Reproduce “Occlusion-Aware Crowd Navigation Using People as Sensors” 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
Ye-Ji Mun, Masha Itkina, Shuijing Liu, Katherine Driggs-Campbell
People-as-Sensors VAE (PaS-VAE) enables a robot to anticipate pedestrians hidden by occlusion by interpreting visible people’s motion—such as slowing or turning—as indirect evidence of unseen agents. In simulation, the method reached 0.91 success and 0.09 collision rates, approaching the 0.97 success of an oracle with sequential ground-truth maps, and it was demonstrated on a Turtlebot 2i in a qualitative real-world test.
Key insights
- Visible pedestrians can act as indirect sensors: slowing, turning, or detouring may reveal the presence and likely location of people hidden from the robot.
- The representation is navigation-oriented rather than a uniformly accurate reconstruction of every occupied cell; it prioritizes hidden agents and regions relevant to the robot’s future path.
- Temporal history alone improves partial-observation navigation, but matching incomplete observations to ground-truth latent representations provides an additional benefit over sequential observation baselines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.