Show HN: Automate Robot Data Quality Improvement
6 months ago
- #data-filtering
- #robot-learning
- #quality-assessment
- A toolkit for quantitatively scoring LeRobot episodes based on multiple quality dimensions.
- Combines Computer Vision heuristics and optional Gemini-powered vision-language checks for scoring.
- Features include automatic scoring on visual clarity, motion smoothness, collision detection, and more.
- Allows filtering low-quality episodes to improve training performance.
- Supports training and comparing baseline vs. filtered dataset models.
- Includes visualization of score distributions and identification of problematic episodes.
- Requires Python 3.8 or higher and pip for installation.
- Supports optional Gemini API for VLM-based scoring.
- Provides detailed output format including per-attribute scores and aggregate scores.
- Includes training and evaluation integration with LeRobot's pipeline.
- Offers troubleshooting tips for common issues.
- Open for contributions under Apache 2.0 License.