Mathis Richter

A robotic architecture for action selection and behavioral organization inspired by human cognition

Robotic agents that interact with humans and perform complex, everyday tasks in natural environments will require a system to autonomously organize their behavior. Current systems for robotic behavioral organization typically abstract from the low-level sensory-motor embodiment of the robot, leading to a gap between the level at which a sequence of actions is planned and the levels of perception and motor control. This gap is a major bottleneck for the autonomy of systems in complex, dynamic environments. To address this issue, we present a neural-dynamic framework for behavioral organization, in which the action selection mechanism is tightly coupled to the agents sensory-motor systems. The elementary behaviors (EBs) of the robot are dynamically organized into sequences based on task-specific behavioral constraints and online perceptual information. We demonstrate the viability of our approach by implementing a neural-dynamic architecture on the humanoid robot NAO. The system is capable of producing sequences of EBs that are directed at objects (e.g., grasping and pointing). The sequences are flexible in that the robot autonomously adapts the individual EBs and their sequential order in response to changes in the sensed environment. The architecture can accommodate different tasks and can be articulated for different robotic platforms. Its neural-dynamic substrate is particularly well-suited for learning and adaptation.

A neural-dynamic architecture for behavioral organization of an embodied agent

How agents generate meaningful sequences of actions in natural environments is one of the most challenging problems in studies of natural cognition and in the design of artificial cognitive systems. Each action in a sequence must contribute to the behavioral objective, while at the same time satisfying constraints that arise from the environment, the agents embodiment, and the agents behavioral history. In this paper, we introduce a neural-dynamic architecture that enables selection of an appropriate action for a given task in a particular environment and is open to learning. We use the same framework of neural dynamics for all processes from perception, to representation and motor planning as well as behavioral organization. This facilitates integration and flexibility. The neural dynamic representations of particular behaviors emerge on the fly from the interplay between task and environment inputs as well as behavioral history. All behavioral states are attractors of the neural dynamics, whose instabilities lead to behavioral switches. As a result, behavioral organization is robust in the face of noisy and unreliable sensory information.

A Software Framework for Cognition, Embodiment, Dynamics, and Autonomy in Robotics: Cedar

We present Cedar, a software framework for the implementation and simulation of embodied cognitive models based on Dynamic Field Theory (DFT). DFT is a neurally inspired theoretical framework that integrates perception, action, and cognition. Cedar captures the power of DFT in software by facilitating the process of software development for embodied cognitive systems, both artificial and as models of human cognition. In Cedar, models can be designed through a graphical interface and interactively tuned. We demonstrate this by implementing an exemplary robotic architecture.

Parsing of action sequences: A neural dynamics approach

Abstract Parsing of action sequences is the process of segmenting observed behavior into individual actions. In robotics, this process is critical for imitation learning from observation and for representing an observed behavior in a form that may be communicated to a human. In this paper, we develop a model for action parsing, based on our understanding of principles of grounded cognitive processes, such as perceptual decision making, behavioral organization, and memory formation.We present a neural-dynamic architecture, in which action sequences are parsed using a mathematical and conceptual framework for embodied cognition—the Dynamic Field Theory. In this framework, we introduce a novel mechanism, which allows us to detect and memorize actions that are extended in time and are parametrized by the target object of an action. The core properties of the architecture are demonstrated in a set of simple, proof-of-concept experiments.

A Neural Dynamic Architecture Resolves Phrases about Spatial Relations in Visual Scenes

How spatial language, important to both cognitive science and robotics, is mapped to real-world scenes by neural processes is not understood. We present an autonomous neural dynamics that achieves this mapping flexibly. Neural activation fields represent and spatially transform perceptual information. An architecture of dynamic nodes interacts with these perceptual fields to instantiate categorical concepts. Discrete time processing steps emerge from instabilities of the time-continuous neural dynamics and are organized sequentially by these nodes. These steps include the attentional selection of individual objects in a scene, mapping locations to an object-centered reference frame, and evaluating matches to relational spatial terms. The architecture can respond to queries specified by setting the state of discrete nodes. It autonomously generates a response based on visual input about a scene.

Autonomous Robot Hitting Task Using Dynamical System Approach

We propose a model that autonomously generates and flexibly organizes sequences of timed actions. The timing of the movements is controlled by non-linear oscillators. Their activation and deactivation is organized by a hierarchical neural-dynamic architecture. We demonstrate the features of our model in an exemplary robotic task where the manipulator arm keeps hitting a ball up an inclined plane. The autonomous generation of movement sequences is tightly coupled to visual sensory information about the ball motion and able to adapt, on-line, to perturbations introduced in the ball trajectory. The performance of the proposed model is evaluated and the reactions to different perturbations are discussed.

Developing Dynamic Field Theory Architectures for Embodied Cognitive Systems with cedar

Embodied artificial cognitive systems, such as autonomous robots or intelligent observers, connect cognitive processes to sensory and effector systems in real time. Prime candidates for such embodied intelligence are neurally inspired architectures. While components such as forward neural networks are well established, designing pervasively autonomous neural architectures remains a challenge. This includes the problem of tuning the parameters of such architectures so that they deliver specified functionality under variable environmental conditions and retain these functions as the architectures are expanded. The scaling and autonomy problems are solved, in part, by dynamic field theory (DFT), a theoretical framework for the neural grounding of sensorimotor and cognitive processes. In this paper, we address how to efficiently build DFT architectures that control embodied agents and how to tune their parameters so that the desired cognitive functions emerge while such agents are situated in real environments. In DFT architectures, dynamic neural fields or nodes are assigned dynamic regimes, that is, attractor states and their instabilities, from which cognitive function emerges. Tuning thus amounts to determining values of the dynamic parameters for which the components of a DFT architecture are in the specified dynamic regime under the appropriate environmental conditions. The process of tuning is facilitated by the software framework cedar, which provides a graphical interface to build and execute DFT architectures. It enables to change dynamic parameters online and visualize the activation states of any component while the agent is receiving sensory inputs in real time. Using a simple example, we take the reader through the workflow of conceiving of DFT architectures, implementing them on embodied agents, tuning their parameters, and assessing performance while the system is coupled to real sensory inputs.

A Neural Dynamic Model Generates Descriptions of Object‐Oriented Actions

Describing actions entails that relations between objects are discovered. A pervasively neural account of this process requires that fundamental problems are solved: the neural pointer problem, the binding problem, and the problem of generating discrete processing steps from time-continuous neural processes. We present a prototypical solution to these problems in a neural dynamic model that comprises dynamic neural fields holding representations close to sensorimotor surfaces as well as dynamic neural nodes holding discrete, language-like representations. Making the connection between these two types of representations enables the model to describe actions as well as to perceptually ground movement phrases-all based on real visual input. We demonstrate how the dynamic neural processes autonomously generate the processing steps required to describe or ground object-oriented actions. By solving the fundamental problems of neural pointing, binding, and emergent discrete processing, the model may be a first but critical step toward a systematic neural processing account of higher cognition.

Learning the Condition of Satisfaction of an Elementary Behavior in Dynamic Field Theory

Abstract In order to proceed along an action sequence, an autonomous agent has to recognize that the intended final condition of the previous action has been achieved. In previous work, we have shown how a sequence of actions can be generated by an embodied agent using a neural-dynamic architecture for behavioral organization, in which each action has an intention and condition of satisfaction. These components are represented by dynamic neural fields, and are coupled to motors and sensors of the robotic agent.Here,we demonstratehowthemappings between intended actions and their resulting conditions may be learned, rather than pre-wired.We use reward-gated associative learning, in which, over many instances of externally validated goal achievement, the conditions that are expected to result with goal achievement are learned. After learning, the external reward is not needed to recognize that the expected outcome has been achieved. This method was implemented, using dynamic neural fields, and tested on a real-world E-Puck mobile robot and a simulated NAO humanoid robot.

A neural dynamics to organize timed movement: Demonstration in a robot ball bouncing task

To address how different movement behaviors may be timed to sensory events while being flexibly organized in sequence, we propose a neural dynamic model of timed movement organization. Two layers of neural dynamics control the activation and de-activation of different elementary movements, while a third layer uses stable limit cycle oscillators to generate timed movement trajectories. Both the organization and the generation of the timed movements are coupled to online sensory information so that the system can compensate for perturbations by updating the movement trajectory while recovering the required movement timing. We formulate and demonstrate the approach in a robotic ball bouncing task. When the ball begins to fall, the robot arm moves to the interception point in a plane and initiates a rotatory motion of the racket timed such as to hit the ball with maximal velocity. When the ball is no longer falling or falling outside the reachable space, the robot moves the racket back toward baseline. A physics simulation is used to assess the model and demonstrate its capacity to handle perturbations of the ball trajectory.