5hadytru/so101_bench_pi05

SO-101 Bench: π0.5 full fine-tune

Full fine-tune of lerobot/pi05_base (LeRobot eb7a54c) on the SO-101 Bench simulated teleop dataset 5hadytru/so101_bench_sim_WM (3,629 episodes, 1,996,120 frames) with four cameras: overhead, wrist, and the settled frames at robot start, overhead_init and wrist_init.

• Official checkpoint: step 18000 (LeRobot micro-steps; 9,000 optimizer updates at global batch 256, 2.30M samples, ~1.15 epochs), the final checkpoint of a 36 h budget on 4x A100-80GB. EMA 0.999 weights, bf16.
• Recipe: OpenPI pi05_libero scaled to the budget (warmup to 5e-5 then constant, AdamW 0.9/0.95, quantile normalization, 30-step chunks; eval executes 15).
• Validation (39-episode SO-101 Bench val set): 16/39 (41%); best validated checkpoint was step 12000 at 18/39.
• Load with LeRobot's PI05Policy.from_pretrained; actions and state are the dataset's native LeRobot joint units.