arkojit1/pi05_pick_block_eef_abs_bs64_lr2.5e-5_20k

π₀.₅ fine-tuned on pick_block_eef_position_abs (batch 64, step 20,000, final)

lerobot/pi05_base fine-tuned with LeRobot 0.6.1 (--policy.type=pi05) on Ameyapores/pick_block_eef_position_abs (35 Franka episodes / 9,181 frames @ 25 fps, 1 task, two 224×224 cameras cam0/cam2; the third π₀.₅ image slot is padded, empty_cameras=1; 4-D state).

This is a full fine-tune of the action expert (SigLIP and Gemma-2B frozen, --train_expert_only, ~0.69B trainable parameters), not a LoRA adapter. The repo holds the complete policy weights. It is the final checkpoint of a 20,000-step run.

Action space — read this before deploying

Each action is [x, y, z, gripper]: the absolute end-effector position to reach at the next frame (same frame and units as observation.state[:3]) plus a binary gripper target (0/1). It is not interchangeable with delta-action models such as arkojit1/pi05_pick_block_eef_delta.

State and action use quantile (q01–q99) normalisation. The training region is narrow: x 0.528–0.559, y 0.056–0.068, z 0.145–0.311, so targets outside it are extrapolation.

Training

setting	value
trainable	action expert + projections (~0.69B); SigLIP and Gemma-2B frozen
batch	64 global (8 GPUs × 8)
optimizer	AdamW, peak LR 2.5e-5, cosine to 2.5e-6 over 20,000 steps, 666 warmup
this checkpoint	step 20,000 (~157 epochs), end of the schedule
other	bf16, gradient checkpointing, torch.compile (default mode), image augmentation on, chunk_size 50, n_action_steps 50

Evaluation

Flow-matching loss on 4 held-out episodes (31–34), not a task success rate:

• this checkpoint: eval 0.0991, train 0.027.
• the run's best eval was 0.0587 at step 1,600 (~13 epochs, train 0.070); after it, train loss kept falling while eval rose, i.e. this checkpoint is overfit relative to step 1,600 by held-out loss. Whether that matters for task success has not been measured.
• point-to-point eval scatter on this 4-episode holdout is about ±0.005–0.01.

Loading

lerobot-eval --policy.path=arkojit1/pi05_pick_block_eef_abs_bs64_lr2.5e-5_20k ...