Mapless Autonomous Navigation through Real Time Perception

623 concepts520 papers3 proposals

Traditional autonomous navigation and robotics operate under a comfortable assumption: the existence of a pre-mapped world. High-Definition (HD) maps, pre-computed 3D point clouds, and semantic annotations act as a spatial "cheat code." They provide agents with perfect geometric priors, explicit lane boundaries, and static obstacle locations long before the robot even turns on its sensors.

However, this reliance introduces a severe prior map dependency. As noted by Keenan Burnett et al. in their development of the mapless autonomous vehicle Zeus (Building a Winning Self-Driving Car in Six Months), HD maps are prohibitively expensive to generate, scale poorly across large geographies, and become instantly obsolete when the physical environment changes. When an agent encounters an unmapped or rapidly evolving region, this structural dependency causes catastrophic failure.

To achieve true autonomy, modern research is shifting toward mapless navigation—systems that trade the brittle completeness of a global map for the adaptive resilience of real-time, onboard perception.


Shifting from Global Priors to Online Topology

If an autonomous agent cannot read lane lines or road structures from a pre-loaded map, it must infer them on the fly. This is particularly challenging for tasks like trajectory forecasting, where predicting how other actors will behave depends heavily on understanding the underlying road geometry.

Traditional Pipeline:  [HD Map Prior] -------> [Trajectory Prediction] -> [Plan]

Mapless Pipeline:      [Raw Sensor Stream] -> [Online Topology Infer] -> [Plan]

To break this bottleneck, Lingyun Luke Li et al. introduced the Interaction Transformer (End-to-end Contextual Perception and Prediction with Interaction Transformer), demonstrating how motion forecasting models can reason about relative spatial relationships and actor interactions without relying on dense geometric map priors.

When lane context is absolutely necessary for safe tracking, the solution is to estimate it dynamically rather than retrieving it from storage. For instance, ByeoungDo Kim et al. developed a lane-aware prediction framework (LaPred: Lane-Aware Prediction of Multi-Modal Future Trajectories of Dynamic Agents) that extracts instance-level lane candidates directly from real-time sensor data. By predicting road topology online, as explored by Hengli Wang (End-to-End Interactive Prediction and Planning With Optical Flow Distillation for Autonomous Driving), vehicles can maintain safe planning capabilities even when operating in entirely unmapped geographic regions.


Exploiting Local and Behavioral Cues

How do humans navigate an unfamiliar city without a map or GPS? We don't calculate exact global coordinates; instead, we read local signs and follow the behavioral flow of the crowds around us. Mapless robotics leverages this exact intuition.

Instead of trying to solve the incredibly complex problem of absolute localization in an unknown space, an agent can treat the surrounding environment as a dynamic guide. Wataru KUMAHARA et al. demonstrated this in Mobile Robot Navigation Based on Local Route Information and Pedestrians' Flow. By modeling crowd movements through a modified Lennard-Jones potential, their system allows a mobile robot to "join the flow" of pedestrians moving toward a shared destination, effectively using human behavior as a proxy for a global map.

Biologically inspired models reinforce why this approach works. Nature rarely constructs high-fidelity internal maps. In a study of hawkmoth flight mechanics, Henghui Zhu et al. (Learning from animals: How to Navigate Complex Terrains) revealed that prioritizing immediate optical flow over precise obstacle mapping provides superior environmental robustness. The moth does not need to know the exact 3D coordinates of a forest to fly through it; it simply balances the visual motion fields across its eyes. Tight perception-control loops eliminate the computational overhead—and the failure points—of persistent mapping.


Bypassing Explicit 3D Reconstruction

The traditional navigation paradigm follows a rigid sequence: Perceive $\rightarrow$ Reconstruct 3D Space $\rightarrow$ Plan Path. Mapless navigation often collapses this pipeline by mapping raw sensory inputs directly to actionable control policies, bypassing explicit 3D metric reconstruction entirely.

Traditional: [Sensors] -> [3D Mesh / Occupancy Grid] -> [Path Planning] -> [Control]

Mapless:     [Sensors] -> [Foundation Model / Pixels] --------------------> [Control]

This direct mapping is made possible by anchoring high-level semantic understanding to low-level pixels. In Bridging Zero-shot Object Navigation and Foundation Models through Pixel-Guided Navigation Skill, Wenzhe Cai et al. introduced PixNav, a system that uses single pixels as "goal specifications." By leveraging foundation models to identify targets within an RGB stream and translating those pixels directly into movement commands, the robot navigates complex spaces without ever building an explicit metric map or requiring expensive depth sensors.

Even when local tracking is required on hardware-constrained devices, relying on pre-captured reference maps can be avoided. Single-author work by Fan Lin (A novel Gravity-FREAK feature extraction and Gravity-KLT tracking registration algorithm based on iPhone MEMS mobile sensor in mobile environment) demonstrates that offloading visual tracking orientation to onboard MEMS sensors (like gyroscopes and accelerometers) allows for robust real-time localization in entirely novel environments.

By shifting the burden of navigation from remembering the world to reacting to it, mapless frameworks allow autonomous systems to step out of pre-mapped test tracks and operate reliably in the wild.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_prior_map_dependency_ec750b97,
  title = {prior map dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/prior-map-dependency/rdr_ec750b9708cb49f195c4c8408437c1a4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_prior_map_dependency_ec750b97_v1,
  title = {prior map dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/prior-map-dependency/rdr_ec750b9708cb49f195c4c8408437c1a4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake