Optimizing Walking Controllers

Contribution:
Evaluation:
Reproducibility:
Improvements:

-- MichielVanDePanne - 24 Nov 2011

Contribution
The study describes a controller for physically simulated characters. The goal of the study is to create an automatic method to define controllers for 3D human-like characters to walk more like humans. Their method does not use any motion capture data nor any reference trajectories. It also produces reasonable results for characters with different mass distributions. The authors achieve these results by parametrizing the controller, and optimizing those parameters with respect to a high-level objective function. Previous studies have used optimization on lower dimensional spaces, but 3D walking is a difficult problem since it involves high-dimensional state space, nonlinear dynamics, and a discontinuous search space due to ground contact. Hence, One of their main contributions is, given a carefully a chosen objective function and a good control parameterization, they show that their method can work for 3D human-like walking.

Evaluation
The results are evaluated based on the computed motions, for several iterations. They also compare their output animations to the related studies, such as their own implementation of SIMBICON, as well as to the mocap data. Hence, their result is compared visually and interpreted by the authors as natural looking. Effects of each individual terms are also evaluated in order to show how they really affect the system.

Reproducibility
The method being implemented can be reproduced since all the functions and the constraints related to optimization, i.e. objective function, gait, head, and body constraints, etc. are given. Moreover, the simulation parameters, and the initial simulation state, as well as the PD controller parameters are also given.

Improvements
As the authors suggest, more sophisticated controllers and objective functions could be designed and tested with the framefork, since the objective function and the optimizer are independent of the control parameters. Such further experiments would have supported the authors' claim that such optimization methods could be utilized for high-dimensional problem of human motion synthesis and control. In terms of paper writing, the sections are well-organized and the flow of the paper is easy to follow.

ozgur

-- Main.ooguz - 24 Nov 2011

(a) What is the contribution of the paper?
The authors develop a parameter-optimized 3D walking controller for humanoid characters. Their emphasis is on producing natural movement that is robust to minor environmental disturbances. Their work is significant with respect to previous work because characters exhibit more natural movement as well as operate beyond the duration for which the characters were optimized.

(b) How are the results evaluated?
Features of human walking: In comparison to SIMBICON, which the authors note was the state of the art at the time, the features of movement of the characters with respect to walking appear more natural. Effects of individual terms: The walker was optimized both with and without energy terms. With these energy terms defined, a more natural motion is accomplished. Comparison with motion capture data & SIMBICON: Motion capture data as well as SIMBICON output were used for comparative analysis. The result is that the model developed by the authors is much closer to mocap data than was their implementation of SIMBICON. Variation in body shape: Movements of various body shapes were also evaluated. The result was that actions appear to have a natural relationship between body shape, speed and walking gait. Robustness: Robustness was evaluated via a pushing experiment. It was found that robustness varies greatly across all controllers. Moreover, the authors' model was found to be less robust than their implementation of SIMBICON. Changing terrain: The characters were subjected to various terrains of varying slopes. The result was that the characters climb the slope in a natural way.

(c) Is the paper reproducible?
The authors do an exceptional job in providing algorithms, required parameters and constraints, allowing their research to (appear to) be reproducible.

(d) How could the paper research or paper writing be improved?
I would have appreciated more qualitative data in the results section. There are many instances where the evaluation was based on the author's subjective opinion (such as the characters 'appear' more natural). Of course, there is little alternative since the goal is to produce movement that is subjectively valid; but it would have been nice to see more numbers in there than there were.

-- DanielTroniak - 24 Nov 2011

Contribution The paper presents a method for optimizing a physics-based controller for a walking biped. The optimization accounts for desirable properties such as angular momentum minimization and head stability, in an effort to produce a more realistic walking motion.

Evaluation The success of the optimization method is evaluated experimentally by optimizing a SIMBICON control scheme and observing the resulting motions. A number of different criteria are used, such as comparison to motion capture data and comparison to the features of real human motion. The flexibility and robustness of the generated controllers are also tested by optimizing for different body shapes and for terrain that isn't flat.

Reproducibility The paper is reproducible, due to the high level of detail provided about the specific controller designs as well as the optimization method.

Improvements The paper is well-written and well-organized, so there is very little to improve in that respect. A minor criticism is that the evaluation of the success of the generated controllers is not the easiest to see directly, due to a lack of quantitative data and the inherent difficulty of demonstrating motion in a paper. Watching the accompanying video alleviates much of this problem, however.

-- Main.cdoran - 24 Nov 2011

Contribution :

Physics based parameters like power, angular momentum and more for biped locomotion are optimized. The optimizations are also applied to SIMBICON and mocap data. Optimized controls for other walking features like speed, shape and terrain are demonstrated.

Results Evaluation

After accounting for optimizations in domains of the controller, and other constraints, the optimization model is tested in different scenarios like features of human walking, motion capture data and by optimizing SIMBICON. Apart from that, controllers accounting for variation in body shape, robustness, terrain change and variations are built and tested.

Reproducibility

I believe the paper is reproducible. All the parameter selections are backed with proper explanations. Appropriate equations and expressions are shown wherever necessary and/or are linked to corresponding references.

Improvements:

Overall the paper is well organized and written. Both qualitative, quantitative and visual explanation of results are done in the end for different kind of optimizations. Scope for improvements and future work is also provided.

-- Main.sumanm - 24 Nov 2011

The paper introduces a modification of SIMBICON optimized under additional constraints and terms inspired by human motion. These constraints include power efficiency, the balancing of angular momentum, head stability, and stability involving foot planting. The optimization of such a tricky objective function was shown to be possible via use of Covariance Matrix Adaptation, and resulted in a smoother and more convincing walking controller for each variation in body type.

The results were contrasted with the original SIMBICON output and mocap data, and tested for robustness testing the controller's ability to recover when pushing the walking character. The paper provided sufficient information regarding the calculation of the objective function and software environment for reproducibility -- although the computational demand may be prohibitive (the researchers used 19 Xeon processors to run the simulation in parallel). The paper can be improved by providing a light overview of the implementation issues involving with the modified SIMBICON and ODE.

-- KevinWoo

* Contribution: The paper optimized the SIMBICON controller [Yin et al.2007] in terms of power minimization, angular momentum minimization etc. It does not rely on motion capture data and can produce human-like qualities of motion of walking.

* Evaluation: The research is evaluated by variety of experiments. They run a number of features of normal human walking in comparison of with and without an individual optimization term, such as power ratio term, angular momentum objective. They also compare the result with motion capture data. Optimization is also applied to the SIMBICON body to show improvements. The experiment also tests the optimization by using different body shapes and changing terrain. Robustness is evaluated by applying an external force.

* Reproducibility: The paper should be reproducible since the implementation and experiment are explained in details.

* Improvement: The structure of the paper is easy to follow and overall I have no problem to read it.

-- Main.Baoxuan - 24 Nov 2011

Contribution: The paper proposes a method for optimizing parameters of SIMBICON controller. The goal of the optimization is to produce more natural walking. The optimization function consists various task constraints and biomechanical features of human walking.

Evaluation: The method was compared to standard SIMBICON, which was claimed to be state-of-the art technology and showed more realistic results, such as ankle push-off strategy instead of hip push-off one. Next, the the influence of each term in the optimized function was demonstrated. Also, the authors carried out a comparison with motion capture data. The claim to be closer to mocap data than SIMBICON, but no real comparison is given (probably because it's just hard to compare?).

Reproducibility: The paper seems to be easily reproducible, but you never know if they unintentionally omitted value of some parameter...

Improvements: The authors note themselves that their model is less stable than SIMBICON, so maybe a different parameterization of controllers will solve this. And of course the motion is still far from motino capture data, but I'm not sure there are any papers demonstrating anything close to that.

-- MikhailBessmeltsev - 24 Nov 2011

Contribution? It introduces a parameter optimization procedure for 3D walking controller synthesis. It makes physics-based walking motion more natural. The optimization can be done on various body shapes.

How are the results evaluated? 1) The resulted walking motion is shown to reflect some features of real human walking. 2) It compares the walking motions with and without some optimization methods. It is shown that motions are more realistic with those optimization methods. 3) The resulted walking motion is compared with mocap data to see how close they are. 4) Walking motions of various body shapes are compared in order to see if they all look natural 5) Robustness of walking are tested. 6) Walking is tested in different terrains.

Reproducible? Yes. The model, equations and parameters are well documented. It should be editable.

How could the paper research or paper writing be improved? There seem to be too many parameters to tune. This might make it difficult for the method to be extended.

-- Main.shuoshen - 24 Nov 2011

Contribution: It provides a way of defining the control parameters for the SIMBICON strategy based on metrics such as energy, angular momentum and velocity during the walking

Results evaluated: It is evaluated by comparing the walking with and without the optimization step, also using motion capture data and under changes on the terrain.

Reproducible: Yes, cma is a standard optimization process and simbicon code is available

Writing: I found it simple to follow and understand

-- Ernesto Torres

-- Main.etorresv - 24 Nov 2011

Contribution: The authors present a procedure for optimizing the stability and other performance metrics of walking controllers based on the SIMBICON system. They describe a number of biomechanically motivated error terms that aim to discourage unnatural gaits, then minimize the combined error using Covariance Matrix Adaptation to tweak a character's DOF parameters.

Evaluation: The authors run several experiments, describing both quantitative and qualitative results. They compare the resulting walking style with and without the influence of various of their error terms. Helpfully, they quantitatively plot some interesting aspects of walking gaits for a normal SIMBICON implementation, for their optimized controller, and for averaged motion capture data. Finally, they test robustness of the walk on varied body types and terrains, and with random pushes on the character.

Reproducibility: The authors provide significant details on their error term formulae, as well as the values of most of their non-optimized parameters. Additionally, their code implementation is available on the project website. As such, reproduction would not be overly difficult for their solution.

Improvements: Generally speaking, the writing was clear and descriptive. To nitpick one aspect, while the authors provide reasonable and straightforward descriptions of the structure of their error terms, the error thresholds (epsilon's) and error combination weights (w's) are given without any justification. How much hand-tuning was required of these parameters before achieving reasonable results?

-- BenHumberston - 24 Nov 2011

Optimizing Walking Controllers - David Matheson

(a) Contribution: This paper introduces parameter optimization for full-body 3d walking controllers that consider body shape, step length as well as key biomechanical properties. Using biomechanical principles to construct more human motion constraints and objective functions results in more realistic looking motion with no reference data (ie no mocap).

(b) Evaluation: In general the evaluation is very good. There is a breakdown of the effect of each term and a quantitive comparison of key metrics with mocap and previous simbicon controllers. The evaluation of robustness and changing terrain was a good summary. More details would have made the paper too long. I would have liked to seen a more quantitative explanation of why CMA was the best optimization algorithm.

(c) Reproducibility: All constraints and terms of the objective function are laid out in great detail. The parameters of the simulation are provided. Therefore it should be possible to reproduce this paper and obtained the same experimental results.

(d) Improvements in Research/Writing: The paper is very clear and easy to follow. It does a good job explaining all the constraints on each joint. It may have been easier to follow all the constraints if they were presented in a table where each row contained a joint and its constraints.

-- Main.davidm - 25 Nov 2011

(a) Contribution: This paper builds on the SIMBICON approach by introducing a semi-automatic way to optimize for parameters that yield a more "natural" looking gait. They also come up with their own objective function and justify/analyze the influence of every term.

(b)Evaluation: Most evaluation is "looking whether it looks natural". They do evaluate the influence of each term in the objective function, but they never once use any kind of numerical data. At least they do compare against motion reconstructed from motion capture data.

(c) They provide code, explicit formulas, and numerical values for their parameters. It seems reproducible enough at first glance, but a fair amount of knowledge of how SIMBICON works may be required.

(d) I think some numerical data might be helpful in tuning their own objective function (not only the parameters, but the actual terms). The point is not to use motion capture data, but they could still do so in order to obtain a better function (and then not use it at runtime) which would be more robust and perhaps simpler.

-- Main.ginestra - 25 Nov 2011