CE33 - Interaction, Robotique – Intelligence artificielle

Morpho-functional Swarm Robotics – MSR

Morpho-functional Swarm Robotics

In this project, we are interested in swarm robotics, where a large number of robots with limited computation and communication power are considered. Our goal is to propose new design methods, with a particular focus on collective decision making using both morphological and logical computation.

To do so, we aim at new kind of swarm bots, the vibebots, where the shape factor guarantees the group dynamics. We will then look for specific educated collective behaviors.

The term swarm intelligence is used to describe complex systems where collective behaviors emerge from simple local interactions between individuals. This phenomenon can be observed in natural systems (e.g., eusocial insects) as well as implemented in artificial systems (e.g., swarm robotics), and involves behaviors such as collective motion, foraging and nest building.<br />We are interested in swarm robotics, where a large numbers of robots with limited com- putation and communication power are considered. Our goal is to better understand swarm intelligent behaviors in nature, and to propose new design methods, with a particular focus on taking into account the role of physical interactions among individuals.<br /><br />The building hypothesis stems from our belief that the role of embodiment has been over- looked in swarm robotics so far. We posit that embodiment is critical and useful for programming self- organizing collective systems. We thus propose to achieve collective decision making using both morpho- logical and logical computation in swarm robotics.

To achieve the present challenging project, the two PI’s need to put together their knowledge, skills and creativity. Our project being truly interdisciplinary, we pay attention to maintain a close collaboration at all levels and for all purpose, as we shall now demonstrate.

At the conceptual level, despite a true difference of vocabulary, which they have already started to reduce, through a common lexicon, the two PI’s share a common culture of collective autonomous systems (self- organization, control parameter tuning, morphological computation, optimization, etc.). This allows a rich cross-fertilization when conducting the experiments and analyzing the results (WP3,WP4,WP5).

Our approach combines (a) In Materio Experiments building on the granular walkers we described earlier and (b) Evolutionary Computation (EC) and Quality Diversity Algorithms (QD) to explore the set of possible hardware designs embedding either no or minimal logic computation that are able to match quantifiable criteria describing a desired collective behavior.

As planned initially, we have started the project using Kilobots. a 3.3 cm tall low-cost swarm robot (Rubinstein et al 2012 IEEE). Kilobots are endowed with an isotropic IR-based broadcast communication device (range of approx. 70mm), an ambient light sensor, and two on-board vibration motors used to modulate translational and angular speed, all controlled through a Atmega 328 processor.

We have rapidly evaluated that adding a vibration platform does not provide any significant improvement to the Kilobots and is therefore not a priority anymore. Having set an arena and an Image Acquisition and Processing system, Kilobots experiments are conducted in an environmentally controlled arena. (D1.1).

We then characterized the individual dynamics of the Kilobots (D3.1). The main conclusion is that Kilobots lack the basic ability to go straight without an external reference robot (or light source). Furthermore, there is a great robot to robot variability, and the calibrations are not robust in time. We have bypassed these shortcomings by adapting the Kilobots morphology complementing them with exoskeletons (see C.3).

We obtained three major results.
First the exoskeleton design allows us to control physically the self-alignment property of the Vibebots (D3.1)
Second we could study the influence of the shape of the exoskeleton on the physical interaction between Vibebots when they collide. (D1.2).
Third we developed a simulation engine to study the dynamics of active agents with aligning interactions (D2.1). We demonstrate two important effects. First, the purely morphological response of a swarm of Vibebots when facing an obstacle is strongly dependent on its self-aligning dynamics. Second, an aggregated cluster of Vibebots resist evaporation when it is composed of anti-aligning Vibebots. Here also, this collective behavior is purely of morphological origin.

Having obtained a robust, and ten times faster Vibebots than the Kilobots, by the design of an exoskeleton.
Having identified a robust correlation between the speed of the Vibebots and its self-alignment property
Having a numerical indication that the self-alignment property can be used as a primitive in controling the interactions of Vibebots with an obstacle or with other Vibebots.
Having designed a distributed learning algorithm for swarm robotics, coping with limited communication. Published at IEEE CEC 2020, Among the best student papers nominee (5 papers nominated).
Having studied how conditional partner choice in robot learning in a cooperative game enables socially optimal cooperation in swarm robotics under specific conditions. Published at ALIFE 2020 for swarm robotics results, and under review in an evolutionary biology journal for the fundamental research.

Nicolas Fontbonne, Olivier Dauchot and Nicolas Bredeche. Distributed On-line Learning in Swarm Robotics with Limited Communication Bandwidth. Proceedings of the IEEE Congress on Evolutionary Computation (CEC), 2020 (1 of 5 best student paper nominees).

Paul Ecoffet, Jean-Baptiste Andre´, Nicolas Bredeche. Learning to Cooperate in a Socially Optimal Way in Swarm Robotics. Published at the Conference on Artificial Life (ALIFE), 2020.

Paul Ecoffet, Nicolas Bredeche, Jean-Baptiste Andre´. Nothing better to do? Environment quality and the evolution of cooperation by partner choice. Preprint on Biorxiv (currently under review at JEB – journal of evolutionary biology)

In this proposal, we aim at better understanding swarm intelligent behaviors in nature, so as to propose new design methods for swarm robotics. We intend to emphasize the role of physical interactions among robots, which we beleive has been under-estimated so far. The building hypothesis for this proposal is that embodiment is critical for programming self-organizing collective robotics systems. The challenge is to reveal how to balance collective decision making between individual's morphological computation (i.e., arising from the physical design of the robots) and logical computation (i.e., the control architecture).

This truly interdisciplinary project combines the partners' expertise in experimental and theoretical physics (collective effects in active matter) with that of computer science and robotics (swarm robotics and evolutionary optimization). Our approach will involve both (1) designing a new kind robot capable a swarm behaviour based solely on physical interactions and (2) the use of exploration and optimization methods to navigate the search space of possible software-based controllers.

More specifically, we will design robots such that the shape factor guarantees group dynamics, with each robot embedding a lightweight system-on-chip, sensors and effectors that can be used to modulate the robot behaviors on-the-fly. We will start by analyzing the spontaneous phases obtained from the purely physical alignment of self-propelled particles. Then, we will characterize operational phases, i.e. swarm behaviors that can be obtained when minimal control is applied using the robots' embedded motors. Finally, we will optimize the robots' controllers to achieve specific collective tasks, either considering that all robots can independently make decisions (swarm of individuals) or that only a subpart of the group can control the whole group (swarm shepherding).

Our method relies on three complementary strategies. The crudest one consists in altering the physical behavior at the level of each individual robot, in such a way that new spontaneous phases emerge. The alteration, although minimal, can be extremely efficient if it alters the symmetry of the physical rules driving the robots. At a more sophisticated level, we will identify the so-called soft-modes of the spontaneous phase and modulate them, taking advantage of their very low stiffness. Finally a third strategy will consist in looking for critical boundaries amongst the possible spontaneous phases and induce switching behavior through minimal control.

Two swarm robotics tasks will be used as a testbed and for validation of our approach. First, we will investigate how to steer a swarm of self-aligned robots towards non-linear trajectories, arising from individual decision only. Second, we will explore how to modulate interactions between several robot swarm, composed of similar individuals but with different settings (velocity, alignment, direction, etc.), so as to control the results from swarm to swarm interaction (e.g. merging, laning, etc.).

Project coordination

Olivier Dauchot (GULLIVER)

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

ISIR Institut des Systèmes Intelligents et Robotiques
GULLIVER

Help of the ANR 596,700 euros
Beginning and duration of the scientific project: December 2018 - 48 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter