| 研究実績の概要 |
In FY2016, we focused on two objectives: First, a literature review about motion planning with focus on humanoid robotics. Our goal is to apply motion planning algorithms to the problem of real-time planning for humanoid robots with multiple contacts. Our conclusion from the review is the following: To achieve real-time planning we need to abstract away unnecessary details. In particular, planning more than a few contacts into the future seems inefficient, since (1) the environment might change, (2) the sensor information might be faulty (objects suddenly appear or disappear), or (3) the robot experiences drift. In all those cases we would reject the planned contacts and start again from the scratch. This represents unnecessary overhead and wasted resources. The idea is to have an exact path near the robot, but the farer we move away, the more abstract the path has to become. The exact formulation of this idea is the focus of my current work.
Second, We familiarized ourselves with two important tools for motion planning, the physical simulator Klamp't developed by Duke University and the motion planning framework OMPL developed by Rice University. The simulator Klamp't is state-of-the-art for physical contact planning, and it is currently the fastest physics simulator to compute contacts between mechanical systems in near real-time. This is usually a demanding process taking several minutes to simulate. The planning framework OMPL is state-of-the-art for general motion planning algorithms.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
We have made a broad review about motion planning methods allowing abstractions that can be applied to changing environments including homotopic change for efficient motion planning, and found that this kind of motion planning has not been well investigated.
Through discussions with the Postdoctoral research fellow together, to achieve that goal of efficient planning of multi-contact motions in challenging environments, we conclude that we will pursuit an approach of adaptive abstractions of the environments. In this approach, we extract the important geometric features of the environments in a reduced expression. Also, we have identified and tested some useful tools such as Klamp't and OMPL to implement and validate the algorithm. We have achieved basic agreement we are in the right direction and started building basic algorithms.
In this manner, we estimate that we are advancing for the goal and expect to construct the basics of the algorithm as planned in the project.
|
| 今後の研究の推進方策 |
The theoretical goal of this research is to develop an efficient motion planning algorithm, which can be applied to any mechanical systems using contacts for locomotion. Depending on the runtime we like to apply it to dynamically changing environments. The key contribution of this motion planning algorithm will be a theory of abstraction, which allows the planner to abstract away certain details of the environment and/or robot geometry, thereby allowing faster and more reactive motions. The current idea is to shrink the workspace of the robot, thereby extracting the topological features to guide the planner. The goal is to formulate this abstraction process mathematically and show its relation to the completeness property of motion planning algorithms.
|
