DylanSh/so101_orange_rectangular_block_blue_cup_act_40k

SO-101 ACT — Orange Rectangular Block to Blue Cup

ACT policy trained for the task:

Pick up the orange rectangular block and place it inside the blue cup.

Training

• Robot: SO-101 follower arm
• Policy: ACT with a ResNet-18 visual backbone
• Demonstrations: 50 successful real-robot episodes, 32,481 frames
• Recording rate: 30 FPS
• Cameras: two RGB streams at 640×480
• Training steps: 40,000
• Action chunk size: 50
• Executed action steps per chunk: 25
• Inputs: two RGB camera streams plus six-dimensional joint state
• Output: six-dimensional joint action

The final 1,000-step window had a mean training loss of approximately 0.0729 ± 0.0015. Training loss is an optimization metric and is not a physical success-rate measurement.

Load with LeRobot

from lerobot.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained(
    "DylanSh/so101_orange_rectangular_block_blue_cup_act_40k"
)

Camera-key warning

The recording setup used these historical feature names:

• observation.images.front: close arm-mounted view
• observation.images.wrist: fixed workspace overview

The names are semantically reversed but must remain unchanged at deployment because they are part of the learned feature mapping.

Intended use and limitations

This checkpoint is intended for the same SO-101 calibration, camera geometry, task wording, table layout, and object appearance used during recording. Run a no-motion calibration/port check before physical evaluation and keep an emergency stop available. Performance may degrade after camera, lighting, calibration, gripper, object, or workspace changes.

The training dataset is currently stored locally and is not included in this model repository.