Control of Running and Walking

Paper One

Marc H. Raibert and Jessica K. Hodgins. Animation of dynamic legged locomotion. In SIGGRAPH ’91: Proceedings of the 18th annual conference on Computer graphics and interactive techniques, pages 349–358, New York, NY, USA, 1991. ACM Press.

This is an example response. To add your own response, click on 'Edit' above. Paragraphs are separated with just a blank line. This paper is interesting because... It is flawed because ... I didn't understand the following bits... Open problems are ... -- Michiel van de Panne

Similar to my comments from last class, I enjoyed this "older" SIGGRAPH paper because it reads like a good textbook as opposed to a research paper. Anyways, the main issue I've always had with control based techniques is that while they ensure physical correctness, they do not necessarily look real (that is, like an actual human). The paper sort of steals my thunder on this by saying "... animals move with a smoothness and coordination that is not required by physical realism alone". Nevertheless, this paper is good in that it provides a good initial exploration into the successful use of applying control algorithms to generate balanced walking and running motions. I'm not quite sure I understand their allometric derivations of the scale factors in Table 1. Clearly, the dimension (not units!) of velocity are LT^{-1}, but why is its scale factor L^{1/2}? -- KenRose

(Animation of Dynamic Legged Locomotion) I strongly agree with the words that the control inputs are more like “suggestions” that must be reconciled with the dynamic state and structure of the system than “commands” since that might make the locomotion more realistic. I also feel very comfortable reading this paper since it provides the basic background of control system in which I don’t have much experience. FSM is very popular in NPC motion control in games, and is simple to implement. While on the other hand, sometimes it will make the NPC act stupidly as the opponent of human player. So nowadays researchers and practitioners are trying to endow the NPCs with more intelligence so as to make the games more interesting.--Zhangbo Liu

The results seem interesting, I hope that we'll be treated to a video during the presentation today. I share Ken's confusion about the scaling factors (wrt to velocity and time). That table is neat in showing that it's not straightforward to scale a dynamic creature as one would maybe expect. On page 355 they note that they're assuming an infinited coefficient of friction to prevent slipping, but that in actuality, their models need a relatively small (?) coefficient to accomplish this. I can't remember much about the magnitudes of realistic coefficients of friction, so are the figures they quote indeed "realistic"? -- Main.Daniel Eaton

(Animation of Dynamic Legged Locomotion) The authors state that the motions described in the paper are physically realistic. It is hard to tell without seeing a animation, but I am guessing that control methods as well as any other methods could generate a non-realistic motion, and it is up to the implementation to dictate the motion “quality”. It would be interesting for me to discuss/analyze that point. Another question is whether physicaly based motion is a realistic one and vice versa. – Hagit Schechter

My only real complaint is that I'm more interested in creatures that learn to walk or develop these locomotion abilities based on some goal. Simply (or complexly) copying the control mechanisms of existing animals is somewhat interesting, but only insofar as it allows us to divine general rules for locomotion of complex systems, so we can extend beyond duplicating motion of existing creatures. (And, looking at it closer now, they do mention in passing that they hope eventually to be able to design better robots using knowledge gained in this way.) "Hard-coding" this control system with finite state machines seems only one step up from hand-animating the motion, except that it will realistically account for physics. Although perhaps learning from the ground up how to control such a complex systgem is another case of the high dimensionality of the system resulting in local minima that don't give a good global solution. --Christopher

Paper Two

Joseph Laszlo, Michiel van de Panne, and Eugene Fiume. Limit cycle control and its application to the animation of balancing and walking. In SIGGRAPH ’96: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 155–162, New York, NY, USA, 1996. ACM Press.

Another paper. Please add your comments below.

This set of papers is significantly different from what I have seen in our class so far, it introduces quite a few new concepts which I expect more elaborations on, such as what does it mean by unstable motion? Linear predictive model (section 5.2), and proportional derivative controller (PD). The idea presented in the paper is very neat, using FSM to determine next state and PD controller to compute the required force and torques which can lead the articulated figure to the desired pose. However, control perturbation is used to solve this control problem and perturbation is like trial-and-error therefore, it would still be an issue of the efficiency. -- Steven Chang

I liked the result of this paper in that it was able to apply control techniques to an articulated figure with a higher number of DOF (19). The previous paper only applied the control algorithms to models with a few DOF, which further limited their realism. I would like to know more about the selection of the regulation variables. Specifically, is this something that is done once for one class of closed-loop motion (e.g., I'll use swing COM for running) or does it have to be done for every motion? Also, the paper mentions that "the evidence for the above linear approximation is empirical". Are there closed loop motions for which perturbations result in non-linear effects? Finally, are there videos of the final animations? The paper mentions that the results still do have a robotic feel to them for the case of the 4 pose FSM. -- KenRose

(Limit cycle control) I am interested to learn if there have been significant results in controlling running and other more complex / less stable types of motion. Also, I would assume that control; methods are shared with the area of robotics research and am interested to compare the different challenges and results in both areas. -- Hagit Schechter

I'm a little confused about the term "open-loop": in section 5.1, this FSM uses foot-contact sensors to decide when to make state transitions, but isn't open-loop sensor-less (I retract this, in my skimming I didn't see the comment that it is only quasi-open-loop)? Also, what happens if one of the stages takes too long? It seems like the timing has been handcrafted to prevent that (giving each stage more than they need to complete), but what if? Could you talk a little more about why the biped falls over in figure 10? -- DanielEaton

What are the equations of motion that are generated by this commercial package? How much harder is it to extend these ideas into the "Real" world. Ie. what difficulties would we face trying to build a robot that does this, as opposed to just an idealized simulation? Why don't we see walking human robots all over the place, yet? This is a nifty technique for generating motion that seems intelligent but doesn't really have a lot of "real" smarts under the hood. How far can we perturb the system, and still have it return to the natural walk cycle without doing anything unrealistic? --Christopher

(Limit Cycle Control And Its Application To The Animation Of Balancing And Walking) “There is no guarantee that the discrete system will be smooth”, this phrase is quite right from my perspective currently, /*especially when the control inputs are randomly assigned in some reinforcement learning problems (but this issue is not related to this paper anyway).*/ This paper is more research-like than the previous one. In deed, physically-based animation, especially walking, is really difficult. Even now we are still trying to use kinds of methods to realize realistic locomotion. Seems in this paper it did not mention any statistic models which are hot in these days. So will the combination of control theory with statistic models be better?--Zhangbo Liu