Vincent Berenz

Coaching robot behavior using continuous physiological affective feedback

In this work we present a new way for human-robot interaction, where a robot is able to receive physiological affective feedback for its actions from a human trainer and learn from it. We capture the human trainers facial expressions using a wearable device that records distal electromyographic signals and uses computational methods of signal processing and pattern recognition in real time. We show how a robot can be coached to perform a certain action when confronted with an object by using the continuous physiological affective feedback from the human trainer. We also show that the robot is able to quickly learn the appropriate actions for different situations from the trainer in a manner modeled after the way children learn from their parents encouragement or reproach. This work shows an effective way to coach a robot using affective feedback and has the advantage of working in multiple lighting conditions and camera angles as well as not increasing the cognitive load of the trainer. Our method has applications in the area of social robotics because it shows that interaction between humans and robots is possible using continuous non-verbal social cues, which are characteristic for human-human interaction.

Emotionally Assisted Human–Robot Interaction Using a Wearable Device for Reading Facial Expressions

Abstract In this paper, we introduce a novel paradigm of emotionally assisted interaction between humans and robots. We present a personal wearable device that can be worn on the side of the face to unobtrusively and continuously detect physiological signals that are a mixture of facial electromyographic signals. Through real-time pattern classification, facial expressions can be identified from them and interpreted as positive and negative responses from a human. We report on successful facial expression identification using Independent Component Analysis and an Artificial Neural Network, and show the design of the interface device that can be used for coaching a real robot.

TDM: A software framework for elegant and rapid development of autonomous behaviors for humanoid robots

Through the use of module based software solutions, programming humanoid robots became simple in the sense that detailed knowledge of the underlying software and hardware layers became largely unnecessary. In this paper we argue that the current situation, while being satisfactory for most users, requires improvement for facing situations in which delivery of a complex autonomous behavior is part of the final target. In such case, implementing a dedicated behavior control architecture remains a complex task. In this paper we propose a behavior oriented software framework to be added above the existing modular architecture. This framework is based on centralized integration of sensory data, schematic representation of objects, resource management and intrinsic motivation. It supports code organization, favors code reuse and allows rapid obtention of behaviors that can be easiliy modified or extended. A version of the framework for Aldebaran Nao was developed and tested.

Targets-Drives-Means: A declarative approach to dynamic behavior specification with higher usability

Small humanoid robots are becoming more affordable and are now used in fields such as human-robot interaction, ethics, psychology, or education. For non-roboticists, the standard paradigm for robot visual programming is based on the selection of behavioral blocks, followed by their connection using communication links. These programs provide efficient user support during the development of complex series of movements and sequential behaviors. However, implementing dynamic control remains challenging because the data flow between components to enforce control loops, object permanence, the memories of object positions, odometry, and finite state machines has to be organized by the users. In this study, we develop a new programming paradigm, Targets-Drives-Means, which is suitable for the specification of dynamic robotic tasks. In this proposed approach, programming is based on the declarative association of reusable dynamic components. A central memory organizes the information flows automatically and issues related to dynamic control are solved by processes that remain hidden from the end users. The proposed approach has advantages during the implementation of dynamic behaviors, but it requires that users stop conceiving robotic tasks as the execution of a sequence of actions. Instead, users are required to organize their programs as collections of behaviors that run in parallel and compete for activation. This might be considered non-intuitive but we also report the positive outcomes of a usability experiment, which evaluated the accessibility of the proposed approach.

Autonomous battery management for mobile robots based on risk and gain assessment

Battery management of mobile robots is an issue that has not been a strong focus of attention and is usually addressed by the simple use of battery thresholds. One of the main causes is that no significant method of assessment of risk of battery depletion has yet been proposed. As a result decision of redirection to a charging station is fixed and takes into account neither a dynamic evaluation of the risk of battery depletion nor an evaluation of the gain, defined as the level of mission accomplishment that could be achieved. In this paper we propose a novel method for evaluation of risk of battery depletion for mobile robots. Uncertainties concerning effective battery capacity, current discharge rate and energy required for reaching the station are addressed by the use of probability density functions. This risk assessment will allow replacing the usage of battery threshold by a customizable risk-taking parameter that will be used to define what level of gain is required for balancing a given level of risk. This risk/gain management of battery will guarantee that decision of redirection to the station corresponds to a favorable compromise between risk and level of mission accomplishment. While the proposed approach has been tested using a simulated and real room cleaning robot, it could be applied on a wider range of mobile robots.

The functional role of automatic body response in shaping voluntary actions based on muscle synergy theory

The functional role of automatic body response in forming voluntary actions remain controversial. We here support the hypothesis that the automatic body responses could be used as a reference to adapt voluntary actions to the environment. We validate this hypothesis by analyzing human body movements from the perspective of muscle synergy. In this study, a horizontal shoulder adduction of the dominant arm of four healthy subjects was examined in various tasks. The tasks include reflex and voluntary movements in regular and modified environments. Preliminary results were encouraging; the number and the consistency between the utilized synergies in automatic and voluntary tasks were fairly correlated. In contrast, there was a lack of the correlation when the environment was abruptly modified (an additional resistance applied to the voluntary movement). This lack of correlation, however, was gradually adjusted through training. Our results suggest that automatic synergy may encode some features which could be used by the central nervous system to shape the voluntary synergy.

Coaching robots with biosignals based on human affective social behaviors

We introduce a novel paradigm of social interaction between humans and robots, which is a style of coaching humanoid robots through interaction with a human instructor, who provides reinforcement via affective/social behaviors and biological signals. In particular facial Electromyography (EMG) to capture affective human response by using a personal wearable device is used as guidance or feedback to shape robot behavior. Through real-time pattern classification, facial expressions can be identified from them and interpreted as positive and negative responses from a human. We also developed a behavior-based architecture for testing this approach in the context of complex reactive robot behaviors.

Usability benchmarks of the Targets-Drives-Means robotic architecture

Even for robots delivered with a middleware, end-user implementation of reaction and deliberation remains a difficult task. We proposed Targets Drives Means (TDM), a behavior based architecture which level of abstraction allows behavior specification through declarative association of reusable components. If behavior based robotic has been broadly used in specialized architectures, TDM implemented it for end-users tools for behavior specification. In this paper TDM is evaluated as a solution for end-user robot programming and results of comparative usability tests are presented.

Risk and gain battery management for self-docking mobile robots

Battery management of mobile robots has not been a strong focus of attention and is usually addressed by the simple use of battery thresholds. As a result decision of redirection to a charging station is fixed and takes into account neither a dynamic evaluation of the risk of battery depletion nor an evaluation of the gain, defined as the level of mission accomplishment that could be achieved. To address this issue, we propose a method of assessment of risk of battery depletion which achieves robustness toward unexpected events through the use of probability density functions. This risk assessment allows replacing the usage of battery threshold by a customizable risk-taking parameter, guaranteeing that redirection to the station corresponds to a favorable compromise between risk and level of mission accomplishment. In this paper, we evaluate this approach for self-docking mobile robots, in particular for situations characterized by discrete gain functions, such as mail delivery and garbage collector robots.

Tacit Learning for Emergence of Task-Related Behaviour through Signal Accumulation

Control of robotic joints movements requires the generation of appropriate torque and force patterns, coordinating the kinematically and dynamically complex multijoints systems. Control theory coupled with inverse and forward internal models are commonly used to map a desired endpoint trajectory into suitable force patterns. In this paper, we propose the use of tacit learning to successfully achieve similar tasks without using any kinematic model of the robotic system to be controlled. Our objective is to design a new control strategy that can achieve levels of adaptability similar to those observed in living organisms and be plausible from a neural control viewpoint. If the neural mechanisms used for mapping goals expressed in the task-space into control-space related command without using internal models remain largely unknown, many neural systems rely on data accumulation. The presented controller does not use any internal model and incorporates knowledge expressed in the task space using only the accumulation of data. Tested on a simulated two-link robot system, the controller showed flexibility by developing and updating its parameters through learning. This controller reduces the gap between reflexive motion based on simple accumulation of data and execution of voluntarily planned actions in a simple manner that does not require complex analysis of the dynamics of the system.