Biped Motion Planning For Humanoid Robots In Complex Environment

The motion planning of humanoid robots is a complex and nonlinear planning problem, and the walking pattern generation is an important challenge in robotics re-search in recent years. We should consider both stability and feasibility for biped mo-tion planning, which means that the robot should keep its balance and avoid unneces-sary collisions with the environment. When we finish the design of walking motions, we also need to adjust walking patterns in order to realize similar walking patterns on real robots as we desired. Although researchers have given several solutions for the aforementioned problems, a lot of detailed checking and adjusting are complicated and rely on developers'experience. Therefore, it is desirable for us to develop intelligence approaches to generate biped walking motions for humanoid robots.In this paper, we propose a universal biped walking motion planner for humanoid robots. The planner plans smooth and flexible walking motions on rough terrains, in-cluding stairs, slopes, and obstacles. In addition to generating collision-free patterns while keeping the balance of the robot, this planner also checks whether the patterns are achievable for the robot. We briefly introduce the developmental history of humanoid robots, the main approaches of walking pattern generation, and the main strategies to plan walking motions in complex environments. Moreover, we introduce the major challenges for humanoid robots to plan biped motions.The reason our motion planner can generate the walking patterns with considering the feasibility is that our approach is base on the'Simplified Walking'mechanism. With this mechanism, we can specify the complete biped motions with the unit of a'walking step', instead of a'time step'of the robots'working time used in usual methods. Then, the continuous planning problem in three-dimension space is converted to the discrete decision problem in the horizontal plane. This leads to a high reduction of the solution space, and make us can plan long-distance biped motions with more constraints in a low computation cost. Although simplified walking is a theoretical solution of biped walking, it can be converted to walking patterns which can be applied on real robots. Part of the CoM trajectory is converted to generate double-support phases. Then, we can implement the vertical movement for the robot to walk on stairs and slopes. We also consider the multi-body physical properties to adjust the CoM trajectory to improve the stability of the biped walking. We demonstrate the converted walking patterns on real humanoid robots. Moreover, we introduce how to design periodic biped motions and how to smoothly connect different walking motions together. The conditions to realize biped motion connection are discussed in this paper.Based on simplified walking, we present two stochastic search algorithms:the first one, named ZMP Sampling Search, calculates the simplified walking solution according to a given feasible footstep plan;the second one, named Sampling Based Footstep Re-vision, revises an unfeasible footstep plan to a feasible one. The ZMP sampling search is an efficient method to determine the walking solution while simultaneously consid-ering dynamic balance and feasibility to make sure each walking pattern is achievable for the robot. The convex expansion technique is proposed to reduce the uncertainty of our stochastic search. The footstep revision algorithm is firstly introduced to improve the intelligence of walking planning, and make robots are able to adjust the footsteps according to their construction limitations. We present our full walking planner and the three-stages planning strategy in this paper. Complex walking tasks are demonstrated on simulated and real humanoid robots.In recent years, a few of international institutes have developed several high per-formance humanoid robots, and realize several flexible biped motions on these robots, such as running, dancing and climbing on stairs. However, these robots are expensive, because their motion capabilities relys on high-grad hardware and the precise produc-tion. But it is showing a new direction for robotics research, which is to explore higher quality motions on low cost robots with the progress of intelligence techniques. There-fore, multiple techniques are introduced to calibrate the walking patterns for low price real robots. The humanoid robot Nao is employed for our research and experiments. Due to the reason there exists a shared joint in Nao's pelvis, it is more difficult for us to solve the inverse kinematics problems. We present the method to calculate the robot's detailed status, including the position, the velocity and the acceleration of each joint by using inverse kinematics. Pi-sigma neural networks are employed to compute the torque of each joint when the robot is walking, and to compensate the elastic deformation in real-time. The simulated annealing algorithm is used for the robot to calibrate the coef-ficients of elasticity. Moreover, we introduce the method to calibrate the CoM trajectory by using a ZMP compensation trajectory. At last, we build up a feedback controller with the observation from CoP measurements, in order to implement the close-loop control for robots. The calibrations and the effect of our feedback controller are demonstrated on real robots.There are mainly four contributions of this paper. First, we introduce the'Simpli- fied Walking' mechanism which describes the biped motions briefly. The mechanism highly reduces the computing cost of torso trajectory generation and makes it is possi-ble to connect biped motion smoothly and to plan long-distance walking motions with considering multiple constraints. Second, we firstly check the walking patterns in de-tail. By using efficient stochastic search algorithms in planning stage, we can make sure the walking patterns generated are achievable and have no potential harm to the robot. Third, we implement a universal biped motion planner which makes robot plan footsteps automatically and generate stable patterns in complex environment. The plan-ner also adjusts unfeasible footsteps for robots to make it possible to generate walking motions. At last, we design multiple calibration approaches for humanoid robots. By using intelligence computing methods, we compensate the elastic deformation and the ZMP difference to realize automatic calibrations, in order to reduce the effect of the hardware differences of robots. Moreover, we also present the lack of our works to provide a reference to the other researchers.