Fxiz0707/act-grab-battery

ACT — Grab Battery

An ACT robot policy trained on all 30 episodes (13,544 frames at 30 FPS) of heidar-an/grab-battery.

Training

• 30,000 optimizer steps; batch size 8; approximately 17.72 epochs.
• ACT with a ResNet18 ImageNet-pretrained visual backbone and action chunks of 100.
• AdamW, learning rate 1e-5; random seed 1000.
• NVIDIA GeForce RTX 4070 Ti SUPER, 16 GB VRAM.
• LeRobot source commit: 3f797de5952432284a7c77ad2760218cb74e2f26.
• Dataset revision: f00757ee521ffd01d7cb95b07df6af8a7a57c2f9.
• Final reported training loss: 0.085; L1 action loss: 0.078.

These are training metrics. No held-out evaluation or physical robot rollout has been performed.

Expected inputs and outputs

• observation.images.front: RGB tensor with shape [3, 360, 640].
• observation.state: six joint positions.
• action: six joint position targets, ordered shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper.
• Dataset robot type: so_follower.

Use the matching robot calibration, joint ordering, camera placement and image configuration.

Files

This repository contains the final model weights, policy configuration, preprocessor/postprocessor definitions and normalization statistics, plus the training configuration. Optimizer state and intermediate checkpoints remain on UCL storage.

Loading

With LeRobot installed and this private repository accessible to your Hugging Face account:

from lerobot.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained("Fxiz0707/act-grab-battery")

For deployment, use LeRobot's rollout tooling so the supplied preprocessing and postprocessing configurations are also loaded:

lerobot-rollout --policy.path=Fxiz0707/act-grab-battery <your robot, camera, task and strategy arguments>