Reproduce “RoboLight: A Dataset with Linearly Composable Illumination for Robotic Manipulation” 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
Shutong Jin, Jin Yang, Muhammad Zahid, Florian T. Pokorny
RoboLight introduces the first real-world dataset for robotic manipulation specifically designed to capture systematic and reproducible illumination variations. It offers both real-world collected data and an extensive synthesized component, leveraging linear light transport principles to enable rigorous analysis of how lighting affects robotic policy performance. The dataset includes 2,800 real-world episodes and 196,000 synthesized episodes, demonstrating that synthetic data effectively trains policies with comparable performance to real data.
Key insights
- The linearity of light transport in HDR image space allows for robust and visually accurate synthesis of diverse lighting conditions, significantly expanding dataset size without additional real-world collection.
- Robotic policies exhibit varying degrees of sensitivity to lighting changes across tasks, with color shifts often leading to consistent performance degradation.
- Failures under shifted lighting conditions primarily manifest as object mislocalization during grasping, indicating the importance of lighting-aware perception for robust robotic systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.