seal-water-bottle-cap Stage1 ResNet18 Reward

Input Format

• RGB three-view vertical stack in Wan world-model format.
• View order: cam_high + cam_left_wrist + cam_right_wrist.
• Tensor shape: [3, 544, 320], CHW.
• Accepted value range: uint8 [0, 255] or float [0, 1]; the model preprocessing converts to float [0, 1] and applies ImageNet normalization.

Output Format

• full_weights.pt is a PyTorch state_dict for the RLinf-compatible ResNet18 binary reward model.
• The model returns one raw scalar logit for whether stage1 has been completed.
• sigmoid(logit) is the stage1 completion probability.
• Recommended inference rule: positive when sigmoid(logit) >= 0.499094.

Files

F1 Threshold

Validation Results