qmddn/insert-mouse-battery-resnet18-reward-v1

Insert Mouse Battery ResNet18 Reward V1

This repository contains a binary ResNet18 reward model for the insert-mouse-battery task.

Input Format

• Image input: RGB three-view vertical stack in Wan world-model format.
• View order: cam_high + cam_left_wrist + cam_right_wrist.
• Image size: [3, 544, 320] as CHW tensor.
• Layout: each view is 180x320; the stacked image is 540x320; the full stack is resized to 544x320.
• Intended preprocessing: the same normalization/preprocessing path used by the RLinf ResNet reward model.

Output Format

The checkpoint is a binary reward model. For a single input image, the model outputs one scalar logit. Applying sigmoid gives the estimated full-task success probability.

Files

File	Description
full_weights.pt	RLinf ResNet18 reward checkpoint.
model_metadata.json	Input/output format and dataset construction metadata.
eval_summary.json	Train/validation/hard-validation metrics.
train_grid.jpg	Sampled training examples.
val_grid.jpg	Sampled validation examples.
hard_val_grid.jpg	Sampled hard-validation examples.

Test Results

split	samples	positives	negatives	accuracy	AUC	loss
	5582	2791	2791	0.9946	0.9999	0.0189
	1388	694	694	0.9856	0.9915	0.1085
	536	268	268	0.9832	0.9921	0.1076

Training Data

The dataset uses three-view videos from expert-data, success-and-hil-data, and failure-data.

• Positive samples: tail frames from expert and success/HIL episodes.
• Negative samples: early and middle frames from expert and success/HIL episodes, plus sampled frames from failure episodes.
• Train/validation splitting is done at the episode level.
• The train and validation sets are class-balanced.