fbsh96/so101-sock-ball-act-dualcam-letterbox-100eps-mi300x-b16-20000steps

SO101 Sock/Ball Pick-Place ACT Dual-Camera Letterbox Checkpoint

This is a LeRobot ACT checkpoint trained on fbsh96/so101_sock_ball_pick_place_formal_100ep.

Training Summary

• Policy: ACT
• Dataset: fbsh96/so101_sock_ball_pick_place_formal_100ep
• Episodes: 100
• Frames: 91,600
• Device: AMD Instinct MI300X via ROCm/PyTorch
• Steps: 20,000
• Batch size: 16
• Input features: observation.state, observation.images.hand_cam, observation.images.front_cam
• Action dimension: 12
• Final logged loss: ~0.071

Camera Preprocessing

The original dataset contains:

• hand_cam: 640x480
• front_cam: 1280x720

For ACT dual-camera training, both camera tensors must share the same shape. The front_cam videos were converted with a non-cropping letterbox transform:

1280x720 -> scale to 640x360 -> pad to 640x480

This preserves the full original front-camera field of view and avoids losing edge objects. The trained checkpoint expects both camera inputs as [3, 480, 640].

Related Baseline

A hand-camera-only baseline is available separately as fbsh96/so101-sock-ball-act-handcam-100eps-mi300x-b16-20000steps.