Motion Planning for Humanoid Robots: Highlights with HRP-2

In this paper we outlook current state of the art in the emerging research field of motion planning for humanoid robots. We will first introduce related research activities in several aspects of motion planning for humanoid robots, which include integration of dynamics and other 3D motion planning for wholebody motion and various tasks. We also mention locomotion planning and local whole-body dynamic motion generation. We next present two related ongoing research projects conducted at JRL-France. The first part shows results 3D wholebody humanoid motion planning for simultaneous locomotion and dynamic manipulation. A two-stage planning method is adopted to generate for stable locomotion and task execution, by integrating geometric and kinematic motion planner and dynamic pattern generator. The second work provides a general framework of task-driven whole-body motion generation including stepping. In that research tasks are given to the IK solver to generate whole-body motions by taking account of monitored criteria and task priorities. A dynamic walking pattern generator provides the stepping motion that is combined with task achieving motion using the same IK solver. Finally we comment perspectives for future direction of motion planning for humanoid robots. I. MOTION PLANNING FOR HUMANOID – INTRODUCTION Humanoid robots are expected to perform complicated tasks thanks to their high mobility and many degrees of freedom (DOF) including legs and arms. Their anthropomorphic configuration gives another advantage that they can easily adapt to machines or environments designed for humans. One of the key issues to fully exploit the capacity of humanoid robots is to develop a methodology that enables them to explore and execute various dynamic tasks, requiring dynamic and smooth whole-body motion including collision avoidance and locomotion, like object carrying tasks. In the field of motion planning, advancement in probabilistic methods has greatly improved planning of the threedimensional (3D) motion of mechanism including complicated geometry and many degrees of freedom. However, most of those methods are based on geometric and kinematic planning in configuration space whereas dynamic control is required for humanoid motion planning in workspace to execute tasks by keeping its balance. Concerning control issues of humanoid robots, stable motion pattern can be generated efficiently thanks to the progress in biped locomotion control theory, basically based on ZMP (zero moment point [1]) control. ∗Staying at LAAS-CNRS as a co-director of JRL-France Planning of 3D humanoid motion for tasks in complex environments must definitely benefit from those two domains. Indeed, integration of dynamics into geometric and kinematic motion planner is a challenging topic. In this paper, we first outlook the recent research trend in motion planning for humanoids in Sections II and III for consideration dynamics and other related issues. Section IV and V show researches on locomotion planning and wholebody dynamic motion generation. After brief introduction of research activities at JRL-France in Section VI, we present two of our research results, a 3D motion planning for simultaneous locomotion and manipulation and whole-body motion generation including support polygon reshaping are presented in Sections VII and VIII. Section IX gives several perspectives. II. HUMANOID MOTION PLANNING WITH DYNAMICS As stated earlier, recent progress in motion planning research now allows us to solve a three-dimensional (3D) collision avoidance motion of robot with complex structure with many DOFs in cluttered environments. This evolution attributes mostly to development of efficient algorithms based on sampling-based planning method such as Rapidly-exploring Random Trees (RRT) [2], [3], Probabilistic RoadMap (PRM) [4] and all their variants (See [5], [6] for two recent overviews). It is natural to have an idea to extend this method for humanoid robots. In this case, the essential issue are to face the high number of of DOFs and to incorporate dynamics to geometric and kinematic motion planning scheme. In [7] Kuffner classified the possible methods into three classes to deal with dynamics. The first one is a two-stage approach. In the first stage, a “path” is generated by geometric and kinematic planning, which is validated by considering dynamics to transform it into a dynamically executable “trajectory.” The second method is based on state-space approach which attempts to include the derivative. This unified approach takes account of dynamic constraints implicitly during planning via state transition equations. The third is any hybrid methods combining those two methods or using sensor information. Although the state-state approach looks nice, this approach doubles the state space and has not really been applied to humanoid motion planning. Two-stage approach has therefore been investigated intensively these years. Fig. 1. Balanced reaching motion generated in [8] Kuffner and his colleagues initiated this area and have been actively working on this topic. They proposed a method to generate collision-free dynamically-stable humanoid motion using RRT [8] through a two-stage method (Fig. 1). In this method first statically stable configurations are pre-computed and the random planner makes a search to find a collision-free path from initial to final configurations. After the planning, the path is then transformed into minimum-jerk trajectory and smoothed by verifying dynamic ZMP constraints. If the collision is detected the trajectory is slowed down towards the statically stable configurations. In this research, support state (double or single support) does not change and locomotion is not included. In our research we developed a framework in order to generate 3D collision-free motions that take account of both the locomotion and task including dynamics [9]. We will provide some detailed description of this method in section VII. III. MOTION PLANNING FOR HUMANOID – OTHER ISSUES There are a number of other topics where motion planning is concerned for humanoid robots. Probabilistic planning method can take advantage of its efficiency for many DOFs for various types of problem. Kagami et al. [10] proposed a motion planner for humanoid arm trajectory using RRT based on stereo vision. As a humanoid robot has a many links, it is important to detect rapidly if the desired motion can be executed safely. For this purpose self-collision avoidance are addressed [11], [12]. Not only collision avoidance but also motion planning taking advantage of contact is now being investigated. Hauser et al. [14] proposed planning of “non-gaited locomotion” of humanoid to go through a rough terrain using locomotion by supporting its body with hands and feet (Fig. 2). Escande et al. [13] presented a method of planning support contact points so that the humanoid robot can accomplish tasks that require multiple contacts, like taking a distant object on the table by supporting its body by one hand. Fig. 2. Non-gaited humanoid locomotion [14] Manipulation planning is another interesting topic. Stilman studied a method for “navigation among movable obstacles” that allows the humanoid to reach the goal by displacing interfering obstacles [15]. Okada et al. developed a software framework for high level motion generation for the robot to move in a daily-life environment including 3D geometric model based action/motion planner and runtime modules contains 3D visual processor, force manipulation controller and walking controller [16]. IV. LOCOMOTION PLANNING It is an important issue how the locomotion is planned for humanoid robots. Research on bipedal walking has a long history and now walking pattern generators have been proposed (for example, [17], [18], [19], [20]) based on ZMP control whose reliability has been verified through hardware experimentations. Kagami et al. and Gutmann et al. proposed on-line foot step planning methods using the environment model built by observation for human-size H7 robot [21] and smaller QRIO robot [22] (Fig. 3). Kuffner et al proposed a footstep planner to determine the footprints in rough terrain [23]. Humanoid robots can also make use of its many DOF to go around various environments by changing its locomotion modes. Fig. 3. On-line path-planning for navigation on rough-terrain [22] A method was first proposed by Shiller et al to allow a digital actor go through narrow spaces by switching walking and crawling locomotion using the environment model which is given a priori [25]. Kanehiro et al. proposed a method that allows the humanoid robot to go through constrained spaces by switching between various predefined locomotion styles (normal biped walking, walking with twisting its waist, side stepping, walking with bending knees deeply and crawling), using a simulated 3D view [26] (Fig. 4) and 3D grid map builder [27]. V. WHOLE-BODY DYNAMIC MOTION GENERATION Finally, the generation of whole-body dynamic motion is closely related to motion planning. Whereas the motion planning takes charge of global plan from initial to goal configurations, a whole-body motion generation concerns how to make valid local motions by taking account of several constraints. So it is important in order to create feasible dynamic trajectories from motions that have been provided by the global motion planner. Khatib and his colleagues have been working on dynamic motion generation for humanoid robots by using task specification in operational space approach [28]. In their work a hierarchical controller synthesizes whole-body motion based on prioritized behavioral primitives including postures and other tasks in a reactive manner. Kajita et al. proposed a “resolved momentum control” to achieve specified momentum by whole-body motion [29]. Mansard et al. [30] proposed a task sequencing scheme to achieve several tasks including walking and reaching at the same time. We have also developed a taskdriven support polygon reshaping [31] that enables a wholebody tasks as introduced later in section VIII. There are several task-specific whole-body motions that have been intensively investigated: pushing [32], [33], [34], and lifting [35], and pivoting [36]. Currently, many researchers are intensively working to integrate those recent developments with global motion planner. Fig. 5 shows an example of planning of pivoting manipulation to move a bulky object to final position. VI. RESEARCH IN JRL USING HUMANOID PLATFORM One of the main research areas in JRL-France at LAASCNRS has been also motion planning for humanoid robots using software and hardware platforms HRP-2 [37] and OpenHRP [38]. We have also been working on human-robot Fig. 4. Passing narrow space by changing locomotion style [26] Fig. 5. Planning result of pivoting with HRP-2 humanoid robot holding the box. Arm configurations are calculated using inverse kinematics from fixed contact points on the box. interaction, software architecture and control issues. Introduction of projects in JRL-France in French is found in a web page [39]. In the following two sections, we present our work on whole-body motion planning for humanoid robots. In the first part, a two-stage iterative planning framework is introduced where geometric motion planner and dynamic pattern generator interact by exchanging the trajectory, to obtain 3D whole-body dynamic motions simultaneous tasks including locomotion, in complex environments. The second part describes a task-driven motion generation method that allows a humanoid robot to make whole-body motions including support polygon reshaping to achieve the given tasks. VII. 3D MOTION PLANNING FOR SIMULTANEOUS MANIPULATION AND LOCOMOTION We have proposed a two-stage planning framework [9] based on the geometrical and kinematic planning technique [40] whose output is validated by dynamic motion pattern generator [19]. Using proposed planning framework, we could obtain 3D whole-body humanoid motions for execution of dynamic task in complex environment. The main contribution of our approach is to cover both manipulation and locomotion tasks in a single unified framework. A. Two-stage planning method Fig. 6 illustrates the two-stage planning method we have proposed. At the first stage of motion planning (upper part in Fig. 6), the geometric and kinematic motion planner takes charge of generating collision-free walking path described by the position and orientation (X , Θ) of the waist for a bounding box approximating the humanoid robot, as well as the upper body motion expressed by joint angles (qu) to achieve desired tasks. Here we assume that robot moves on a flat plane with obstacles. Then at the second stage, those outputs is given to the dynamic pattern generator [19] (lower part in Fig. 6) of humanoid robots to transform the input planar path into a dynamically executable motion described by waist position and orientation (Xd, Θd) and joint angles of whole body (q) at sampling time of 5[ms] by taking account of dynamic Dynamic pattern generator Upper body motion