zimplex/pi05-base-ee-genesis-hr-bench-step100000

pi05 fine-tuned on genesis-hr-bench (ee, step 100000)

Fine-tune of lerobot/pi05_base on the genesis-hr-bench/genesis_hr_bench_lerobot_v30 genesis-hr-bench dataset, action-space variant ee.

End-effector delta actions (xyz + axis-angle rot + gripper, action_dim=7).

Training

Base model	lerobot/pi05_base
Dataset	genesis-hr-bench/genesis_hr_bench_lerobot_v30
Action variant	ee (action_dim = 7, chunk_size = 50)
Steps	100,000 (final)
Per-GPU batch size	8
GPUs	8 × H200
Effective batch	64
Optimizer	AdamW (lr 2.5e-5, weight_decay 0.01)
Scheduler	cosine_decay_with_warmup (warmup 1k, decay 30k, peak 2.5e-5, floor 2.5e-6)
Precision	bfloat16
Image resolution	224 × 224
Image augmentation	disabled
Seed	1000
Hardware	1 node, FAIR Cloud (h200 partition), slurm job 1371539
Wandb run	0abw7au2

The full training config (lerobot v2 train_config dump) is included in train_config.json.

Files

File	Purpose
model.safetensors	Policy weights (bf16).
config.json	Policy config (architecture + input/output features).
policy_preprocessor.json + policy_preprocessor_step_*_normalizer_processor.safetensors	Input normalizer (image / state).
policy_postprocessor.json + policy_postprocessor_step_*_unnormalizer_processor.safetensors	Action un-normalizer.
train_config.json	Full lerobot training config used to produce this checkpoint.

The optimizer / RNG state was not uploaded — this repo is inference-only.

Usage

from lerobot.policies.factory import make_policy_from_pretrained

policy = make_policy_from_pretrained(
    "zimplex/pi05-base-ee-genesis-hr-bench-step100000",
    device="cuda",
)
policy.eval()

Or directly via lerobot.scripts.lerobot_eval:

python -m lerobot.scripts.lerobot_eval \
    --policy.path=zimplex/pi05-base-ee-genesis-hr-bench-step100000 \
    --dataset.repo_id=genesis-hr-bench/genesis_hr_bench_lerobot_v30

Provenance

• Wandb run: https://wandb.ai/multi-agent-world-model/pi05_genesis_hr_bench/runs/0abw7au2
• Slurm job: 1371539 (h200_mrs_2)
• Local exp: runs/pi05/finetune/pi05_base_ee_20260524
• Training code: scripts/finetune_lerobot_v2.sh

Part of the {pi0, pi0.5} × {ee, qpos, qpos_abs} sweep — see CHECKPOINT_SUMMARY.md for the full 6-checkpoint matrix.