Autonomous navigation remains one of the toughest challenges in robotics. While many robots rely on detailed maps or highly accurate localisation systems, these assumptions often fail in real-world environments—particularly indoors, where GPS is unavailable and visual conditions can change quickly.
A new AI-powered approach, developed by researchers at Cardiff University and Hohai University, introduces a breakthrough in robot navigation without maps. By integrating deep reinforcement learning (DRL) with real-time feedback from simultaneous localization and mapping (SLAM) systems, the model enables robots to move intelligently, avoid collisions, and maintain awareness of their position—even in complex, unfamiliar spaces.
This development represents a major step forward for autonomous robotics in digital construction, logistics, and service industries, where reliable navigation is essential for safe and efficient operation.
The challenge of Mapless Robot Navigation
Traditional robot navigation methods face significant limitations:
- Dependence on maps: Many systems require detailed, pre-built maps that are costly and impractical in dynamic or evolving environments.
- Over-reliance on localisation accuracy: Robots often assume continuous access to accurate position data, which is not always possible indoors.
- Weaknesses of visual SLAM: Visual SLAM systems can fail in textureless areas, under poor lighting, or when the robot moves abruptly.
- Rigid AI models: Existing AI-based navigation tends to focus only on collision avoidance, ignoring whether the robot can stay localised along the way.
For autonomous robots in construction sites, warehouses, and healthcare facilities, these challenges make current navigation approaches unreliable in real-world scenarios.
AI-powered navigation: a localisation-aware approach
The Cardiff and Hohai University research team has developed a localisation-aware navigation policy that combines:
- Deep Reinforcement Learning (DRL) – allowing the robot to learn adaptive decision-making strategies.
- RGB-D camera input – providing visual and depth information about the environment.
- Real-time ORB-SLAM2 feedback – delivering continuous updates on localisation quality.
- Compact state representation – encoding spatial distribution of visual map points into 24 angular regions around the robot for smarter route planning.
Key innovation: Relative Pose Error (RPE) reward function
Instead of static penalties, the system introduces a dynamic reward function based on Relative Pose Error (RPE). This gives the robot instant feedback on whether a chosen action improves or worsens its ability to localise. As a result, the robot prioritises safer, more reliable paths—even if they are longer—rather than simply aiming for the shortest distance.
Performance evaluation and results
The researchers tested the model using the iGibson simulation environment, comparing it against four baseline navigation strategies. Results showed clear advantages:
- 49% success rate vs. 33% for traditional SLAM-based navigation
- Lower localisation errors and improved adaptability in unfamiliar environments
- Preference for safer paths, avoiding areas with poor visual features
- Robust navigation in challenging, mapless indoor scenarios
This proves that AI-enhanced navigation with localisation awareness can outperform both conventional SLAM and existing AI-based models.
Applications in digital construction and beyond
The ability to navigate without detailed maps has significant implications across multiple sectors:
Digital Construction
- Autonomous site inspection: Robots can move through evolving construction sites without requiring constant re-mapping.
- Material transport: Automated guided vehicles (AGVs) can adapt to temporary obstacles and layout changes.
- Worker safety monitoring: Mobile robots can maintain reliable localization in cluttered, unpredictable environments.
Warehousing and logistics
- Indoor delivery robots can reliably navigate changing storage environments.
- Automated forklifts and AGVs can operate in dynamic spaces with varying visual conditions.
Expert insight
Dr. Ze Ji, senior author of the study, explained: “Navigation isn’t only about avoiding obstacles—it’s about maintaining confidence in your position every step of the way. By integrating perception and planning, our model enables smarter, safer movement in uncertain spaces. This could pave the way for autonomous robots that handle the complexities of the real world without constant human oversight.”
This breakthrough in AI-driven, mapless robot navigation marks a significant milestone in the journey toward fully autonomous systems. By teaching robots not just to move but to understand and adapt to their localisation confidence, the Cardiff and Hohai University model paves the way for reliable, safe, and flexible navigation in real-world environments.
For industries such as digital construction and warehousing, where robots must operate in dynamic and often unpredictable spaces, this innovation offers a powerful step forward. As the research progresses to real-world testing, the technology could become a cornerstone of next-generation robotics, enabling smarter navigation, improved safety, and increased autonomy.
Read next: Win-win approach to designing cities for cars and pedestrians – Build in Digital
Are you a building professional? Sign up for a FREE MEMBERSHIP to upload news stories, post job vacancies, and connect with colleagues on our secure social feed.