Insert Mouse Battery ResNet18 Reward V2
This repository contains a binary ResNet18 reward model for the insert-mouse-battery task, retrained from scratch on the local processed LeRobot data.
Input Format
- Image input: RGB three-view vertical stack from the processed LeRobot videos.
- View order:
observation.images.top_head + observation.images.hand_left + observation.images.hand_right. - Image size:
[3, 1440, 640]as CHW tensor. - Layout: each view is
480x640; the stacked image is1440x640. - Preprocessing: RGB conversion, CHW float tensor in
[0, 1], ImageNet mean/std normalization.
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 | ResNet18 reward checkpoint trained from scratch. |
model_metadata.json | Input/output format and dataset construction metadata. |
eval_summary.json | Train/validation metrics from the training run. |
Test Results
| split | samples | positives | negatives | accuracy | precision | recall | F1 | loss |
|---|---|---|---|---|---|---|---|---|
| validation | 192 | 96 | 96 | 0.9948 | 0.9897 | 1.0000 | 0.9948 | 0.0063 |
| best validation | 192 | 96 | 96 | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 0.0174 |
Training Data
The dataset uses three-view videos from the existing processed LeRobot splits.
- Positive samples: tail frames from successful episodes.
- Negative samples: head, early, and middle frames from the same successful episodes.
- Train/validation splitting follows the existing processed split directories and stays episode-level.
- The train and validation sets are class-balanced.
- No pretrained weights from
qmddn/insert-mouse-battery-resnet18-reward-v1were used.