Monocular Robot Navigation Using Cross-Sensor Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current indoor autonomous navigation algorithms for mobile robots, such as unmanned aerial vehicles, face challenges in achieving stable navigation in complex and dynamic indoor environments with varying lighting and pedestrian presence, particularly with monocular cameras, due to issues like increased response delay, discrete action spaces, and unstable path generation.
Innovation Solution
An indoor monocular navigation method based on cross-sensor transfer learning, which uses simulation single-line laser radar data to train an autonomous navigation model via deterministic policy gradient deep reinforcement learning, and combines this with monocular camera data using a Resnet18 network and pre-trained YOLO v3 network to improve navigation accuracy by determining heading angles and performing navigation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If trajectory planning-based method is used for mobile robot navigation, then global optimal path can be obtained with controllable and stable results, but the algorithm is long in time and heavily relies on calculation resources
Solution Approach 1:
The patent pre-trains a deep reinforcement learning model in a simulation environment before actual navigation tasks. This preliminary training in virtual space allows the model to learn optimal navigation strategies beforehand, so that during real-world execution, the model can make rapid decisions without heavy real-time computation, thus resolving the contradiction between navigation reliability and decision-making time.
2Loss of time
If end-to-end decision algorithm is used for mobile robot navigation, then decision-making time is reduced with better generalization capability, but the decision is not necessarily a global optimal solution
Solution Approach 1:
The patent performs comprehensive training and optimization in the simulation environment before deployment. The end-to-end decision model is pre-trained to learn globally optimal navigation strategies through extensive simulation experiences, allowing it to make rapid sub-optimal or near-optimal decisions in real-time without sacrificing too much navigation quality, thus balancing decision speed and optimality.
3Productivity
If simulation environment training is used for autonomous navigation, then learning efficiency is improved, but direct application to real environment shows performance gap due to domain difference
Solution Approach 1:
The patent uses simulation environment as an intermediary training space between theoretical algorithms and real-world application. The simulation provides a safe, controllable environment for efficient learning, while domain adaptation techniques serve as mediators to bridge the gap between simulation and reality, allowing the model to transfer learned knowledge to real environments effectively.
Solution Approach 2:
The patent creates a virtual copy of the real environment in simulation for training purposes. By replicating real-world scenarios, physics, and sensor models in the simulation, the model learns from realistic data without physical risks. Domain adaptation then adjusts this simulated knowledge to account for remaining differences, enabling effective real-world deployment.
4Device complexity
If monocular camera is used for navigation, then device complexity is reduced, but navigation accuracy decreases due to lack of depth information
Solution Approach 1:
The patent replaces multi-sensor hardware systems (such as stereo cameras or laser rangefinders) with a monocular camera combined with deep learning algorithms. The neural network learns to infer depth and spatial relationships from single-image cues through training, substituting physical depth-sensing mechanisms with computational inference, thus maintaining low device complexity while improving navigation accuracy.
Data Source
AI summary
The present disclosure relates to an indoor monocular navigation method based on cross-sensor transfer learning and a system thereof. Determining an preliminary autonomous navigation model according to simulated laser radar data; acquiring actual single-line laser radar data and monocular camera data of the mobile robot simultaneously in an actual environment; determining the heading angle of the mobile robot according to the actual laser radar data; determining a laser radar monocular vision navigation model, according to the generated heading angle of the mobile robot and the monocular camera data at a the same moment and by using a Resnet18 network and a pre-trained YOLO v3 network; determining a heading angle of the mobile robot at the current moment, according to the acquired monocular camera data and by using the laser radar monocular vision navigation model; performing navigation of the mobile robot.