Vision-Based Navigation in Humanoid and Mobile Robotics

Vision-based navigation integrates visual sensing with autonomous control to enable humanoid and wheeled or legged robots to perceive, localise and plan trajectories in unstructured settings. Systems typically combine simultaneous localisation and mapping (SLAM), visual odometry and obstacle detection to construct and update environment representations in real time. Geometry-based approaches extract keypoints and estimate camera motion via epipolar geometry, whereas appearance-based and deep-learning frameworks employ convolutional neural networks to learn features directly from imagery. Humanoid platforms introduce dynamic balance and gait adaptation constraints that demand precise egomotion estimation, while mobile robots prioritise energy efficiency and path smoothness. Recent advances in sensor fusion, robust estimation and end-to-end learning have expanded capabilities in indoor service, search and rescue, industrial inspection and planetary exploration, fostering more resilient, adaptable and scalable navigation solutions.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

One study introduced a deep learning framework for visual map generation in both indoor and outdoor arenas, using a pretrained local feature transformer constrained by a fundamental matrix to link consecutive key images. Robust outlier rejection via a marginal sample consensus algorithm enhanced map consistency across six benchmark datasets, outperforming classic handcrafted descriptors and streamlining subsequent localisation and planning stages. A second contribution proposed an appearance-based localisation pipeline for a humanoid robot that leveraged transfer learning from large-scale image datasets. By integrating a global average pooling layer with an L₂-norm constraint, the approach boosted feature discriminability and achieved top-three recognition accuracy above 90 per cent across bespoke visual maps. Finally, a vision navigation control system for mobile robots combined a novel marking-line detection method with fuzzy filtering to improve path-tracking stability. After image preprocessing and target-line recognition, the system computed real-time offsets in angle and distance, attaining angular deviations within ±1.5° and demonstrating robust operation in varied lighting and surface conditions.

Vision-Based Navigation in Humanoid and Mobile Robotics publication trend

The graph below shows the total number of articles in vision-based navigation in humanoid and mobile robotics across all publications each year (not limited to Nature Index journals).

Technical terms

Simultaneous Localisation and Mapping (SLAM): A process by which a robot incrementally builds a map of an unknown environment while estimating its own position within it.

Appearance-based Localisation: A method of determining robot pose by matching current camera views against a database of previously captured images.

Convolutional Neural Network (CNN): A deep learning architecture that applies convolutional filters to extract hierarchical features from images.

Transfer Learning: The reuse of a neural network pretrained on a large dataset for a new, related task, often with fewer labelled examples.

Fundamental Matrix: A mathematical matrix that encapsulates the epipolar geometry between two camera views and enables correspondence estimation.

Marginal Sample Consensus (MAGSAC): A robust estimator that identifies and rejects outliers when fitting geometric models to data.

References

1. Transfer Learning for Humanoid Robot Appearance-Based Localization in a Visual Map. IEEE Access (2021).
2. A Deep Learning-Based Visual Map Generation for Mobile Robot Navigation. Eng (2023).
3. Recognition and Localization of Target Images for Robot Vision Navigation Control. Journal of Robotics (2022).