Spring, the makings of realistic social robotics

Changed on 21/12/2021

How do we develop robots capable of non-pre-programmed social interactions with several people at once? This is the question that the European project SPRING will attempt to answer in the field of healthcare.

Interactions avec le robot humanoïde ARI

Robots capable of interacting with humans

What if robots could take on the repetitive tasks involved in receiving the public? Science fiction? For the moment, maybe... We know there are already forms of artificial intelligence which are capable of interacting with humans. One such example is the therapeutic robot seal Paro, which is used in the healthcare field to provide the benefits of zootherapy to patients with cognitive disorders. However these mediation tools have but rudimentary possibilities, or require remote control by an engineer.

“They lack the ability for multi-person interaction. If two people are having a conversation or if they both ask the robot a question, present-day robots are incapable of understanding,” says Xavier Alameda-Pineda, head of the RobotLearn team at the Inria Grenoble Rhône-Alpes centre.

Numerous perceptive skills are required

While there are “butler” robots which can provide the weather forecast or give geographical directions, they are not able to execute complex social tasks autonomously, such as escorting users around a building. To be able to carry out such tasks, a social robot must be capable of perceiving and distinguishing signals emitted by different speakers, understanding these signals and identifying that they are addressed to the robot, and then react accordingly. This is a daunting challenge, because it requires numerous perceptive abilities and a capacity for automatic learning in order to execute autonomous decision-making.

To address this issue, Xavier Alameda-Pineda, a specialist in audio-visual perception, teamed up with seven partners of renowned expertise in complementary fields to his own (see slide).Together, they run SPRING (Socially Pertinent Robots for Gerontological Healthcare), a 4-year, € 8.3-million project funded by the European programme Horizon 2020 which the young researcher is responsible for coordinating.

SPRING in video

A multi-disciplinary project

Four of the partners are from the academic world (Prague Technical University, Czech Republic; the University of Trento, Italy; Heriot-Watt University, Edinburgh, UK and Bar-Ilan University, Israel).

Alongside Robotlearn, the task of these research teams is to develop the dialogue, audio processing, human behaviour analysis and spatial awareness system of ARI, a humanoid robot developed by one of the project’s technology partners, Pal Robotics. ERM Automatismes, the other technology partner, is in charge of integrating the various advances made by the researchers in the robot software. Lastly a hospital, AP-HP (Ile-de-France University Hospital), which already uses the robots Paro and Nao in the Broca Living Lab, will provide SPRING with an experimental framework consisting of patients and medical staff.

The project partners

AP-HP (Ile-de-France University Hospital)

APHP's group works on the ethics of the project, on robot acceptability issues and on the experimental setups at a geriatric day-care hospital. Two important questions can be raised when attempting to put a social robot in a healthcare environment: first (pre-experiment), the experimental protocol must respect people 's rights and national as well as EU legislation, in particular regarding personal data treatment. Second (post-experiment), understanding people's perceptions of their interaction with the robot in order to improve both its acceptbility and its usefulness.

PAL Robotics

PAL is the provider of the hardware platform ARI. It develops the robot according to specs requested by the different partners such as type and position of cameras and microphones, CPU/GPU/memory, etc. It products 6 robots.

ERM Automatismes

ERM is in charge of the exploitation roadmap of the project: selection of use cases and generation of scenarii of interaction and integration of all software blocks.

Inria

The Inria team works toward interfacing all of the elements produced by the other partners, and looks to create decision and action schemes for the robot when interacting with several (or a group of) humans. Specifically, this requires to achieve the fusion od audio and visual signals for an optimum understanding of the robot environment, and to generate via simulation and training numerous scenarii for action, movement (exploration), behaviour and attitude (expcepted verbal interaction). This step is crucial for the project to push the robot's performances beyond the current scientific state-of-the-art. Inria is also in charge of the overall coordination of the project.

Prague Technical University (Czech Republic)

CVUT is in charge of navigation aspects. Indeed the robot requires environmental awareness in order to be able to move and seamlessly interact, in dialog, with elements of its environment. This is valid for fixed elements (walls, doors, benches etc) as well as for mobile ones (wheelchairs, plants, foldable chairs, etc). A first 3D mapping of the environment is performed in advance and installed in the robot's memory. Additional, mobile elements are detected and categorised on-the-fly by analysing signals from its cameras through a specially trained AI.

University of Trento (Italy)

The UNITN team is specialised in understanding of human behaviour. In short: what are the attitudes and feelings of the people interacting with the robot? To perform this, the team works on a dedicated AI that collects, analyses and exploits visual signals. For example by detecting a smile on a person's fce one can infer that he or she has rather positive feelings/attitude; on the other hand a person frowning might be categorised as shocked or rejecting the robot. The information is of high importance if one wants to provide the robot with autonomous decision making in social relations.

Heriot-Watt University (United Kingdom)

HWU's expertise is at the very end of the human-robot interaction chain: dialogue management. To hold a normal conversation one needs to understand the meaning of uttered words, generate a logical answer and transcribe it in clear understandable language. The first and last step are performed through automatised semantic analysis (once to decode -- understand words directed at the robot-- and once to encode --transform meaning in a logical suite of words understandable by humans). Generating a meaningful response is done through a fine tuned IA that has to take into account information from all of the other blocs developed by the project partners, in real-time.

Bar-Ilan University (Israël)

BIU is focused on audio signal treatment and cleaning. Natural dialogue between humans in a realistic, closed space implies a whole chain of actions that we humans do seemingly effortlessly: identify who speaks, isolate speech from surrounding noise and potential echo; and if the person is not in front of us, understand where sound comes from and move towards it while having in mind that moving will modify the sound signal. All of this is needed for the robot, and is done through real time analysis of audio signals collected by its microphone array using in-house algorithms based on sound propagation equations.

An artificial intelligence system capable of self-correction

But how do we enable a robot to identify from a set of conversations which request is addressed to it; to understand that it is being asked where a person may sit; to look around and find a vacant seat, determine the path to accompany the speaker to their seat while avoiding other patients and staff on the premises, and then perceive the relevance of offering distraction in the form of conversation? There are numerous technological difficulties and hurdles to overcome in order to accomplish this type of complex task.

With regard to movement, RobotLearn opted to implement the reinforcement learning method. In order to determine its speed, approach angle and other parameters of movement, the robot is trained through an artificial intelligence system which calculates the adequacy between optimal action and the action actually undertaken, and attributes “rewards” for successful outcomes. This training phase enables the robot to come across a wide variety of possible cases in full autonomy, without human intervention to correct pathways. Once placed in real conditions, ARI continues to learn and identify the optimal action for each situation. This opens up the possibility of its use in a hospital setting.

Using a hospital setting to test the assets of a social robot

This is the aim of the second phase of SPRING, which is due to start in 2022: to validate the use of the robot in a hospital and to assess its impact on users and their habits, in addition to its acceptability. Entrusting even simple social tasks to a robot is nevertheless far from innocuous and raises numerous ethical and organisational issues, which Etienne Berger, a research engineer at the Broca Living Lab, is looking into (see insert).

In any case, one thing is clear for Xavier Alameda-Pineda: “I believe it’s neither possible nor desirable to replace humans by robots.”It is more a matter of complementarity, and of taking advantage of what the social robot can provide, i.e., the ability to carry out repetitive tasks with a constant level of quality and a lack of judgement which can reassure users.

Ethics and acceptability

To supervise the research being led by the SPRING consortium, a strict ethical framework based on existing recommendations and directives was put in place. Approved by the competent bodies (CPP, CNIL, etc.), it ensures the protection of subjects involved in the hospital trials.

At the same time, Etienne Berger, who is responsible for monitoring these issues for SPRING, is leading a study on the ground to assessthe acceptability of the robot by various service stakeholders (users and accompanying persons, medical and administrative staff) and to measure the organisational impact of the robot on the running of the service.

Beyond the protection of personal data (in particular, medical data), this project raises questions with regard to the management of psycho-social effects of the robot on the various stakeholders of the hospital service (patients and accompanying persons, medical and administrative staff). According to the researcher, the watchword is “to anticipate risk situations and take the vulnerability of users into account.” Initial surveys in the field reveal a mixture of apprehension and expectation among future users, in addition to curiosity and pragmatism. The implementation of these new forms of gerontechnology would appear to be inevitable.

Xavier Alameda-Pineda in five key dates

2010: 2nd year of Masters in Computer Science in France, then thesis with the Perception team at INRIA Grenoble Rhône-Alpes.
2014 - 2016: Post-doctorate at Trento University, Italy.
2016: Joins the Perception team once again, as research fellow.
2018: Receives the SIGMM Rising Star Award from the scientific society ACM, for his contribution to the understanding of multi-modal social behaviour.
July 2021: Becomes head of the RobotLearn team, successor to Perception.

To find out more

Des robots au service des malades et des élèves, National Geographic, 30/10/2020.
Sommes-nous prêts à cohabiter avec les robots ? CNRS Le journal, 21/09/2020.