Adam-YAN/pi05-liberox-base

pi0.5 fine-tuned on LIBERO-X (full-parameter)

Reproduction of the pi0.5 baseline from LIBERO-X: Robustness Litmus for Vision-Language-Action Models (arXiv:2602.06556), trained with openpi (commit 215abfb).

setting	value
init weights	gs://openpi-assets/checkpoints/pi05_base/params
data	meituan/LIBERO-X (all 2,520 demos)
fine-tuning	full-parameter, pi05=True, action_horizon=10
global batch / steps	256 / 30,000 (~8.6 epochs)
lr	cosine, peak 5e-5, warmup 10k (openpi pi05_libero defaults; paper does not state)
optimizer	AdamW, clip 1.0, EMA 0.999
normalization	quantile (openpi default for pi0.5)
hardware	4x H200, FSDP, 22.6 h

Files

• params/ - orbax checkpoint of the EMA weights (load with openpi policy_config.create_trained_policy)
• assets/meituan/LIBERO-X/norm_stats.json - dataset normalization stats

openpi TrainConfig (add to src/openpi/training/config.py)

TrainConfig(
    name="pi05_liberox",
    model=pi0_config.Pi0Config(pi05=True, action_horizon=10, discrete_state_input=False),
    data=LeRobotLiberoDataConfig(repo_id="meituan/LIBERO-X", base_config=DataConfig(prompt_from_task=True), extra_delta_transform=False),
    batch_size=256,
    lr_schedule=_optimizer.CosineDecaySchedule(warmup_steps=10_000, peak_lr=5e-5, decay_steps=1_000_000, decay_lr=5e-5),
    optimizer=_optimizer.AdamW(clip_gradient_norm=1.0),
    ema_decay=0.999,
    weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
    num_train_steps=30_000,
)

LIBERO-X LEVEL1 evaluation (600 tasks x 10 trials, official eval_template.py, 20 Hz, 1200-step cap)

time limit	this model	paper pi0.5
1.1x human time	62.0	65.2 (Table III)
1200-step cap	77.2	-

The 1.1x limit is reconstructed per task as 1.1 x mean demo length in the training set (the released code has no time-limit data).