Neurorobotik
Research
Neurorobotik 

Research

Neurorobotics is an emerging science at the intersection of computational neuroscience and robotics. It studies the interaction between brain, body, and environment in closed perception-action loops. At the core are robots controlled by simulated nervous systems that model the structure and function of biological brains at varying levels of granularity. In a typical neurorobotics experiment, a robot will perceive its current environment through a set of sensors that will transmit its signals to a simulated brain. The brain model may then produce signals that will cause the robot to move, thereby changing the agent's perception of the environment. Observing how the robot then interacts with its environment and how the robot's actions influence its future sensory input allows scientists to study how the brain and body work together to produce an appropriate behavioural response. Thus, neurorobotics links robotics and neuroscience, enabling a seamless exchange of knowledge between these two disciplines.

Nermo Mouse Robot

This biomimetic robot is modular and low-cost; it was created to mimic the locomotion of a rodent and has the size of a common rat (Rattus norvegicus). The robot is untethered, easy to use, and simple to produce; it thus can be used as a universal research platform. It is based on tendon-driven actuation, which enables the implementation of a compliant leg and body design and thus enables adaptive and dynamic walking motions. Small biomimetic robots can be helpful for behavioural/social studies in combination with animals, or for investigating new, efficient types of locomotion for exploration systems. Combined with digital twins, they are useful to reduce the reality gap between simulation and the real world.

Franka Emika Panda

Panda is a collaborative robot with 7 degrees of freedom developed by the German company Franke EMIKA. This robot allows direct control and the possibility to program it and connect it with external sensors (packages and libraries for C++, ROS, and Movelt!). The robot can be used for several interesting applications, such as exploring the use of Spiking neural networks to control the robot or deploying the robot to help the elderly.

Distributed Deep Reinforcement Learning for Multi-agent UAV control systems

Here multiple drones work together as Reinforcement Learning Agents to achieve a task. Each agent learns to adapt its behaviour to its peers. Knowledge and experience acquired by one drone are shared with the other drones.

Central Pattern Generators With iCUB

iCub is a humanoid robot from the theRobotCub project, which aims to be an open research platform for cognition.

Central Pattern Generators are a set of neurons found in the spinal cords of various animals. These generators can generate rhythmic patterns to control the joints of animals to perform tasks such as walking. In this project, CPGs were used to control the shoulder and elbow joints of the iCUB robot.

Event-Based Camera

Event-based cameras are a type of optical sensor that only reacts to changes in the environment. This property allows them to have a very high response rate and very low power consumption. Each pixel inside the event camera operates independently and asynchronously, reporting changes in brightness as they occur. This makes them a good candidate for neurorobotic applications, for example when coupled with spiking neural networks.

C.Elegans inspired autonomous delivery robot

Neural Circuit Policies (NCPs) are designed sparse recurrent neural networks loosely inspired by the nervous system of the organism C. elegans. The idea of this project is to use NCPs to attempt self-driving based on visual data. Currently, the robot can maintain lanes in simulation, and the next step is to build a physical version of the robot and develop the model further.

Parametric Study for a Supervised Spiking Neural Network

Spiking neural networks (SNNs) are a type of neural network that utilizes biologically plausible spiking neurons to process temporal or spatiotemporal information. However, developing stable and efficient supervised learning algorithms for SNNs is difficult due to their complex and nonlinear mechanisms. This research aims to review previous efforts in classifying, identifying, and formulating supervised algorithms for SNNs, as well as to present an experimental study of a proposed SNN architecture. The proposed architecture includes encoding and decoding components in addition to a fully connected convolutional spiking layer and uses a surrogate gradient to represent the gradient during training. The individual components of the architecture are validated independently before being assembled and tested on the Fashion-Mnist dataset. The architecture is then advanced to be used as part of a behavior cloning algorithm for training a model of a Franka Emika Panda robot on the ManiSkill2 Challenge. The results indicate that the Atan and fast sigmoid functions are the best surrogate gradient functions in terms of performance and that the backpropagation through time algorithm is the best backpropagation algorithm.