Reproduce “Learning-Based Navigation for Indoor Mobile Robots” 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
Tri-Tin Nguyen, Tien-Dat Nguyen, Gia-Uy Le, Vinh Nguyen, Vinh-Hao Nguyen
A hybrid navigation stack for indoor differential-drive robots combines a cost-aware neural global planner with a learning-augmented Dynamic Window Approach. Trained from A* demonstrations and refined with PPO, the system improves path length, tracking, and motion smoothness in several tests, while accepting longer traversal times and relying on GPU acceleration for practical neural global planning.
Key insights
- Autoregressive rollout quality does not follow one-step prediction accuracy directly: the best global-planner checkpoint reached 91.94% rollout success despite later epochs achieving similar but lower classification performance.
- Learning within a constrained candidate space improves smoothness and, in several scenarios, path length and tracking, but the resulting controller is more conservative and takes longer to reach the goal.
- The hybrid architecture offers a practical compromise between model-based safety and learning-based adaptation, though its benefits depend on GPU inference and the diversity of training environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.