Human-robot interaction

Serena Ivaldi, a research director at Inria at the Laboratoire lorrain de recherche en informatique et ses applications (LORIA – CNRS/Université de Lorraine), presents the AS3 structured action ‘Decision-making, learning and social interaction’ within the Organic Robotics (O2R) research programme. She explores how robots learn through interactions with humans.

You describe laying the foundations for new, society-centred algorithms for decision-making, learning and interaction as a major scientific challenge. How do you technically enable robots to learn and make decisions?

Serena Ivaldi: Robots learn and make decisions using a combination of sensors, algorithms and computing power.

The selection of an action, as well as how to execute it, generally results from a process of estimating the robot’s state within its environment. This process transforms data from multiple sensors — such as cameras, lidars, microphones or force sensors — into information that can be used for decision-making, for example to detect the presence of people nearby, interpret verbal instructions or estimate the force exerted by a person. 

Based on this estimate, planning algorithms determine the actions to be carried out according to the robot’s current state and its objectives. Other algorithms then translate these decisions into concrete actions: every millisecond, real-time calculations generate the commands needed to activate the robot’s motors.

Learning takes place at various levels of granularity and temporal scale, in order to refine movements, improve internal models and adapt decisions to new interaction scenarios. Within the AS3 project, we are studying, in particular, supervised learning, reinforcement learning and diversity optimisation algorithms (see quality diversity).

You mention that you are focusing initially on mobile robotic manipulators to develop these new models and algorithms. Why did you choose this type of robot in the first place? Could you give us two examples of how they physically assist humans?

S.I.: We have selected platforms that are representative of service and assistive robotics. Prototypes fitted with numerous sensors are being used to develop our algorithms, with the ultimate aim of deploying and evaluating them on the PI1, PI2 and PI3 robots.

Mobile manipulator robots, widely used in service robotics, possess the essential capabilities to assist humans in both public and private environments: navigation, perception, manipulation, as well as verbal and non-verbal interaction. They provide a suitable framework for studying realistic scenarios, such as catering services or the delivery of medical equipment in hospital settings. In particular, the robot must be capable of managing the forces exchanged when passing an object, reacting to an unexpected collision, or adapting when a human comes into physical contact with it.

In this first phase of the project, social interaction capabilities take precedence over physical interaction capabilities. This priority will be reversed in the second phase, which will focus on developing algorithms for exoskeletons.

You state that you are seeking to understand whether a robot can find its place in society without causing mistrust and anxiety, whether it is highly intelligent or user-friendly for the general public and equipped with language tailored to end-users in France. What initial insights have non-specialists in robotics (from the fields of neuroscience, neurocybernetics, anthropology, sociology and linguistics) contributed to the project so far?

S.I.: Firstly, they shed crucial light on the social acceptability of robots and on how people interact with service robots, based on extensive field observations.

Secondly, the entire AS3 team is working on the issue of trust in robots. This does not depend solely on the robot’s level of intelligence, but above all on its ability to fit into existing social practices, whilst respecting cultural roles and norms. Social theory highlights the difficulty of transposing the concept of interpersonal trust – specific to human relationships – to interactions between humans and robots. Psychology emphasises the need for the robot to detect breaches of trust and adapt its behaviour accordingly.

Neuroscience has also contributed to this research, demonstrating that trust is closely linked to the anxiety generated by a robot’s movements when it is in close proximity to a human.

Overall, these disciplines point to the directions in which our decision-making, learning and interaction algorithms must evolve in order to design robots that are better suited to social contexts.