pi05_libero_gr00t_h10_bs32_seed42

openpi pi05_libero_gr00t checkpoint(s): step(s) [29999]. Each <step>/params is the EMA parameter tree saved by openpi.training.checkpoints.save_state (load with openpi.models.model.restore_params(<step>/params) or openpi.policies.policy_config.create_trained_policy(config, <step>)); <step>/assets holds the normalization stats. Training code: https://github.com/jiminlx/openpi_atq (configs in src/openpi/training/config.py, design in docs/atq.md).

Config

{
  "train": {
    "batch_size": "32",
    "num_train_steps": "30000",
    "lr_schedule": "CosineDecaySchedule(warmup_steps=10000, peak_lr=5e-05, decay_steps=1000000, decay_lr=5e-05)",
    "optimizer": "AdamW(b1=0.9, b2=0.95, eps=1e-08, weight_decay=1e-10, clip_gradient_norm=1.0)",
    "ema_decay": "0.999"
  },
  "model": {
    "action_dim": 32,
    "action_horizon": 10,
    "max_token_len": 200,
    "dtype": "bfloat16",
    "paligemma_variant": "gemma_2b",
    "action_expert_variant": "gemma_300m",
    "pi05": true,
    "discrete_state_input": false,
    "pytorch_compile_mode": "max-autotune"
  },
  "data": "LeRobotLiberoGr00tDataConfig(repo_id='libero_gr00t_delta', repo_root='/s3data/libero_gr00t_delta/v1', assets=AssetsConfig(assets_dir=None, asset_id=None), base_config=DataConfig(repo_id=None, repo_root=None, video_backend=None, loader_in_order=True, asset_id=None, norm_stats=None, repack_transforms=Group(inputs=(), outputs=()), data_transforms=Group(inputs=(), outputs=()), model_transforms=Group(inputs=(), outputs=()), use_quantile_norm=False, action_sequence_keys=('actions',), prompt_from_task=True, rlds_data_dir=None, action_space=None, datasets=()), conf_sidecar=None)"
}