Duy Nguyen-Tuong

Safe Exploration for Active Learning with Gaussian Processes

In this paper, the problem of safe exploration in the active learning context is considered. Safe exploration is especially important for data sampling from technical and industrial systems, e.g. combustion engines and gas turbines, where critical and unsafe measurements need to be avoided. The objective is to learn data-based regression models from such technical systems using a limited budget of measured, i.e. labelled, points while ensuring that critical regions of the considered systems are avoided during measurements. We propose an approach for learning such models and exploring new data regions based on Gaussian processes GPs. In particular, we employ a problem specific GP classifier to identify safe and unsafe regions, while using a differential entropy criterion for exploring relevant data regions. A theoretical analysis is shown for the proposed algorithm, where we provide an upper bound for the probability of failure. To demonstrate the efficiency and robustness of our safe exploration scheme in the active learning setting, we test the approach on a policy exploration task for the inverse pendulum hold up problem.

Learning Throttle Valve Control Using Policy Search

The throttle valve is a technical device used for regulating a fluid or a gas flow. Throttle valve control is a challenging task, due to its complex dynamics and demanding constraints for the controller. Using state-of-the-art throttle valve control, such as model-free PID controllers, time-consuming and manual adjusting of the controller is necessary. In this paper, we investigate how reinforcement learning (RL) can help to alleviate the effort of manual controller design by automatically learning a control policy from experiences. In order to obtain a valid control policy for the throttle valve, several constraints need to be addressed, such as no-overshoot. Furthermore, the learned controller must be able to follow given desired trajectories, while moving the valve from any start to any goal position and, thus, multi-targets policy learning needs to be considered for RL. In this study, we employ a policy search RL approach, Pilco [2], to learn a throttle valve control policy. We adapt the Pilco algorithm, while taking into account the practical requirements and constraints for the controller. For evaluation, we employ the resulting algorithm to solve several control tasks in simulation, as well as on a physical throttle valve system. The results show that policy search RL is able to learn a consistent control policy for complex, real-world systems.

Fusing vision and odometry for accurate indoor robot localization

For service robotics, localization is an essential component required in many applications, e.g. indoor robot navigation. Today, accurate localization relies mostly on high-end devices, such as A.R.T. DTrack, VICON systems or laser scanners. These systems are often expensive and, thus, require substantial investments. In this paper, our focus is on the development of a localization method using low-priced devices, such as cameras, while being sufficiently accurate in tracking performance. Vision data contains much information and potentially yields high tracking accuracy. However, due to high computational requirements vision-based localization can only be performed at a low frequency. In order to speed up the visual localization and increase accuracy, we combine vision information with robots odometry using a Kalman-Filter. The resulting approach enables sufficiently accurate tracking performance (errors in the range of few cm) at a frequency of about 35Hz. To evaluate the proposed method, we compare our tracking performance with the high precision A.R.T. DTrack localization as ground truth. The evaluations on real robot show that our low-priced localization approach is competitive for indoor robot localization tasks.

Model-based policy search for automatic tuning of multivariate PID controllers

PID control architectures are widely used in industrial applications. Despite their low number of open parameters, tuning multiple, coupled PID controllers can become tedious in practice. In this paper, we extend PILCO, a model-based policy search framework, to automatically tune multivariate PID controllers purely based on data observed on an otherwise unknown system. The systems state is extended appropriately to frame the PID policy as a static state feedback policy. This renders PID tuning possible as the solution of a finite horizon optimal control problem without further a priori knowledge. The framework is applied to the task of balancing an inverted pendulum on a seven degree-of-freedom robotic arm, thereby demonstrating its capabilities of fast and data-efficient policy learning, even on complex real world problems.

Sparse Gaussian process regression for compliant, real-time robot control

Sparse Gaussian process (GP) models provide an efficient way to perform regression on large data sets. The key idea is to select a representative subset of the available training data, which induces the sparse GP model approximation. In the past, a variety of selection criteria for GP approximation have been proposed, but they either lack accuracy or suffer from high computational costs. In this paper, we introduce a novel and straightforward criterion for successive selection of training points used for GP model approximation. The proposed algorithm allows a fast and efficient selection of training points, while being competitive in learning performance. As evaluation, we employ our approach in learning inverse dynamics models for robot control using very large data sets (e.g. 500.000 samples). It is demonstrated in experiments that our approximated GP model is sufficiently fast for real-time prediction in robot control. Comparisons with other state-of-the-art approximation techniques show that our proposed approach is significantly faster, while being competitive to generalization accuracy.

Efficient sparsification for Gaussian process regression

Abstract Sparse Gaussian process models provide an efficient way to perform regression on large data sets. Sparsification approaches deal with the selection of a representative subset of available training data for inducing the sparse model approximation. A variety of insertion and deletion criteria have been proposed, but they either lack accuracy or suffer from high computational costs. In this paper, we present a new and straightforward criterion for successive selection and deletion of training points in sparse Gaussian process regression. The proposed novel strategies for sparsification are as fast as the purely randomized schemes and, thus, appropriate for applications in online learning. Experiments on real-world robot data demonstrate that our obtained regression models are competitive with the computationally intensive state-of-the-art methods in terms of generalization and accuracy. Furthermore, we employ our approach in learning inverse dynamics models for compliant robot control using very large data sets, i.e. with half a million training points. In this experiment, it is also shown that our approximated sparse Gaussian process model is sufficiently fast for real-time prediction in robot control.