Research On Autonomous Behavior Planning Of Robotics With Behavior Trees

With the development of software and hardware,robots are being integrated into all aspects of everyday life.Intelligent robots have become the summit of the global scientific and technological revolution,of which behavior generation is one key technology.In recent years,the rapid development of artificial intelligence has brought not only broad prospects,but also new challenges for intelligent robots.The behavior generation is no longer combining simple actions,but constructing complex behaviors by coordinating a wide variety of intelligent skills.The behaviors are no longer pure hard-coded pre-designs,but require robots to autonomously plan their behaviors to be adaptive to a wider range of unmanned scenarios.Therefore,it is of great significance to study the autonomous behavior planning of robotics.However,current behavior models mainly rely on manual design,and have limited capability for autonomous generation due to lack of representation and computation mechanisms.This thesis studies computational behavior models and their automated planning algorithms.The major contributions of this paper are listed as follows:(1)The computational model and theory for autonomous behaviors of robotics.The computational behavior model and theory are the basis for autonomous behavior planning of intelligent robots.This thesis proposes computational behavior model of intelligent robot based on behavior trees(BTs).With the classical representation and the state-space formulation,the characteristics and advantages of BTs are systematically analyzed,and the computational theory of model state and model capability for behaviors is proposed.For the first time,this thesis studies the group behavior model of distributed multi-robot systems based on BT sets,and the computational theory of group behavior state and capability is proposed.These research achievements can provide not only theoretical and computational basis for robot autonomous behavior planning,but also a theoretical analysis framework for the application of BTs in related fields.(2)A complete algorithm for autonomous behavior planning in single-robot tasks.Besides the computational model,the autonomous behavior generation of intelligent robots also requires task-oriented planning algorithms.For the first time,this paper establishes a theoretical framework for autonomous behavior planning problem based on classical planning and the computational BT models.A complete algorithm for autonomous behavior planning is proposed with intensive theoretical analysis.With the help of the computational theory of behavior capability,the advantages of the proposed method over classical planning are analyzed.It is proved that robots autonomously planning BTs are theoretically robust to any resolvable disturbances,and the performance of the algorithm is verified by experiments.Experimental results show that the proposed algorithm surpasses the state-of-the-art baseline by virtue of its complete theoretical basis,and achieves significantly improved performance.(3)Multi-robot autonomous behavior planning algorithms and improved polynomial time solution.Multi-robot autonomous behavior planning is important for complex cooperative tasks.Based on multi-agent planning and the multi-robot computational BT model,this paper extends the theoretical framework from single-robot to multi-robot problems focusing on the correctness and efficiency.To reduce the computational complexity,improved algorithms are proposed with subgoals based on task heuristics and the decoupling of global and local goals,which reduces the time complexity from exponential to polynomial.The effectiveness of the proposed algorithms is verified through experiments in unmanned logistic domains.The results show that the optimized algorithm significantly reduces the planning time and communication cost compared with the baseline.(4)Integrating learning to autonomous behavior planning.The autonomous behavior planning of intelligent robots also needs the ability of self-learning to adapt to scenarios lack of prior knowledge for classical planning.Based on the collaborative information collection of unmanned aerial vehicles(UAVs),this paper proposes a framework integrating planning and learning of robot behaviors.Robots are able to learn basic skills and to generate BT nodes unsupervisedly,and then autonomously plan to generate BT structures guided by rewards.The effectiveness of the proposed method is verified in UAV simulation experiments,which indicates that the UAV group can autonomously plan and learn the behavior model needed to complete the task,and its performance is competitive to control systems designed by experts,with improved robustness.