Study On Approaches Of Automated Planning Of Service Robot For Complex Environment

Service robots have become a hot topic in the world.The real human settlement environment is a dynamic and unstructured environment: pedestrian interference often occurs in the environment,and the states of the environment may change and be partially observable.The tasks that robots have to accomplish are also increasingly complex,diverse,ambiguous,variable and time-limited.Therefore,the robot task planning layer faces the problems that the prior task domain knowledge is difficult to cover all application situations,the environmental dynamic changes cause task execution failure,and the execution efficiency for complex tasks are low.Path planning and navigation decision-making layers often face challenges such as that the trajectory planned in a multi-pedestrians environment collides with pedestrians easily and the robots' inefficiency when greatly adjust the movement due to avoiding pedestrians.It is hoped that robots can learn and store human experience in dealing with these tasks,use these experiences to adapt to different environments autonomously,and handle abnormal situations,and ultimately complete tasks efficiently and safely.Although the research on robots' task experience learning and generalization,task planning and robot navigation in multi-pedestrians environment has made great progress,the common problems that restrict the adaptability,safety and efficiency of robots still exist: The experience of planning activities in different situations of similar tasks cannot be merged;The experience-based task re-planning repair and optimal planning solution have not been realized;It is difficult for robots to make decisions and find the optimal feasible trajecory in multi-pedestrians environment.In view of the above problems,we mainly study the robot self-planning method for complex environments.The specific contents and results are as follows:In order to describe the abstract activity experience of multiple situations of similar tasks with an activity schema,the abstract method and the concept of integrated activity schema with abstract method layers are proposed.At the same time,it is introduced into the experience-based task planning model to form a task planning model description that supports the integrated activity schema.According to the model requirements and task requirements,we build a robotic cognition and planning system based on exprience learning.In addition to the blackboard,ontology knowledge base,action knowledge base and other traditional modules,a new activity schema repository and an improved task planner have been included in the system,as a result,experience learning and integration,task planning and task execution are integrated in the same framework.In order to test the feasibility of the system,the task learning and planning scenarios for a robot are built.The test results show that the basic modules of the system can meet the model requirements and task requirements.In order to solve the problem of integration of activity experience from different situations of a same task,a learning algorithm for abstract methods and integrated activity schemas is proposed.Firstly,the human demonstration activity experiences for a task are collected and extracted,then the structure of the integrated activity schema for the current activity experience is obtain through the generalization of the example parameters in the experience and the hierarchical mapping by abstract operators and abstract methods.Second,the preconditions of the abstract methods are extracted through the correlation analysis and constraint detection technology.Finally,an algorithm for integration with the learned integrated activity schemas is designed.Integration of activity experience from different situations of similar tasks is achieved,so that the learning task experience can be applicable to multiple scenarios of the same task automatically.Planning problems.At the same time,the activity experience is conceptualized based on ontology knowledge.In the simulation with a PR2 robot,three scenarios of serve-a-coffee task for learning was conducted.The results show that the integrated activity schema successfully integrated the activity experience of different situations.In order to realize the experience-based task re-planning repair and sovlethe optimal planning solution,a task planning and re-planning method based on integrated activity schema is proposed.The first step of the method is to obtain a similar plan solution with the empirical activity schema through the depth-first and active feature matching-first search traversal algorithm;the second step is to analyze the relationship between the actions in the initial solution,and the maximum-weight clique problem is used for merging optimization to seek the optimal planning solution..The mapping relationship from task tree to integrated activity schema is established through the abstract method layer,and partially backtracking and replanning algorithm is introduced to recover from the experienced abnormal failure.A simulation with a PR2 robot and a physical experiment was conducted to validate the proposed method,The results show that hat the proposed task planning method has better adaptability and efficiency.To improve the safety and efficiency of robot navigation in crossing low-density and medium-density crowds,we propose a robot navigation method based on human trajectory prediction and multiple travel modes.First,an improved socially conscious model is used to learn and predict the pedestrian's future trajectory,in which the effect of other pedestrians' moving direction is considered,and the effect of a pedestrian's own history trajectory on the pedestrian's future trajectory is enhanced.Second,local trajectory for obstacle avoidance optimization based on the predicted pedestrian trajectory and chasing cost judgment is performed to generate safer and more efficient trajectories.Third,we have designed the five following travel modes for robot navigation in different traffic states to guide the robot finding the optimal navigation strategies: free mode,high-speed-follow mode,travelable mode,low-speed-follow mode,and probing mode.We have demonstrated the performance of our approach outperforms state-of-the-art approaches with public datasets,in low-density and simulated medium-density crowded scenarios.