert, and the SigLIP vision tower (lora_on_vision=True; vision base frozen)
• Training: batch 256, AdamW; single-domain runs 50k steps, combined run 150k steps (≈50k effective steps per domain; the three datasets are concatenated and globally shuffled, norm stats shared since states/actions are identical across re-renders)

Data domains

Same trajectories (identical states/actions), different rendering:

Contents (sa_0824/)

Each step folder contains:

• model.safetensors — trainable weights (LoRA A/B + regression head)
• optimizer.pt — optimizer state (for resuming training)
• metadata.pt — openpi training metadata
• assets/physical-intelligence/libero/norm_stats.json — normalization stats

Usage

These checkpoints expect the PyTorch openpi fork with the regression action head and the configs listed above (paths point at local re-rendered LeRobot datasets; adjust local_dataset_root(s) to your layout or rely on the cached norm stats included here).

from openpi.training import config as train_config
from openpi.policies import policy_config

cfg = train_config.get_config("pi0_libero_lora_regression_combined")
ckpt = "reg_combined_visionft/110000"  # downloaded from this repo
policy = policy_config.create_trained_policy(cfg, ckpt)
actions = policy.infer(obs)  # obs with base/wrist images, state, prompt

Download everything:

hf download lozoRUST/pi0_libero_regression_sa_ckpts --local-dir sa_ckpts