Reproduce “Self-supervised Learning Of Visual Pose Estimation Without Pose Labels By Classifying LED States” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nicholas Carlotti, Mirko Nava, Alessandro Giusti
A deep learning model is presented that performs monocular RGB relative pose estimation for ground robots by classifying the on/off states of onboard LEDs. This self-supervised approach eliminates the need for explicit pose labels or detailed CAD models, achieving 70% pose accuracy (Γ_45°1m) and 17 pixels localization error while enabling autonomous data collection.
Key insights
- Observable and controllable LED states can serve as the sole source of supervision for learning complex spatial understanding, such as full 2D robot pose.
- A carefully designed loss function, incorporating spatial weighting, orientation-based occlusion reasoning, and multi-scale processing, allows an FCN to infer pose details from basic LED state classification.
- Autonomous data collection by pairs of robots toggling LEDs provides a scalable and infrastructure-free method for generating training data in diverse environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.