arkojit1/act_pick_block_eef_delta

ACT on Ameyapores/pick_block_eef_delta

LeRobot ACT policy (policy.type=act, 52M parameters) trained on the Franka pick-block dataset (35 episodes @ 25 fps, two 224×224 cameras, 4-dim state, 4-dim end-effector-delta action).

This is the step 12,000 checkpoint (~12 epochs) of a 100,000-step run, chosen as the lowest held-out loss: eval L1 = 0.2311 on normalised actions, over the last 4 episodes held out. Later checkpoints overfit (0.284 at 100k).

Recipe. LeRobot's reference ACT settings: batch 8, AdamW at a constant 1e-5 (backbone included), weight decay 1e-4, grad clip 10, fp32. ResNet18 backbone from ImageNet, fine-tuned (FrozenBatchNorm). chunk_size 100 / n_action_steps 100 (4 s at 25 fps), KL weight 10, image augmentation on. 1× AMD MI300X.

from lerobot.policies.act.modeling_act import ACTPolicy
policy = ACTPolicy.from_pretrained("arkojit1/act_pick_block_eef_delta")

Pre- and post-processor configs (with the dataset's normalisation stats) are included.