SigOpt offers our software free to academics, nonprofits, or labs who are performing research they plan to publish. Dozens of publications across a wide variety of disciplines have utilized SigOpt through this program. Learn more about the SigOpt Academic Program or sign up for free access to SigOpt.
Researchers from the Lauflabor Locomotion Laboratory at TU Darmstadt applied SigOpt as part of an experimental approach to robotic hopping at specific hopping heights that resulted in more energy-efficient and human-like hopping patterns than previous researches on the same robot. In this interview, we discuss this approach and their findings, both published in an IEEE paper, with this team.
What is your research subject?
Creating dynamic legged motion in robots is a challenging task that requires knowledge on both the mechanics and control, plus optimizing a variety of parameters for getting a desired behavior. Our goal with this research is to explore the potentials of having an electric-pneumatic actuation as a bioinspired actuation method in terms of creating human-like and also energy-efficient motion. More specifically, we explore a method that combines the high bandwidth of an electric motor with the variable compliance of a pneumatic artificial muscle (PAM) to make up the so-called Electric-Pneumatic Actuation (EPA) system.
There are a few novel components of this approach. First, this system is the first to include a ground reaction force (GRF) based controller for the knee joint, named FMCK. Second, it uses a PAM as an adjustable compliance to improve movement efficiency. Finally, the approach to experimentation was also novel, as we used SigOpt as a hosted solution for Bayesian optimization that was applied to obtain parameters minimizing energy consumption while maintaining stable hopping at specific target heights.
For whom is this research most valuable?
This research is of interest to a variety of audiences. First, anyone in legged bio-inspired robotics can benefit from it and build upon it, as it proposes the advantages of having both electrical motors along with PAMs. From a control system standpoint, this research also offers a novel validated control approach to robotic control systems that could be applied to a wide variety of other robots with higher degrees of freedom. Finally, this research is also relevant for anyone who needs to implement a sample-efficient approach to experiment design. By applying Bayesian optimization via SigOpt to these experiments, the team was able to reduce the number of trials required to optimize the control parameters, which improved wall-clock time and resource efficiency. This type of methodology is easily transferable to other domains, so could have broad applicability.
What made you interested in it?
Legged locomotion in robotics is a particularly challenging task that requires deep understanding of both mechanics and control in a fluid and complementary way. Given this complexity, there is plenty of room for improvement in how robots move and how human-like or energy-efficient their movement is. And methods that draw on biological inspiration have begun to gain traction, but have not been thoroughly explored. So this research subject is particularly novel and broadly applicable across legged robotics, hitting on a variety of relevant themes at once.
How did you design the experiment?
We designed our experiments to show the controller’s ability to achieve stable hopping at different heights and investigate the influence of PAM on energy efficiency. To do so, we optimized the control parameters with SigOpt to achieve maximum energy efficiency while maintaining the desired hopping height across nine different experiments. In this context, we defined two distinct objective metrics that we wanted to optimize and balance at the same time. The first is hopping height and the second is energy consumption. For each experiment, we were trading off height and energy consumption at the same time. These nine experiments explored the pairwise combination of three different PAM pressures (0, 400, 600 kPa) with three different hopping heights (6, 10, and 14cm). In each experiment, the robot hopped at least 15 different times (e.g., 15+ trials per experiment). We also implemented verification trials that repeated and reproduced the results from the optimal parameter trial at least ten times. For all of these trials, we applied SigOpt to select the next configuration of parameters to test so we could more efficiently uncover the best-performing control parameters.
What were the results of your experiments?
There were three categories of results from this set of experiments. First, we found optimal parameters for the FMCK controller to validate the optimization procedure and capability of the FMCK control functionality. Second, we showed that this bioinspired system is capable of demonstrating human-like and efficient hopping – thanks to the presence of PAMs – which points to future potential direction for additional studies. Finally, we also showed that you can optimize the tunable parameters of this control system to minimize energy consumption while achieving the desired hopping height.
Why did you select SigOpt to optimize these models?
Bayesian optimization is a well-known method for sample-efficient experimental design. By incorporating information from prior trials into suggestions for future trials, this method leads to experiment convergence much faster than other naive experiment design methods. This sample efficiency is particularly important when running a real-world trial that is time-consuming and can be resource intensive.
SigOpt offers a hosted solution for Bayesian optimization that makes it easy to design these experiments, run them efficiently and visualize the results in a compelling web dashboard. Without SigOpt, we would have had to install an open source Bayesian optimization package, set up servers to run these packages, troubleshoot any errors or bugs, set up our own system for logging, and create custom visualizations to understand each trial. SigOpt did all of this for us, making it easy to run and gain insights from these trials on a much faster timeline.
We did not explore whether SigOpt’s implementation of Bayesian optimization was more sample efficient or better performing than other methods, but we did find that it led to convergence for our problem more quickly than we expected, in fewer than 25 trials for certain hopping heights.
Finally, we found the capacity SigOpt has for handling multiple metrics to be a huge advantage for real world trials. You are typically trading off multiple metrics rather than exploiting just one, so doing so easily in this product was a significant benefit.
How would you characterize the benefit of using SigOpt?
I’ll build on what I stated above. First, SigOpt implements Bayesian optimization, which is a sample-efficient approach to experimental design that allows us to optimize these control parameters with fewer trials than is otherwise possible. Each trial takes considerable time, energy, and resources from our team, so minimizing the number of trials is a huge benefit.
Second, SigOpt tracks the full experiment history in a dashboard. This made it easy for our team to view, analyze, comment on, and collaborate on these experiments asynchronously without having a manual process for developing and sharing results from each experiment. This dashboard also made it easy to validate the results from our experiment and share them with external collaborators who were reviewing our work.
Finally, SigOpt is a hosted solution that you implement via API, which makes it much easier to support than other methods in this space. Most other Bayesian optimization packages require that you install them and run a server to support their computation. The fact that SigOpt solves all of this for you saves you compute resources, but, more importantly, time that you would otherwise spend managing and debugging the optimizer itself.
What would you most love to see as an improvement to SigOpt’s product to assist you in further research?
In this case, we applied SigOpt to optimize multiple metrics that we could easily compute for these Bayesian optimization experiments. I know you can use SigOpt to optimize two metrics at the same time, which is useful in real-world settings like robotics. But I’d also love to be able to run purely exploratory experiments that do not optimize a specific metric, but instead explore model configurations that satisfy a variety of metrics at the same time. We often have 5+ metrics that we care about for a given experiment, and having the ability to uncover a variety of models that satisfy my particular constraints for each of these metrics would be a useful addition to the SigOpt library (beyond what is possible with metric constraints today). This type of search would go beyond Bayesian optimization into more of an active search area.
How do you expect to continue evolving this research in the future?
This research holds interesting implications for how biological insights can inform robotics design. We demonstrated that drawing inspiration from biology in how we design and control the system for controlling hopping movement can lead to more human-like, efficient locomotion. We think there is opportunity to go much deeper on this research front and look forward to collaborating with other researchers who push this space forward.