The CIAD laboratory is searching for a PhD candidate to work on cooperative control and planning, cooperative environmental perception,machine learning, multi-agent systems and robot behavior analysis.
The PhD project is funded by the Technology University of Belfort- Montbeliard for 36 months.Starting period is September/October 2021.
Application details are into the attached file or available on https://www.linkedin.com/feed/update/urn:li:activity:6794917064424341505/
1) Introduction / background: Several applications in the field of robotics require interactions between robots toaccomplish their task. These interactions can be conflictual as in the case of space sharing,or collaborative as during handling operations. The movements of the robots in both casesmust be synchronized to perform their tasks safely. Due to the uncertain environment,especially in the presence of humans, these movements can experience delays, hence theneed to share the perception of the environment. There are two possible solutions tomeet this need. The first one consists of building a global dynamic representation mapshared and updated by all robots. This assumes that it must be managed centrally. In thesecond approach, which is decentralized, the robots communicate interfering elementswith each other. To do this, they must be able to classify the states of the environmentand jointly define the different sources of delay to synchronize accordingly. Two scientificbuilding blocks are identified in the proposed thesis subject.Cooperative planning and control: This involves studying interaction models and analyzing theproperties of control or trajectory planning. In addition to the properties of the solutions, themodel will be used to deduce the relevant information to be exchanged between the robots.Other control or planning techniques can be exploited. Through these analyzes, the student will beable to address the thorny issue of multi-agent reinforcement learning in the context ofcontinuous decision-making [1]. The aim here is to test the potentials of Deep ReinforcementLearning (DRL) in the context of the learning of several agents [2]. Also, other distributed controlstrategies can be deduced, explored and compared.Dynamic cooperative perception: This involves sharing the perception of a robot's environmentwith other robots and vice versa in a collaborative context [3]. The objective is to increase theperception of each of the robots in order to offer them broader perspectives to carry out theirindividual and collective tasks as well as possible [4]. In general, each robot, equipped with one ormore sensors (cameras, Lidars, etc.), must be able to locally perceive its surrounding space, thenintegrate all the information useful for the mission of each robot in its perception or knowledgemap [5]. The student will focus particularly on creating a dynamic representation of the perceptionof each robot by exploiting its own perception and those shared by other robots. The objectivehere is to understand the dynamic content of the environment by recognizing situations or eventsthat may cause difficulties to the robot itself, but also to other robots participating in the collectivemission. This representation requires a spatial and temporal registration, which can be complexdepending on the type of information shared.2) Planned works:The two scientific topics presented above will have to be treated and exploited jointly.Cooperative control and planning can benefit fromdynamic cooperative perception and vice versa.Indeed, the results of perception will be exploited to optimize the control of the robots, and inreturn, the perception process will exploit the robots control or planning to improve theirperception in terms of prediction for example. From a practical point of view, the sharing andupdating of the perception map of each robot can be done at the request of the robot concerned(to other robots) or can be detected automatically as part of a strategy defined by the missionitself and made known to all robots participating in the mission.For experiment and testing, the student will benefit from an application in a concrete case ofcollaboration between several real robots and a computing platform. The data will be generatedthrough real and augmented tests. [1] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. 2017. Multi-agentactor-critic for mixed cooperative-competitive environments. In Proceedings of the 31stInternational Conference on Neural Information Processing Systems (NIPS'17). Curran AssociatesInc., Red Hook, NY, USA, 6382–6393.[2] OROOJLOOYJADID, Afshin et HAJINEZHAD, Davood. A review of cooperative multi-agent deepreinforcement learning. arXiv preprint arXiv:1908.03963, 2019.[3] SCHMUCK, Patrik et CHLI, Margarita. CCM ‐ SLAM: Robust and efficient centralized collaborativemonocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics,2019, vol. 36, no 4, p. 763-781.[4] QUERALTA, Jorge Pena, TAIPALMAA, Jussi, PULLINEN, Bilge Can, et al. Collaborative Multi-Robot Search and Rescue: Planning, Coordination, Perception, and Active Vision. IEEE Access,2020, vol. 8, p. 191617-191643.[5] YANG, Chule, WANG, Danwei, ZENG, Yijie, et al. Knowledge-based multimodal informationfusion for role recognition and situation assessment by using mobile robot. Information Fusion,2019, vol. 50, p. 126-138.--
Laboratoire Connaissance et Intelligence Artificielle Distribuées
CIAD UMR 7533 Prof. Dr. Stéphane GALLAND
Full Professor of Computer Science and Multiagent Systems
Deputy Director of CIAD
French Head of ARFITEC ARF-17-11 & ARF-19-11 "Energy, Transport, Industry, Challenges for tomorrow"
Senior member of the Multiagent Group
Member of AFIA Université de Technologie de Belfort-Montbéliard - UBFC
13, rue Ernest Thierry-Mieg
90010 Belfort Cedex, FRANCE CIAD Lab: www.ciad-lab.fr
Web: www.ciad-lab.fr/author-10836
Phone: +33 384 583 418 (work office)