Balance learned and classical modules in navigation systems

Establish principled frameworks to combine learned components and classical modules (e.g., mapping, localization, planning) in visual navigation systems so as to retain the generalization and explainability of classical stacks while leveraging the performance of end‑to‑end reinforcement learning.

Background

Navigation research spans end‑to‑end learning approaches and modular pipelines that integrate mapping and planning. While end‑to‑end RL excels in simulation, modular designs often offer better generalization and interpretability in the real world.

The survey emphasizes that determining the proper integration boundaries and roles of learned versus classical components is still unresolved and central to building robust, deployable navigation systems.

References

Striking the right balance between learned and classical modules remains an open challenge.

— Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes  (2408.03539 - Tang et al., 2024) in Trends and Challenges in Navigation (Subsection "Navigation")

Dynamic actors, perception uncertainty, broader object counts and topologies, and closed-loop validation of the vertices-plus-LiDAR planner remain for future work.

— NeuralParker: A Reinforcement Learning Planner for Irregular Parking Environments  (2608.24485 - Wang et al., 25 Aug 2026) in Conclusion, Section VI

It therefore cannot establish that every geometric controller requires a VLM, nor can the entire difference from the online system be attributed to the semantic policy.

— Reliability-Aware Sparse Route Memory for Round-Trip Vision-Language Navigation  (2609.34163 - Long et al., 28 Sep 2026) in Discussion and Conclusion, limitations paragraph

Future work should focus on three open problems: establishing standardized cross-benchmark evaluation protocols that fairly compare modular and end-to-end systems; developing hybrid architectures with formally verifiable safety properties; and reducing the sim-to-real gap for end-to-end learned policies through domain randomization and adaptive transfer techniques.

— End-to-End Learning vs. Modular Architectures: Comparative Insights into Autonomous Driving Systems  (2610.01746 - Kapse, 1 Oct 2026) in Section IV, Conclusion