arkojit1/pi05_pick_block_eef_abs_lora_bs128_lr5e-5_20k

π₀.₅ LoRA adapter — pick block, absolute end-effector position (batch 128, 20k steps)

LoRA adapter for lerobot/pi05_base, fine-tuned with LeRobot 0.6.1 on Ameyapores/pick_block_eef_position_abs (Franka, 35 episodes / 9,181 frames @ 25 fps, one task, two 224×224 cameras). Actions are 4-D: absolute end-effector x, y, z (the next frame's position) plus a binary gripper target. State and action use quantile (q01/q99) normalisation.

This repo holds only the adapter (1,287,168 trainable parameters, 5 MB) plus the policy config and pre/post-processors. The base weights are loaded from lerobot/pi05_base.

Training

setting	value
adapter	LoRA, r = 16, lora_alpha = 8, no fully-trained modules
targets	LeRobot's π₀.₅ default: action-expert attention q_proj/v_proj, action_in_proj, action_out_proj (the default regex does not match π₀.₅'s time_mlp_in/out, so they are not adapted)
base weights	all frozen
batch	128 (8 GPUs × 16)
optimizer	AdamW, peak LR 5e-5, cosine to 2.5e-6 over 20,000 steps, 666 warmup
steps	20,000 (~315 epochs) — this is the final checkpoint
other	bf16, gradient checkpointing, image augmentation on, chunk_size 50, n_action_steps 50, 1 empty camera slot

Evaluation

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

• step 20,000 (this checkpoint): eval 0.0675, train 0.060
• best over the run: 0.0616 at step 18,400
• eval was flat from step 5,000 to the end (quarter means 0.0668 / 0.0667 / 0.0666); train loss fell below eval from step 7,200 without eval rising

Loading

lerobot-eval --policy.path=arkojit1/pi05_pick_block_eef_abs_lora_bs128_lr5e-5_20k ...

LeRobot reads use_peft: true from config.json, loads lerobot/pi05_base, then applies this adapter. Requires peft>=0.18. The Hub may show a warning that task_type in adapter_config.json must be a string: LeRobot leaves it null because a robot policy is not one of PEFT's task types. It does not affect loading.