| 研究課題/領域番号 |
13F03796
|
|---|
| 研究機関 | 独立行政法人産業技術総合研究所 |
|---|
研究代表者 |
吉田 英一 独立行政法人産業技術総合研究所, 知能システム研究部門, 連携研究体長 (30358329)
|
|---|
| 研究分担者 |
ESCANDE Adrien 独立行政法人産業技術総合研究所, 知能システム研究部門, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2013-04-01 – 2016-03-31
|
| キーワード | 最適化 / ヒューマノイド |
|---|
| 研究実績の概要 |
The work done from May 2014 (April was a month off for parental leave) to March 2015 is as follows.
For my main topic of hierarchical optimization, I continued to investigate hierarchical optimization with non-linear constraints and the development of a dedicated solver. The results of FY2013 were encouraging and many choices were validated but in case of conflict between constraints (what is the case of interest in hierarchical optimization), the convergence speed was linear, which is not very good (classical SQP have super-linear convergence speed). Investigation showed that second order derivatives could not be ignored and I wrote a better version taking them into account. As a result, convergence speed was vastly improved.
I also investigated the stopping criterion in case of conflicts because in this case the classical criterion of constraint optimization is not necessary correct. Hierarchical non-linear optimization also has a lot of links with priority-based control schemes. I show those schemes can fail and using those links, I studied the theoretical reasons behind these failures.
Concerning constrained optimization on manifolds, with a PhD student, I developed a sequential quadratic program (SQP) solver which is able to work with variables living in non-Euclidean spaces. This eliminates the problem of singularities in the parametrization of such variables, resulting in more stable applications based on optimization scheme.
With another PhD student we finalize the investigation on the parametrization and use of forces in multi-contact trajectory generation.
|
| 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
This is the status for the 3 main points of my original JSPS proposal:
1) Robust initialization: a bit behind plan
2) Collision constraints: behind plan
3) Dedicated search spaces: more than planned
In summary, I estimate that the research plan advances as expected.
|
| 今後の研究の推進方策 |
The plan till the end of the grant (end of October) is divided in three main topics that mirror the topic of the original proposal.
The remaining work of (i) hierarchical optimization here is to write the paper, with two main parts: the writing in itself and the generation of good examples to illustrate the various points within. In particular, the targeted applications in the paper are generalized inverse kinematics and model predictive control. Some work needs to be done for the coding of such applications. Less time consuming works include reading a few more references and polishing some mathematical formulations.
Concerning (ii) posture generation, aside from some more rework of the SQP solver on manifold itself the main goal now is to use the solver for robotics applications. One of these applications is finding admissible postures of the robot respecting a set of constraints. We started working on these so-called posture generator. It should have a good implementation of it with a framework allowing easy extensions at multiple levels. Second part is to find new formulation of the constraints, in particular the contact constraints, to allow for more generic usage.
(iii) Collision constraints will be a preliminary work. I intend to begin with studies on simple examples to show the interactions between an optimization routine and the computation of distance for collision avoidance constraints. I will then investigate what kind of pseudo-distance could be used instead of the Euclidean distance, to decrease the computation time, and see what accelerating data structures.
|
