arkojit1/pi05_pick_block_eef_abs

π₀.₅ fine-tuned on pick_block_eef_delta with absolute xyz actions

lerobot/pi05_base fine-tuned with LeRobot (--policy.type=pi05) on a locally rebuilt copy of Ameyapores/pick_block_eef_delta (35 Franka episodes, 1 task, two 224×224 cameras cam0/cam2; the third π₀.₅ image slot is padded, empty_cameras=1; 4-D state).

Action space — read this before deploying

The source dataset's action is [dx, dy, dz, gripper], where dx, dy, dz is exactly the change in observation.state[:3] (end-effector position) from frame t to t+1. This model was trained on the absolute form:

action[t] = [state_x[t] + dx[t], state_y[t] + dy[t], state_z[t] + dz[t], gripper[t]]

i.e. each predicted action is the end-effector position to reach at the next frame (same frame and units as observation.state[:3]), plus the unchanged binary gripper target (0/1). It is not interchangeable with the delta-action model arkojit1/pi05_pick_block_eef_delta. The built dataset is not published; train_config.json refers to it by its local path.

Action normalisation is quantile (q01–q99) over the absolute targets; 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

This is the step 1,100 checkpoint (~35 epochs), tied for the lowest held-out eval loss in the run (0.0633, also reached at step 600).

Trainable	action expert only (train_expert_only=true; SigLIP + Gemma-2B frozen)
Global batch	256 (32/GPU × 8 MI300X)
LR	2.5e-5 peak, cosine pinned to 4,000 steps (warmup 133), floor 2.5e-6; bf16
Normalisation	quantile (state and action)
Augmentation	LeRobot image transforms on train frames
chunk_size / n_action_steps	50 / 50
Eval split	last 4 of 35 episodes held out
Eval loss	0.0633 (flow-matching loss on held-out episodes)

Eval loss is a training-objective number on 4 held-out episodes, not a task success rate, and it is not comparable with the delta-action model's eval loss (different target and scale).

from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("arkojit1/pi05_pick_block_eef_abs")