Learning to Co-Drive. Brain Architectures and Mental Imagery Mechanisms that Help Improving Agents for Autonomous Driving and Enable Natural Human-Robot Interactions

Details

14:30 - 15:00 | Sun 9 Jun | Room L218 | SuBT2.7

Session: BROAD: Algorithmic, Legal, and Societal Challenges for Autonomous Driving

Abstract

This talk describes the position of the H2020 Dreams4Cars research project (a Research and Innovation Action funded under the EU Robotics banner) that deals with the architecture, and the abilities, of agents that should be capable of autonomously learning reliable driving and natural human-robot interactions. Hence, the goals of D4C are developing cognition abilities for 1) automatic discovery of significant situations and 2) automatic learning from those situations. I will start from an introductory survey of theories for artificial cognition in embodied robots. The traditional sense-think-act paradigm will be recalled and critically reviewed, focusing on the issues that limit this design approach. Some considerations regarding recent examples of Deep Neural Network implementations (end-to-end trained networks) will also be given. Then, I will introduce and motivate a biologically inspired layered control architecture that consists of a network-of-networks. I will show how this agent architecture can be created with use (and re-use) of simple artificial neural networks building components that can be tested in isolation and hence the aggregated behaviour can be certified. I will show examples of emerging safe behaviours that are difficult to obtain with the sense-think-act approach. I will then focus on the lower levels of sensorimotor control, taking inspiration from how the human brain efficiently solves the problem of learning the forward and inverse dynamics of its body and of the manipulated objects and how these learned models are used for a variety of in-line and offline purposes. I will provide examples of application of these principles to the engineering of artificial agents; in particular examples of vehicle dynamics models learned with artificial neural networks (contrasted to parametric analytical models) and their use for sensory anticipation, state estimate and motor control. I will then talk about sensorimotor imagery. Predictions in the brain and neural network architectures for similar efficient prediction in artificial agents. Episodic simulations and embodied simulation. Inline use of sensory anticipation. Forward/inverse model adaptation. Learning efficient motor control at chassis and tactical level. Offline use of sensorimotor imagery. Creating episodes and learning from episodes. Finally, I will talk about the use of the same agent architecture for creating emergent human-robot interactions (in the “like-me” fashion). I will recall the mirror neuron theory and how mirroring emerges from the agent as just another type of action-selection process. How an agent with mirroring can understand and predict human intentions and how it can collaborate with humans.