Reproduce “MetricNet: Recovering Metric Scale in Generative Navigation Policies” 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
Abhijeet Nayak, Débora Oliveira Makowski, Samiran Gode, Cordelia Schmid, Wolfram Burgard
MetricNet allows generative navigation policies to predict a real-world scale factor, grounding unscaled 2D waypoint trajectories into metric space for robot execution. This enables a 'position control' strategy to follow full metric-scaled trajectories and is integrated into MetricNav, a guidance framework that improved real-world goal-reaching rates by up to 20 percentage points and reduced collisions by actively guiding robots away from obstacles and towards goals.
Key insights
- Learning a context-dependent metric scale factor via MetricNet is more effective than using a fixed, handcrafted constant, leading to improved navigation performance.
- Executing full metric-grounded trajectories through 'position control' significantly enhances robot navigation safety and efficiency compared to 'velocity control' that targets only a single waypoint.
- Gradient-guided diffusion, leveraging metric-scaled trajectories and explicit cost functions for collision and goal proximity, can effectively steer generative navigation policies towards safer, more goal-directed paths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.