arkojit1/pi05_haply_pick_block_bs64_5k_20k

π₀.₅ fine-tuned on Ameyapores/haply_pick_block_eef_delta — batch 64, step 5k of 20k

lerobot/pi05_base (revision a538eb27) fine-tuned with LeRobot 0.4.4 on Ameyapores/haply_pick_block_eef_delta (revision af57690, corrected gripper state) — 40 Franka episodes collected with a Haply teleoperation device, 1 task ("Pick up the block."), 4-D state, 4-D end-effector delta action.

Camera slots. The published dataset lists wrist before scene; this model was trained on a copy with scene first, so observation.images.scene → base_0_rgb, observation.images.wrist → left_wrist_0_rgb, third slot padded and masked (empty_cameras=1). Feed observations under those keys.

This is the best-eval (chosen) checkpoint, step 5,000 (35.1 epochs) of a 20,000-step run; the run's lowest eval was 0.1399 at step 2,500 (no checkpoint); this step was chosen as the run's representative early checkpoint, before the eval loss climbs steadily.

Trainable	action expert only (train_expert_only=true; SigLIP + Gemma-2B frozen)
Hardware / batch	8 × MI300X (8 per GPU), global batch 64
Steps	20,000 planned (142.3 steps per epoch = 140.5 epochs)
Optimiser	AdamW, LR 2.5e-5 peak, betas (0.9, 0.95), wd 0.01, grad clip 1.0
Schedule	cosine with warmup, pinned to 20,000 steps (warmup 666), floor 2.5e-6; bf16
Normalisation	quantile (q01–q99) for state and action
Augmentation	none
chunk_size / n_action_steps	50 / 50
Eval split	last 4 of 40 episodes (36–39) held out
Eval loss	0.1494 at this step (flow-matching loss on held-out episodes)
Train loss	0.105

Eval loss is a training-objective number on 4 held-out episodes, not a task success rate; differences of a few thousandths are within its noise.

Loading needs LeRobot 0.4.4 with HF's transformers fork for π₀ (branch fix/lerobot_openpi).

from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("arkojit1/pi05_haply_pick_block_bs64_5k_20k")