A Biomimetic Learning Control Scheme for control of Modular Robots

Informations projet

BIOMODULAR

N° de convention de subvention: 705100

DOI

10.3030/705100

Projet clôturé

Date de signature de la CE 22 Février 2016

Date de début 1 Février 2017

Date de fin 31 Janvier 2019

Financé au titre de

• EXCELLENT SCIENCE - Marie Skłodowska-Curie Actions

Coût total

€ 212 194,80

Contribution de l’UE

€ 212 194,80

Coordonné par

DANMARKS TEKNISKE UNIVERSITET
Denmark

Periodic Reporting for period 1 - BIOMODULAR (A Biomimetic Learning Control Scheme for control of Modular Robots)

Période du rapport: 2017-02-01 au 2019-01-31

Résumé du contexte et des objectifs généraux du projet

• What is the problem/issue being addressed?
The problem is to build biologically plausible models (yet functional) to study working hypothesis behind certain characteristics of the neurobiological substrate. Thus, we use robotics to study how the neural system works in the framework of perception action closed-loops. As an example, we aim not to build high performance robotics directly, but rather systems that are able to adapt and learn from actual training towards improving their performance. In this way, we make them very flexible for a wide variety of tasks and scenarios. Furthermore, this allows the control engine to adapt to a robot and optimizing its performance at each stage.

• Why is it important for society?
BioModular has promoted the development of artificial adaptive learning systems embedding spiking neural network and machine learning mechanisms. These learning systems could have wide applicability in robotics. At the same time, understanding the brain will aid the design of biological plausible control schemes to be generalized to any robot, to any conditions. As a matter of fact, mimicking the biological functionality of the central nervous system will lead to create autonomous intelligent robot agents as the next generation of robots. Robots of the future will perform in real environments, maybe partially unknown and/or changing, as living systems do, under control paradigms that go beyond the scope of conventional control algorithms in terms of self-adaptation, self-learning, and self-recognition. Robots that will operate safely in proximity with people and in different fields away from the tightly controlled environment of the factory floor. This will enhance the future generation of self-learning and compliant robots operating in a safely human environment.

• What are the overall objectives?
Objective 1. To design and model an optimized cerebellar-machine learning (CML) network.
Objective 2. To design and implement an efficient framework for online model-based learning control embedded in bio-inspired closed-loop architectures.
Objective 3. To design and develop a novel bio-inspired composite architecture for online model-based learning control.
Objective 4. To validate the composite architecture applicability in modular robotic systems.

Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent

"Overview of the results
The main outcome of this project is a modular spiking implementation of the cerebellar-like circuit that was embedded in three novel bio-inspired control architectures that embed forward and/or inverse internal models: the feed-forward, recurrent, and composite control schemes. Modular stands for the creation of distinct cerebellar microcircuits that are responsible for the learning of the internal model of a robot module or robot joints. Depending on the internal model, inverse or forward, the microcircuits receive varied sensorimotor information that are integrated towards giving motor or sensory outcomes. The architectures were tested on both a physical and real robot to perform a task with changing kinematics and dynamics conditions. The composite architecture combines the advantages of forward and inverse models and the learning of the latter is improved by the prediction error of the first. Tests showed that the architectures have different adaptation scales based on the disturbances and contexts. Different modular robotic configurations were tested for the validation and generalization of the control and learning mechanisms both in manipulation and locomotion tasks within the NRP developed in the framework of the Human Brain Project. The NRP integrates all the tools necessary for embedding artificial brain models to robotic systems. In particular, it facilitates the coordination between spiking neural networks and continuous time system. Moreover, the platform resulted to be a convenient framework for accelerating the experimental set up and analyses thanks to all the available plugins.

Main dissemination activities
• School of Brain Cells & Circuits “Camillo Golgi”, Ettore Majorana Foundation and Centre for Scientific Culture. Lecture title: Integration of cerebellar models into robotic control loops. Erice (Italy), 11-15 December 2018.
• Participation at the ICT Networking Dinner in Vienna, Austria, December 3, 2018. ROBOINSIGHTS’18 The First Global annual NEW Technology Conference & Expo, Elsinore, Denmark, 26-27 Sept. 2018. Talk: ""Neuro-Robotics, connecting body, environment and brain"".
• Participation at the workshop organized by the Technical University of Munich (TUM) for discussing the current work on bio-inspired motor control schemes within the HBP. Munich, Germany, August 20-21, 2018.
• 4th EU-Japan Workshop on Neurorobotics, Embodiment of Informatics with Neurorobotics. Talk title: Using robots for understanding the cerebellar role in sensorimotor control, AIST Tokyo Waterfront, Tokyo, April 18-20, 2018.
• BioModular project was mentioned and acknowledged during the talk presented by Marie Claire Capolei and Carlos Corchado Miralles (research assistants at DTU for HBP). Talk title: “The Human Brain Project - how robotics can exploit brain science”, H.C. Ørsted Institute, Copenhagen, Denmark, March 19, 2018.

Summary of the exploitable results.
-A spiking cerebellar network model with granular and molecular layers and PC-DCN-IO cells was implemented and tested.
-An optimal input-space representation for the granular layer by means of a machine learning engine was provided.
-Bio-inspired control loops were built and tested with a physical robot and neuromorphic hardware SpiNNaker to perform tasks controlled by the cerebellar machine learning network.
-Learning of forward and inverse internal models of the robot was encoded inside an ULM embedded into the bio-inspired control loops to adapt the corrections and make predictions.
-The forward and inverse internal models were combined and embedded in a composite architecture.
-Different modular robotic configurations were tested for the validation of the bio-inspired control architecture in manipulation and locomotion tasks within the Neurorobotics Platform. The inclusion of input from vision was evaluated to enhance the learning and so, the behavior of the control process."

Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)

One of the main advantages of the developed approach is that physical systems can operate in unknown and changing environments overall when the dynamical robot model is unknown or not well defined. Implementing this model for different tasks and different modular robotic platforms, allows a validation of the relationship between the architecture and the human modular cerebellar control. Compared with state-of-the-art, this approach reduces the amount of (sometimes implausible) a priori information the control system needs, leading to a real-time robotic control system that is capable of learning autonomously how to perform determinate physical tasks and of adapting to changing conditions. This approach introduces the fast and flexible modular control architecture that can be applied to different robotic platforms without any/excessive customization. Hence, we developed the bio-mimetic modular learning control architecture that is based on the modular cerebellar learning concept, machine learning optimization, and on the traditional adaptive feedback control.

Cerebellar-like control with the Fable robot, the Spinnaker board, and the NRP.