CASH only: Constitutive autonomy through motorsensory self-programming

Goal Oriented Behavior With a Habit-Based Adaptive Sensorimotor Map Network

We present a description of an ASM-network, a new habit-based robot controller model consisting of a network of adaptive sensorimotor maps. This model draws upon recent theoretical developments in enactive cognition concerning habit and agency at the sensorimotor level. It aims to provide a platform for experimental investigation into the relationship between networked organizations of habits and cognitive behavior. It does this by combining (1) a basic mechanism of generating continuous motor activity as a function of historical sensorimotor trajectories with (2) an evaluative mechanism which reinforces or weakens those historical trajectories as a function of their support of a higher-order structure of higher-order sensorimotor coordinations. After describing the model, we then present the results of applying this model in the context of a well-known minimal cognition task involving object discrimination. In our version of this experiment, an individual robot is able to learn the task through a combination of exploration through random movements and repetition of historic trajectories which support the structure of a pre-given network of sensorimotor coordinations. The experimental results illustrate how, utilizing enactive principles, a robot can display recognizable learning behavior without explicit representational mechanisms or extraneous fitness variables. Instead, our model's behavior adapts according to the internal requirements of the action-generating mechanism itself.

Quantifying Reinforcement-Learning Agent’s Autonomy, Reliance on Memory and Internalisation of the Environment

Intuitively, the level of autonomy of an agent is related to the degree to which the agent’s goals and behaviour are decoupled from the immediate control by the environment. Here, we capitalise on a recent information-theoretic formulation of autonomy and introduce an algorithm for calculating autonomy in a limiting process of time step approaching infinity. We tackle the question of how the autonomy level of an agent changes during training. In particular, in this work, we use the partial information decomposition (PID) framework to monitor the levels of autonomy and environment internalisation of reinforcement-learning (RL) agents. We performed experiments on two environments: a grid world, in which the agent has to collect food, and a repeating-pattern environment, in which the agent has to learn to imitate a sequence of actions by memorising the sequence. PID also allows us to answer how much the agent relies on its internal memory (versus how much it relies on the observations) when transitioning to its next internal state. The experiments show that specific terms of PID strongly correlate with the obtained reward and with the agent’s behaviour against perturbations in the observations.

Artificial Interactionism: Avoiding Isolating Perception From Cognition in AI

We discuss the influence upon the fields of robotics and AI of the manner one conceives the relationships between artificial agents' perception, cognition, and action. We shed some light upon a widespread paradigm we call the isolated perception paradigm that addresses perception as isolated from cognition and action. By mobilizing the resources of philosophy (phenomenology and epistemology) and cognitive sciences, and by drawing on recent approaches in AI, we explore what it could mean for robotics and AI to take distance from the isolated perception paradigm. We argue that such a renouncement opens interesting ways to explore the possibilities for designing artificial agents with intrinsic motivations and constitutive autonomy. We then propose Artificial Interactionism, our approach that escapes the isolated perception paradigm by drawing on the inversion of the interaction cycle. When the interaction cycle is inverted, input data are not percepts directly received from the environment, but outcomes of control loops. Perception is not received from sensors in isolation from cognition but is actively constructed by the cognitive architecture through interaction. We give an example implementation of artificial interactionism that demonstrates basic intrinsically motivated learning behavior in a dynamic simulated environment.