Reproduce “Understanding while Exploring: Semantics-driven Active Mapping” 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
Liyan Chen, Huangying Zhan, Hairong Yin, Yi Xu, Philippos Mordohai
ActiveSGM, developed by researchers from Stevens Institute of Technology and Goertek Alpha Labs, introduces the first dense active semantic mapping system built upon a 3D Gaussian Splatting backbone. The system efficiently constructs accurate geometric and semantic maps of unknown environments by integrating a sparse semantic representation with an uncertainty-guided exploration strategy, achieving strong performance even with noisy semantic inputs.
Key insights
- Combining a sparse (top-k) semantic representation with 3D Gaussian Splatting enables efficient, dense semantic mapping, overcoming memory limitations of storing full probability distributions.
- Guiding active exploration with a joint consideration of geometric completeness and semantic uncertainty (entropy) leads to more efficient and comprehensive map construction compared to passive or purely geometric strategies.
- Applying an entropy-based mask to filter high-uncertainty pseudo-ground truth labels, and restricting semantic loss backpropagation to semantic attributes, allows for robust semantic map refinement from noisy off-the-shelf segmentation models without degrading geometric accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.