Periodic Reporting for period 2 - NEUSEQBOT (NEUro cerebellar recurrent network for motor SEQuence learning in neuroroBOTics)
Periodo di rendicontazione: 2023-01-11 al 2024-01-10
The new generation of compliant robots, designed to safely interact with humans in unstructured environments, requires control systems able to naturally deal with their “biological features”. These robots can be efficiently controlled using biologically inspired control systems based on brain regions such as the cerebellum. This nucleus plays a key role in fluent body movements, being essential for adaptive motor control and coordination of body movements. The cerebellum has been traditionally modeled as a feed-forward network with two inputs and one output. Nevertheless, recent experimental studies have demonstrated the existence of multiple recurrent connections in the cerebellum: 1) nucleo-cortical connections (NCCs), and 2) nucleo-olivary connections (NOCs). These recurrent connections back-propagate the cerebellar output activity to the cerebellar inputs, thus shifting the feed-forward toward a recurrent approach. NEUSEQBOT project studied both recurrent connections, trying to undertand how they could contribute to the motor sequence learning capabilities in the cerebellum. This multidisciplinary study combined neuroscientific experiments in animals, cerebellar modelling and neurorobotic applications. Firstly, we experimentally studied the NCC effect in the cerebellar dynamics during reflexive eyelid movements in optogenetically modified mice. These experimental results were used during the cerebellar modelling process. Finally, the resulting cerebellar model was validated in a neurorobotic object manipulation task using a compliant robotic arm. Within the objectives of H2020, NEUSEQBOT project aimed to advance our understanding of how the cerebellum (as a recurrent network) processes the sensorimotor information to generate the required motor command sequences, applying this knowledge to develop biologically inspired control systems for neurorobotic applications with compliant robots. This work has enabled the experienced researcher to enhance his position at the forefront of advances in these fields.
This project can be divided in three main tasks: 1) experimentation in mice to study the NCCs contribution to the generation of movement sequences, 2) development of new and more advanced cerebellar computational models based on recurrent networks, and 3) development of biologically inspired control systems based on cerebellar models for robot control.
Regarding the experimentation in mice, we designed and performed three different experiments to study how the cerebellum, and more concretely the NCCs, could contribute to the generation of movement sequences. We used a modified version of the eye blink conditioning (EBC) experiment for this study. In the original version, the mouse had to correlate a conditional stimulus (CS), a bright light, with an unconditional stimulus (US), an air puff in the eye, which is generated after the CS. After several training sessions, the cerebellum in the mouse learns to correlate both events, generating a conditional response (CR) that closes the eyelid after the CS and just before the US. In this project we modified this original experimental protocol, introducing another puff in the other eye (US2) after the first puff (US1). Thus, the mouse should generate a sequence of movements after the CS, closing both eyes (CR1 and CR2) in a coordinated way just before the US1 and US2 respectively.
Analyzing the brain anatomy, the CR1 could propagate between both sides of the cerebellum following different pathways: a) NCCs, b) red nuclei, and c) red and facial nuclei. We firstly performed optogenetics experiments in several genetically modified mice, implanting an optical fiber in the cerebellum to disable the generation of the CR1 and checking the effect in the CR2. The experimental results demonstrated that the first cerebellum used the CS to generate the CR1, while the second one just used the CR1 and US1, but not the CS, to generate CR2.Thus we demonstrated the capacity of both sides of the cerebellum to generate and coordinate movement sequences.
After the optogenetics study, we performed two additional pharmacological experiments to respectively inhibit the red and facial nuclei, “disabling the propagation” of the CR1 and checking the effect in CR2. In both cases, the suppression of the CR1 also produced the suppression of the CR2 (the same results that we observed in the optogenetics experiment).
Regarding the cerebellar modeling, we developed and published several advanced recurrent cerebellar models, using recurrent NOCs, to modulate the activity in the Inferior Olive (one of the cerebellar inputs) during simulated Vestibulo-Ocular Reflex (VOR) experiments. We have also implemented a full recurrent cerebellar model implementing both NCCs and NOCs, able to reproduce and explain the experimental results obtained during the first task.
Finally, for the neurorobotic experimentation, we have integrated the full recurrent cerebellar model in the sensorimotor system of a robot. Thanks to the new NCCs, this cerebellar model was able to inplicitely identify and model the manipulated object by the robot, adjusting the motor commands to perform the desired task with whatever object. This learning process is non-destructive in the new cerebellar topology, solving one of the main issues of feed-forward cerebellar models.
Regarding the experimentation in mice, we designed and performed three different experiments to study how the cerebellum, and more concretely the NCCs, could contribute to the generation of movement sequences. We used a modified version of the eye blink conditioning (EBC) experiment for this study. In the original version, the mouse had to correlate a conditional stimulus (CS), a bright light, with an unconditional stimulus (US), an air puff in the eye, which is generated after the CS. After several training sessions, the cerebellum in the mouse learns to correlate both events, generating a conditional response (CR) that closes the eyelid after the CS and just before the US. In this project we modified this original experimental protocol, introducing another puff in the other eye (US2) after the first puff (US1). Thus, the mouse should generate a sequence of movements after the CS, closing both eyes (CR1 and CR2) in a coordinated way just before the US1 and US2 respectively.
Analyzing the brain anatomy, the CR1 could propagate between both sides of the cerebellum following different pathways: a) NCCs, b) red nuclei, and c) red and facial nuclei. We firstly performed optogenetics experiments in several genetically modified mice, implanting an optical fiber in the cerebellum to disable the generation of the CR1 and checking the effect in the CR2. The experimental results demonstrated that the first cerebellum used the CS to generate the CR1, while the second one just used the CR1 and US1, but not the CS, to generate CR2.Thus we demonstrated the capacity of both sides of the cerebellum to generate and coordinate movement sequences.
After the optogenetics study, we performed two additional pharmacological experiments to respectively inhibit the red and facial nuclei, “disabling the propagation” of the CR1 and checking the effect in CR2. In both cases, the suppression of the CR1 also produced the suppression of the CR2 (the same results that we observed in the optogenetics experiment).
Regarding the cerebellar modeling, we developed and published several advanced recurrent cerebellar models, using recurrent NOCs, to modulate the activity in the Inferior Olive (one of the cerebellar inputs) during simulated Vestibulo-Ocular Reflex (VOR) experiments. We have also implemented a full recurrent cerebellar model implementing both NCCs and NOCs, able to reproduce and explain the experimental results obtained during the first task.
Finally, for the neurorobotic experimentation, we have integrated the full recurrent cerebellar model in the sensorimotor system of a robot. Thanks to the new NCCs, this cerebellar model was able to inplicitely identify and model the manipulated object by the robot, adjusting the motor commands to perform the desired task with whatever object. This learning process is non-destructive in the new cerebellar topology, solving one of the main issues of feed-forward cerebellar models.
The experimental results obtained from the experimentation in mice have demonstrated the capacity of the cerebellum to generate and coordinate sequences of body movements.
The new computational cerebellar models published in two papers represent a step forward in the understanding of the computational mechanism used by the cerebellum to process the sensory information and generate the corresponding motor commands. This knowledge was used in the development of our last cerebellar model integrating all the additional knowledge acquired during experimentation phase in mice. We hope this new recurrent cerebellar model, integrating the feedback NCCs and NOCs, will help to better understand how the cerebellum is able to deal with our own body movement in order to manipulate different objects with different properties. This knowledge is important by itself, but additionally, it will allow us to improve the development of biologically inspired control systems based on the cerebellum able to operate the new generation of neuromorphic robots.
In addition to a cerebellar model, a biologically inspired control system requires many other components to communicate with and control a real robot in real time. We have already developed and published in a paper a complete set of tools, including a real-time spiking neural network simulator, able to perform neurorobotic experimentation controlled by artificial cerebellums. Anyone could now test their cerebellar models in a realistic neurorobotic task using this set of tools.
The new computational cerebellar models published in two papers represent a step forward in the understanding of the computational mechanism used by the cerebellum to process the sensory information and generate the corresponding motor commands. This knowledge was used in the development of our last cerebellar model integrating all the additional knowledge acquired during experimentation phase in mice. We hope this new recurrent cerebellar model, integrating the feedback NCCs and NOCs, will help to better understand how the cerebellum is able to deal with our own body movement in order to manipulate different objects with different properties. This knowledge is important by itself, but additionally, it will allow us to improve the development of biologically inspired control systems based on the cerebellum able to operate the new generation of neuromorphic robots.
In addition to a cerebellar model, a biologically inspired control system requires many other components to communicate with and control a real robot in real time. We have already developed and published in a paper a complete set of tools, including a real-time spiking neural network simulator, able to perform neurorobotic experimentation controlled by artificial cerebellums. Anyone could now test their cerebellar models in a realistic neurorobotic task using this set of tools.