An Embodied Agent Learning Affordances With Intrinsic Motivations and Solving Extrinsic Tasks With Attention and One-Step Planning

Interdependence as the key for an ethical artificial autonomy

Currently, the autonomy of artificial systems, robotic systems in particular, is certainly one of the most debated issues, both from the perspective of technological development and its social impact and ethical repercussions. While theoretical considerations often focus on scenarios far beyond what can be concretely hypothesized from the current state of the art, the term autonomy is still used in a vague or too general way. This reduces the possibilities of a punctual analysis of such an important issue, thus leading to often polarized positions (naive optimism or unfounded defeatism). The intent of this paper is to clarify what is meant by artificial autonomy, and what are the prerequisites that can allow the attribution of this characteristic to a robotic system. Starting from some concrete examples, we will try to indicate a way towards artificial autonomy that can hold together the advantages of developing adaptive and versatile systems with the management of the inevitable problems that this technology poses both from the viewpoint of safety and ethics. Our proposal is that a real artificial autonomy, especially if expressed in the social context, can only be achieved through interdependence with other social actors (human and otherwise), through continuous exchanges and interactions which, while allowing robots to explore the environment, guarantee the emergence of shared practices, behaviors, and ethical principles, which otherwise could not be imposed with a top-down approach, if not at the price of giving up the same artificial autonomy.

A Biological Inspired Cognitive Framework for Memory-Based Multi-Sensory Joint Attention in Human-Robot Interactive Tasks

One of the fundamental prerequisites for effective collaborations between interactive partners is the mutual sharing of the attentional focus on the same perceptual events. This is referred to as joint attention. In psychological, cognitive, and social sciences, its defining elements have been widely pinpointed. Also the field of human-robot interaction has extensively exploited joint attention which has been identified as a fundamental prerequisite for proficient human-robot collaborations. However, joint attention between robots and human partners is often encoded in prefixed robot behaviours that do not fully address the dynamics of interactive scenarios. We provide autonomous attentional behaviour for robotics based on a multi-sensory perception that robustly relocates the focus of attention on the same targets the human partner attends. Further, we investigated how such joint attention between a human and a robot partner improved with a new biologically-inspired memory-based attention component. We assessed the model with the humanoid robot iCub involved in performing a joint task with a human partner in a real-world unstructured scenario. The model showed a robust performance on capturing the stimulation, making a localisation decision in the right time frame, and then executing the right action. We then compared the attention performance of the robot against the human performance when stimulated from the same source across different modalities (audio-visual and audio only). The comparison showed that the model is behaving with temporal dynamics compatible with those of humans. This provides an effective solution for memory-based joint attention in real-world unstructured environments. Further, we analyzed the localisation performances (reaction time and accuracy), the results showed that the robot performed better in an audio-visual condition than an audio only condition. The performance of the robot in the audio-visual condition was relatively comparable with the behaviour of the human participants whereas it was less efficient in audio-only localisation. After a detailed analysis of the internal components of the architecture, we conclude that the differences in performance are due to egonoise which significantly affects the audio-only localisation performance.

A computational model of language functions in flexible goal-directed behaviour

The function of language in high-order goal-directed human cognition is an important topic at the centre of current debates. Experimental evidence shows that inner speech, representing a self-directed form of language, empowers cognitive processes such as working memory, perception, categorization, and executive functions. Here we study the relations between inner speech and processes like feedback processing and cognitive flexibility. To this aim we propose a computational model that controls an artificial agent who uses inner speech to internally manipulate its representations. The agent is able to reproduce human behavioural data collected during the solution of the Wisconsin Card Sorting test, a neuropsychological test measuring cognitive flexibility, both in the basic condition and when a verbal shadowing protocol is used. The components of the model were systematically lesioned to clarify the specific impact of inner speech on the agent’s behaviour. The results indicate that inner speech improves the efficiency of internal representation manipulation. Specifically, it makes the representations linked to specific visual features more disentangled, thus improving the agent’s capacity to engage/disengage attention on stimulus features after positive/negative action outcomes. Overall, the model shows how inner speech could improve goal-directed internal manipulation of representations and enhance behavioural flexibility.

A generative spiking neural-network model of goal-directed behaviour and one-step planning

In mammals, goal-directed and planning processes support flexible behaviour used to face new situations that cannot be tackled through more efficient but rigid habitual behaviours. Within the Bayesian modelling approach of brain and behaviour, models have been proposed to perform planning as probabilistic inference but this approach encounters a crucial problem: explaining how such inference might be implemented in brain spiking networks. Recently, the literature has proposed some models that face this problem through recurrent spiking neural networks able to internally simulate state trajectories, the core function at the basis of planning. However, the proposed models have relevant limitations that make them biologically implausible, namely their world model is trained ‘off-line’ before solving the target tasks, and they are trained with supervised learning procedures that are biologically and ecologically not plausible. Here we propose two novel hypotheses on how brain might overcome these problems, and operationalise them in a novel architecture pivoting on a spiking recurrent neural network. The first hypothesis allows the architecture to learn the world model in parallel with its use for planning: to this purpose, a new arbitration mechanism decides when to explore, for learning the world model, or when to exploit it, for planning, based on the entropy of the world model itself. The second hypothesis allows the architecture to use an unsupervised learning process to learn the world model by observing the effects of actions. The architecture is validated by reproducing and accounting for the learning profiles and reaction times of human participants learning to solve a visuomotor learning task that is new for them. Overall, the architecture represents the first instance of a model bridging probabilistic planning and spiking-processes that has a degree of autonomy analogous to the one of real organisms.

Goal-directed behaviour relies on brain processes supporting planning of actions based on their expected consequences before performing them in the environment. An important computational modelling approach proposes that the brain performs goal-directed processes on the basis of probability distributions and computations on them. A key challenge of this approach is to explain how these probabilistic processes can rely on the spiking processes of the brain. The literature has recently proposed some models that do so by ‘thinking ahead’ alternative possible action-outcomes based on low-level neuronal stochastic events. However, these models have a limited autonomy as they require to learn how the environment works (‘world model’) before solving the tasks, and use a biologically implausible learning process requiring an ‘external teacher’ to tell how their internal units should respond. Here we present a novel architecture proposing how organisms might overcome these challenging problems. First, the architecture can decide if exploring, to learn the world model, or planning, using such model, by evaluating how confident it is on the model knowledge. Second, the architecture can autonomously learn the world model based on experience. The architecture represents a first fully autonomous planning model relying on a spiking neural network.