Mutual Information-based Robot Exploration In Unknown Environments

Autonomous robots are irreplaceable in complex environments such as deep-sea exploration,planetary exploration,cave,and field rescue.In the exploration of unknown and complex environments where optical/acoustic features are sparse or even missing and sensor performance is limited,robots often face the challenge of how to actively plan and select viewpoints using onboard sensors,as well as maximizing the collected environment information,and completing autonomous localization,environment mapping,and other tasks,without relying on external navigation devices and prior environmental maps while considering multiple constraints such as accessibility,safety,and resource consumption.In this context,using the range-sensing robotic platforms,this thesis conducts a series of research on two key issues: achieving accurate environmental perception modeling and efficient autonomous exploration in unknown,complex environments.The main research contributions of this thesis are as follows:(1)A method is proposed to address the problem of robot environmental map construction in the presence of sparse sensor observations and complex environmental characteristics.The confidence-rich grid mapping method for limited range-sensing robots is proposed.The sensor cause model is established for limited range-sensing sensors,modeling the occupancy of each map grid cell as a non-parametric continuous occupancy distribution(confidence).By considering the measurement dependency between map grid cells caused by the range measurement,the proposed method effectively restores environmental details and can handle sparse observation resulting from low-resolution range sensors in unstructured and cluttered environments.(2)To handle the imperfect accuracy and vulnerability to noise in traditional exploration methods based on occupancy grid mutual information,this thesis presents confidence-rich mutual information(CRMI)based on continuous belief distributions to measure the information gain of control actions accurately.The proposed CRMI framework addresses the limitations and noise susceptibility of traditional information-driven exploration methods.First,conduct raycasting on the current environment map at candidate actions to obtain the virtual environmental measurement,and then recursively update the belief distribution of virtual map occupancy using the beam-based measurements.The change of map entropy(i.e.,mutual information)given the virtual environment measurement is derived using information theory,accurately capturing the environmental mutual information changes during the robot’s exploration,which drives the robot to complete fine exploration in unknown complex environments.The mechanisms of different sensor models on environmental mutual information and their influence on robot exploration behavior are also studied in depth using case studies.(3)To address the expensive computational cost of evaluating the CRMI of candidate exploration actions when exploring unknown unstructured environments,an efficient information evaluation method based on Gaussian process-based Bayesian optimization is proposed.The proposed method constructs a continuous surrogate model of candidate actions and their information gain,optimizing the decision-making process with fewer samples.It incorporates an upper confidence bound function to ensure optimality and infer the optimal action,maximizing the expected information gain.By introducing sparse approximation techniques,a new information evaluation method based on Bayesian kernel inference and optimization is proposed to perform a high approximation and optimization of the surrogate model,significantly reducing computational complexity to approximate logarithmic levels.This approach further enhances the efficiency of information utility inference and the real-time performance of robot exploration.(4)To address the potential problem of significant biases or even failures in CRMI-based exploration paths due to ignoring pose uncertainty in unknown environments,a Rao-Blackwellized particle filter-based confidence-rich localization and mapping method using improved closedform particle weight updates is proposed.The proposed method achieves accurate and robust full-state estimation,including state uncertainty,for the robot exploring unknown complex environments.The uncertainty-aware information metric called uncertain confidence-rich mutual information(UCRMI)is defined by approximating the robot’s pose uncertainty using weighted particles.UCRMI balances the robot’s behavior of actively selecting actions to reduce its pose uncertainty and the behavior of exploring unknown areas,thereby improving safety and exploration precision in unknown environments.The methods proposed in this thesis address various challenges of range-sensing robots exploring unknown complex environments,such as insufficient richness of environmental occupancy information,inadequate localization precision and robustness,imperfect exploration performance,and undesirable exploration efficiency.This thesis provides solutions for accurate environmental perception and modeling,and efficient autonomous exploration,offering theoretical references and technical support for the development of autonomous robot exploration technology.