My research is centered around human-robot interaction and control of bipedal robotics. More specifically, my research has three different branches:
- Utilizing preference-based learning to optimize exoskeleton user comfort and characterize the corresponding preference landscapes
- Implementing preference-based learning directly in the gait generation process for bipedal robots
- Control of a 18 degree-of-freedom lower-body exoskeleton, Atalante.
Preference-Based Learning to Optimize and Characterize Exoskeleton User Preferences
First, we developed a novel algorithm, CoSpar, that utilized methods from preference-elicitation to learn the underlying preference function of exoskeleton users. This framework was first verified in simulation, and then experimentally conducted for 3 able-bodied subjects. The experiments explored obtaining both 1 and 2 dimensional bayesian posteriors. The first publication of this work received Best Overall Paper at ICRA 2020, as well as Best Paper in Human-Robot Interaction. Later, we also extended this work to optimize over higher-dimensional action spaces using dimensionality reduction techniques.
We also extend preference-based learning to an active-learning setting to characterize the entire landscape of exoskeleton user preferences. The result of this method enables a better understanding of the underlying utility function dictating user preferences. This was accomplished by leveraging Information Gain as the method of posterior sampling. The final algorithm was termed Region of Interest Active Learning (ROIAL). ROIAL was first verified in simulation, and then experimentally conducted for 3 able-bodied subjects in which the preference-landscapes were obtained over 4 exoskeleton gait parameters.
- Tucker, M., Novoseller, E., Kann, C., Sui, Y., Yue, Y., Burdick, J., & Ames, A. D. (2019). Preference-Based Learning for Exoskeleton Gait Optimization. In 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020. IEEE ICRA Best Overall Paper Award. IEEE ICRA Best Paper in Human-Robot Interaction Award.
- Tucker, M., Cheng, M., Novoseller, E., Cheng, R., Yue, Y., Burdick, J. W., & Ames, A. D. (2020). Human Preference-Based Learning for High-dimensional Optimization of Exoskeleton Walking Gaits. In 2020 IEEE International Conference on Intelligent Robots and Systems (IROS), 2020.
- Li, K., Tucker, M., Bıyık, E., Novoseller, E., Burdick, J.W., Sui, Y., Sadigh, D., Yue, Y. and Ames, A.D. (2020). ROIAL: Region of Interest Active Learning for Characterizing Exoskeleton Gait Preference Landscapes. Under Review.
Preference-Based Learning Towards Gait Generation
Through careful construction of the Hybrid Zero Dynamics (HZD) gait generation framework, preference-based learning can be used to tune various optimization constraints. So far, we have experimentally demonstrated this framework on the planar robot AMBER-3M with two different leg configurations: rigid point feet and compliant point feet. For both leg configurations the model used for gait generation was the rigid point foot model. However, the framework was still capable of realizing dynamic, stable, and robust walking for both leg configurations – demonstrating the power of utilizing preference-based learning in the gait generation process.
- Tucker, M., Csomay-Shanklin, N., Ma, W.L. and Ames, A.D. (2020). Preference-Based Learning for User-Guided HZD Gait Generation on Bipedal Walking Robots. Under Review.
Towards Variable Assistance via Controlled Set Invariance
In this work, we proposed and demonstrated a method of acheiving variable assitance on the exoskeleton. This framework used tools from controlled set invariance, specifically control barrier functions. This framework was validated through two separate experiments. The results of these experiments showed that lower levels of exoskeleton assistence resulted in higher metabolic expenditure rates for 8 able-bodied subjects.
- Gurriet, T., Tucker, M., Duburcq, A., Boeris, G., & Ames, A. D. (2019). Towards Variable Assistance for Lower Body Exoskeletons. IEEE Robotics and Automation Letters, 5(1), 266-273.