Andrej Gams

Task-Specific Generalization of Discrete and Periodic Dynamic Movement Primitives

Acquisition of new sensorimotor knowledge by imitation is a promising paradigm for robot learning. To be effective, action learning should not be limited to direct replication of movements obtained during training but must also enable the generation of actions in situations a robot has never encountered before. This paper describes a methodology that enables the generalization of the available sensorimotor knowledge. New actions are synthesized by the application of statistical methods, where the goal and other characteristics of an action are utilized as queries to create a suitable control policy, taking into account the current state of the world. Nonlinear dynamic systems are employed as a motor representation. The proposed approach enables the generation of a wide range of policies without requiring an expert to modify the underlying representations to account for different task-specific features and perceptual feedback. The paper also demonstrates that the proposed methodology can be integrated with an active vision system of a humanoid robot. 3-D vision data are used to provide query points for statistical generalization. While 3-D vision on humanoid robots with complex oculomotor systems is often difficult due to the modeling uncertainties, we show that these uncertainties can be accounted for by the proposed approach.

On-line learning and modulation of periodic movements with nonlinear dynamical systems

The paper presents a two-layered system for (1) learning and encoding a periodic signal without any knowledge on its frequency and waveform, and (2) modulating the learned periodic trajectory in response to external events. The system is used to learn periodic tasks on a humanoid HOAP-2 robot. The first layer of the system is a dynamical system responsible for extracting the fundamental frequency of the input signal, based on adaptive frequency oscillators. The second layer is a dynamical system responsible for learning of the waveform based on a built-in learning algorithm. By combining the two dynamical systems into one system we can rapidly teach new trajectories to robots without any knowledge of the frequency of the demonstration signal. The system extracts and learns only one period of the demonstration signal. Furthermore, the trajectories are robust to perturbations and can be modulated to cope with a dynamic environment. The system is computationally inexpensive, works on-line for any periodic signal, requires no additional signal processing to determine the frequency of the input signal and can be applied in parallel to multiple dimensions. Additionally, it can adapt to changes in frequency and shape, e.g. to non-stationary signals, such as hand-generated signals and human demonstrations.

Coupling Movement Primitives: Interaction With the Environment and Bimanual Tasks

The framework of dynamic movement primitives (DMPs) contains many favorable properties for the execution of robotic trajectories, such as indirect dependence on time, response to perturbations, and the ability to easily modulate the given trajectories, but the framework in its original form remains constrained to the kinematic aspect of the movement. In this paper, we bridge the gap to dynamic behavior by extending the framework with force/torque feedback. We propose and evaluate a modulation approach that allows interaction with objects and the environment. Through the proposed coupling of originally independent robotic trajectories, the approach also enables the execution of bimanual and tightly coupled cooperative tasks. We apply an iterative learning control algorithm to learn a coupling term, which is applied to the original trajectory in a feed-forward fashion and, thus, modifies the trajectory in accordance to the desired positions or external forces. A stability analysis and results of simulated and real-world experiments using two KUKA LWR arms for bimanual tasks and interaction with the environment are presented. By expanding on the framework of DMPs, we keep all the favorable properties, which is demonstrated with temporal modulation and in a two-agent obstacle avoidance task.

On-line motion synthesis and adaptation using a trajectory database

Autonomous robots cannot be programmed in advance for all possible situations. Instead, they should be able to generalize the previously acquired knowledge to operate in new situations as they arise. A possible solution to the problem of generalization is to apply statistical methods that can generate useful robot responses in situations for which the robot has not been specifically instructed how to respond. In this paper we propose a methodology for the statistical generalization of the available sensorimotor knowledge in real-time. Example trajectories are generalized by applying Gaussian process regression, using the parameters describing a task as query points into the trajectory database. We show on real-world tasks that the proposed methodology can be integrated into a sensory feedback loop, where the generalization algorithm is applied in real-time to adapt robot motion to the perceived changes of the external world.

On-line frequency adaptation and movement imitation for rhythmic robotic tasks

In this paper we present a novel method to obtain the basic frequency of an unknown periodic signal with an arbitrary waveform, which can work online with no additional signal processing or logical operations. The method originates from non-linear dynamical systems for frequency extraction, which are based on adaptive frequency oscillators in a feedback loop. In previous work, we had developed a method that could extract separate frequency components by using several adaptive frequency oscillators in a loop, but that method required a logical algorithm to identify the basic frequency. The novel method presented here uses a Fourier series representation in the feedback loop combined with a single oscillator. In this way it can extract the frequency and the phase of an unknown periodic signal in real time and without any additional signal processing or preprocessing. The method determines the Fourier series coefficients and can be used for dynamic Fourier series implementation. The proposed method can be used for the control of rhythmic robotic tasks, where only the extraction of the basic frequency is crucial. For demonstration several highly non-linear and dynamic periodic robotic tasks are shown, including also a task where an electromyography (EMG) signal is used in a feedback loop.

On-line periodic movement and force-profile learning for adaptation to new surfaces

To control the motion of a humanoid robot along a desired trajectory in contact with a rigid object, we need to take into account forces that arise from contact with the surface of the object. In this paper we propose a new method that enables the robot to adapt its motion to different surfaces. The initial trajectories are encoded by dynamic movement primitives, which can be learned from visual feedback using a two-layered imitation system. In our approach these initial trajectories are modified using regression methods. The data for learning is provided by force feedback. In this way new trajectories are learned that ensure that the robot can move along the object while maintaining contact and applying the desired force to the object. Active compliance can be used more effectively with such trajectories. We present the results for both movement imitation and force profile learning on two different surfaces. We applied the method to the ARMAR-IIIb humanoid robot, where we use the system for learning and imitating a periodic task of wiping a kitchen table.

Effects of Robotic Knee Exoskeleton on Human Energy Expenditure

A number of studies discuss the design and control of various exoskeleton mechanisms, yet relatively few address the effect on the energy expenditure of the user. In this paper, we discuss the effect of a performance augmenting exoskeleton on the metabolic cost of an able-bodied user/pilot during periodic squatting. We investigated whether an exoskeleton device will significantly reduce the metabolic cost and what is the influence of the chosen device control strategy. By measuring oxygen consumption, minute ventilation, heart rate, blood oxygenation, and muscle EMG during 5-min squatting series, at one squat every 2 s, we show the effects of using a prototype robotic knee exoskeleton under three different noninvasive control approaches: gravity compensation approach, position-based approach, and a novel oscillator-based approach. The latter proposes a novel control that ensures synchronization of the device and the user. Statistically significant decrease in physiological responses can be observed when using the robotic knee exoskeleton under gravity compensation and oscillator-based control. On the other hand, the effects of position-based control were not significant in all parameters although all approaches significantly reduced the energy expenditure during squatting.

Generalization of example movements with dynamic systems

In the past, nonlinear dynamic systems have been proposed as a suitable representation for motor control. It has been shown that it is possible to learn desired complex control policies by a nonlinear transformation of an existing simpler control policy, which is based on a canonical dynamic system. The resulting control policies were termed dynamic movement primitives. The main result of this paper is an approach to learning parametrized sets of dynamic movement primitives based on a library of example movements. Learning was implemented by applying locally weighted regression where the goal of an action is used as a query point into the library of example movements. The proposed approach enables the generation of a wide range of movements that are adapted to the current configuration of the external world without requiring an expert to appropriately modify the underlying differential equations to account for percepetual feedback.

Real-time 3D marker tracking with a WIIMOTE stereo vision system: Application to robotic throwing

In this paper we describe the use of a standard game console joystick, namely the Nintendo WIIMOTE, for an active real-time 3D marker tracking. We show the ease of applicability of inexpensive and robust standard game controllers for 3D object tracking, e.g. to track an infrared source in 3D space. Recovering the 3D information using stereo vision is still one of the major research areas in computer vision and has given rise to a great deal of literature in the recent past. In this paper we present the method for calibrating a WIIMOTE stereo pair without knowing any parameters of the build-in infrared cameras in advance. The results are two matrices which includes both the intrinsic and extrinsic parameters for left and right cameras. The comparison between the stereo and the mono WIIMOTE tracking system is presented. Furthermore, to demonstrate the use of the WIIMOTE stereo system we considered the task of throwing a ball with robotic hand, to the target identified with an infrared source. The throwing task was divided into two separate parts: the tracking part and the throwing part.

Imitating human acceleration of a gyroscopic device

Spinning up a Power®ball—a hand-held gyroscopic toy or exerciser that exhibits rotor spin-up when applying appropriate torque to its casing—is a fairly easy task for a human but rather complex to perform with a robot. To accomplish the task of spinning up the rotor of the Power®ball with a robot, we measured the motion of a human and identified the conditions an individual uses for a successful spin-up. Several control approaches were applied to the device mainly using feedback information from the velocity counter and force/torque sensor to synchronize the torque exerted by the device and the motion of the robot. Best human imitation was achieved with two modified learning methods with highest rotor speeds in excess of 1480 rad/s, rating among top 100 world Power®ball players.