CON-HUMO - Control based on Human Models
CON-HUMO focuses on novel concepts for automatic control, based on data-driven human models and machine learning. This enables innovative control applications that are difficult if not impossible to realize using traditional control and identification methods, in particular in the challenging area of smart human-machine interaction. In order to achieve intuitive and efficient goal-oriented interaction, anticipation is a key. For control selection based on prediction, a dynamic model of the human interaction behavior is required, which, however, is difficult to obtain from first principles.
In order to cope with the high complexity of human behavior with unknown inputs and only sparsely available training data, we propose to use machine-learning techniques for statistical modeling of the dynamics. In this new field of human interaction modeling – data-driven and machine-learned – control methods with guaranteed properties do not exist. CON-HUMO addresses this niche. Key methodological innovation and breakthrough is the merger of probabilistic learning with model-based control concepts through model confidence and prediction uncertainty.
The ambitious goals of CON-HUMO are structured in three research blocks:
- The design of Bayesian system identification for human interaction behavior.
- The analysis of control properties of learned Bayesian dynamical systems.
- The synthesis of constrained control based on learned Bayesian dynamical systems.
Bayesian non-parametric models of human interaction behavior
Classical system identification methods approach the characterization of dynamics by assuming a parametric model which is adapted depending on observed data. In contrast, Bayesian non-parametric models are a suitable alternative for the identification of non-linear dynamical systems. Instead of computing optimal parameters, for instance Gaussian process(GPs) based models only define the correlation between observations and predictions naturally result by applying Bayesian inference. They are also compelling due to their ability to provide prediction confidence bounds. Additionally, incorporating available knowledge gives GP priors a structure in accordance with the real function to approximate. In this research area we aim for studying Bayesian models that incorporate a priori properties of human behavior, from well-known sensorimotor structures to desirable control properties.
Control properties of learned Bayesian dynamical systems
Over the last decade the use of Bayesian models such as Gaussian Processes (GPs) has grown significantly in the control community. Due to several advantages of GPs like the small number of parameters and the data smoothing behavior, it is an easy-to-use technique for the interpolation of nonlinear functions. This approach has been proven beneficial for model-based control algorithms, e.g. Model Reference Adaptive Control (MRAC) or Model Predictive Control (MPC). However, the analysis of the closed-loop dynamics requires a closer look to the inherent properties of GP-based dynamical systems. Although widely applied for control, the current characterization of their control-related properties is very limited. Here, we aim for a systematic specification of control-relevant characteristics like stability, controllability and observability of GPs. These results are instrumental for the design of robust controllers based on GP-based Dynamical Systems.
Constrained control design based on Bayesian dynamical systems
The deployment of automated systems based on Bayesian dynamical models poses novel challenges in control design. When Bayesian priors are involved, control strategies relying only on expected dynamics/constraints are likely to perform poorly as they neglect valuable information encoded in model uncertainty. Besides expected values, Bayesian dynamical models also provide confidence levels reflecting potential variations. This level of uncertainty is an a priori indicator of potential prediction errors and should form a decisive factor for control. To account for this valuable additional measure, we explore systematic control approaches that account for variability while maintaining the desired control-related properties such as stability and constraint satisfaction.
T. Beckers; J. Umlauft; D. Kulić; S. Hirche: Stable Gaussian Process based Tracking Control of Lagrangian Systems. Proceedings of the 56th Conference on Decision and Control (CDC), IEEE, 2017
T. Beckers; S. Hirche: Equilibrium Distributions and Stability Analysis of Gaussian Process State Space Models. Proceedings of the 55th Conference on Decision and Control (CDC), IEEE, 2016
J. Umlauft; T. Beckers; M. Kimmel; S. Hirche: Feedback Linearization using Gaussian Processes. Proceedings of the Conference on Decision and Control (CDC), IEEE, 2017
J. Umlauft; A. Lederer; S. Hirche: Learning Stable Gaussian Process State Space Models. American Control Conference (ACC), IEEE, 2017
A. Dietrich, M. Kimmel, T. Wimböck, S. Hirche, A. Albu-Schäffer (2014). Workspace Analysis for a Kinematically Coupled Torso of a Torque Controlled Humanoid Robot. 2014 IEEE International Conference on Robotics & Automation (ICRA), 3439–3445. [PDF] [BibTeX] [mediaTUM]
V. Koropouli, A. Gusrialdi, S. Hirche, D. Lee (2016). An extremum-seeking control approach for constrained robotic motion tasks. In Control Engineering Practice, vol. 52 (52), pp. 1–14. [PDF] [BibTeX] [mediaTUM]
M. Lang, O. Dunkley, S. Hirche (2014). Gaussian process kernels for rotations and 6D rigid body motions. In 2014 IEEE International Conference on Robotics and Automation (ICRA) (pp. 5165–5170). IEEE. [PDF] [BibTeX] [mediaTUM]
M. Lang, S. Endo, O. Dunkley, S. Hirche (2014). Object Handover Prediction using Gaussian Processes clustered with Trajectory Classification. In RO-MAN Workshop on Wearable Technology and Human – Wearable Robot Interaction. [PDF] [BibTeX] [mediaTUM]
M. Lang, M. Kleinsteuber, O. Dunkley, S. Hirche (2015). Gaussian Process Dynamical Models over Dual Quaternions. Proceedings of European Control Conference (ECC). Linz, Austria. [PDF] [BibTeX] [mediaTUM]
T. Lorenz, B. Vlaskamp, A. Kasparbauer, A. Mörtl, S. Hirche (2014). Dyadic movement synchronization while performing incongruent trajectories requires mutual adaptation, in Frontiers in Human Neuroscience, vol. 461 (8). [PDF] [BibTeX] [mediaTUM]
J.R. Medina Hernández, T. Lorenz, S. Hirche (2015). Synthesizing Anticipatory Haptic Assistance Considering Human Behavior Uncertainty, in IEEE Transactions on Robotics, vol. 31 (1), 2015, pp. 180 - 190 [PDF] [BibTeX] [mediaTUM]