This page contains the database of possible research projects for master and bachelor students in the Biorobotics Laboratory (BioRob). Visiting students are also welcome to join BioRob, but it should be noted that no funding is offered for those projects. To enroll for a project, please directly contact one of the assistants (directly in his/her office, by phone or by mail). Spontaneous propositions for projects are also welcome, if they are related to the research topics of BioRob, see the BioRob Research pages and the results of previous student projects.
To limit the list to the projects matching a given keyword, click on it.
Animal locomotion critically depends on the precise coordination of the activities of neurons in the spinal cord. This coordination varies across locomotor behaviors. This project will investigate how such coordination works by modelings three types of gaits in the salamander: swimming, trotting and struggling. In each of these gaits spinal neurons display a stereotypical coordinated activity pattern that results in an appropriate limb and spine coordination. During swimming, axial motoneurons generate a wave of left-right antiphase oscillatory activity that propagates rostro-caudally (positive phase-lag), while the limbs are folded. During trot the motoneurons active during swing (stance) on diagonal limbs are synchronised, and drive the axial motoneurons to generate a standing wave (zero phase-lag). During struggling, the axial motoneurons produce a caudo-rostral activity propagation (negative phase-lag). This project will systematically investigate the contribution of ascending and desceding excitatory and inhibitory synaptic interactions in shaping the axial propagation speed using a biologically-realistic spiking neural network model paired with a mechanical model of the salamander using numerical simulations. The model will help to elucidate the role of spinal interneurons' connections, proprioception and desceding modulation from higher brain areas to shape the phase lag observed in the salamander experiments.
The project will involve the following steps: 1) Analysis of kynematic data obtained from swimming salamanders through DeepLabCut 2) Replicating the kynematic recordings in a simulated salamander model in MuJoCo 3) Optimizing an hydrodynamic model (e.g. Lighthill model) to match the experimental data The project will involve the collaboration with biology experimentalists (in Canada) who will gather data from swimming salamanders.
V1 interneurons are a class of ipsilaterally-projecting inhibitory neurons. While they are connected to the rhythm-generating neurons of the locomotor circuits (namely V2a, V0d and dI6 interneuronns), their role in the central pattern generation is poorly understood. In this project, neural data from electrophysiology experiments on adult zebrafishes will be used to develop spiking neural models of V1 neurons. The developed neurons will be connected to an existing model of the zebrafish locomotor circuits in order to systematically study the role of the new connections in the resulting locomotor patterns.
This project aims at exploring the possible role of vorticity in entraining the activity of the spinal locomotor circuits of the adult zebrafish. The neural network will be modeled with spiking neural models (e.g. Adaptive Exponential Integrate and Fire) similarly to Roussel et al. 2021, with the addition of newly discovered axial organs of proprioception sensory neurons (Picton et al. 2021). The mechanical model (derived from CT scans of an adult zebrafish) will be simulated on MuJoCo with the FARMS simulation framework. The simulations will aim at understanding whether and to what extent can vortices present in the environment influence the activity of the locomotor circuits and the swimming performance overall. The experiments will be used to provide possible predictions for future experimental analyses.
This project aims at replicating the experimental setups used for the analysis of the locomotor circuits in the Zebrafish in order to understand the role of the newly discovered axial organ of proprioception (Picton et al. 2021) during swimming in adult individuals. The neural network will be modeled with spiking neural models (e.g. Izhikevic model) similarly to Roussel et al. 2021, with the addition of proprioceptive sensory neurons. The mechanical model (derived from CT scans of an adult zebrafish) will be simulated on MuJoCo with the FARMS simulation framework. The simulations will aim at optimizing the parameters of the network to match the biological experiments. Additionally, the validated network will be used to provide possible predictions for future experimental analyses.
Recurrent neural networks are widely used tools to model complex cognitive and motor tasks in neuroscience. With the forthcoming neural wiring diagram of the fruit fly brain, it is now possible to use this connectivity to reverse-engineer sensory processing and motor control. In this project, we will train RNNs using PyTorch with an architecture constrained by the connectome to recapitulate the experimental data from freely behaving fruit flies. We will then systematically investigate the trained networks to understand the network activity's low-dimensional dynamics. If you’re interested in this project, please send your CV and transcript to Gizem Ozdil via e-mail pembe.ozdil@epfl.ch
Animal behavior arises from a complex interplay between the nervous system, sensory feedback from interactions with the environment, and the musculoskeletal properties of the body. Computational simulations have been instrumental in improving our understanding of these components in the generation and execution of movement as they allow for rapid experimentation. Recent studies in Drosophila neuroscience made the full nervous system connectomes available, as well as the 3D reconstruction of the muscles in the animal’s legs. Thus, there is more data than ever to model the biological neural networks and neuromechanics of the musculoskeletal system in Drosophila to complement the experimental work.In this project, we will focus on Hill-type muscle modeling in Drosophila legs by leveraging the recently published anatomical and physiological papers. We will model the muscles in our recently published biomechanical model NeuroMechFly^ using OpenSim and MuJoCo. This project will have the potential outcomes:
Developing a pipeline to map muscle attachment accurately points from one animal to another (Python, Blender).
Developing a pipeline to build hill type muscle model based on the anatomical data and an optimization framework to estimate the unknown muscle parameters (OpenSim, Python).
Carrying out muscles-in-the-loop simulations in MuJoCo to replicate real animal kinematics.
Familiarity with OpenSim and Muscle models is preferable. It is a plus if the student already took CMC (CS-432) or AML (BIOENG-404). If you’re interested, please send your CV and transcript to Gizem via e-mail pembe.ozdil@epfl.ch
^Lobato-Rios, V., Ramalingasetty, S.T., Özdil, P.G. et al. NeuroMechFly, a neuromechanical model of adult Drosophila melanogaster. Nat Methods 19, 620–627 (2022). https://doi.org/10.1038/s41592-022-01466-7
This project concerns the analysis of several metrics of the balance of exoskeleton users of the autonomyo lower-limb exoskeleton. Autonomyo is developed for people who have walking impairments due to muscular weakness or motor deficits. The laboratory BIOROB and the research group REHAssist have established expertise in human locomotion and lower limb exoskeletons. Autonomyo has been designed at EPFL. Currently, Center of Pressure (CoP) based metrics are used to investigate the balance of different locomotion tasks. However, by using an external measurement system we can also investigate center of mass (CoM) based metrics as well. This project involves different steps: – Familiarization with existing hardware/software and firmware of the exoskeleton – Conduct experiments with autonomyo – Prepare functional code blocks that is synchronizing two measurement systems – Analyze the collected data for various balance metrics Skills: – Experience with Matlab for data analysis – Experience with C++ References: Gordon, D. F., Henderson, G., & Vijayakumar, S. (2018). Effectively quantifying the performance of lower-limb exoskeletons over a range of walking conditions. Frontiers in Robotics and AI, 5. https://doi.org/10.3389/frobt.2018.00061
Last edited: 12/07/2023
Quadruped robotics
A small excerpt of possible projects is listed here. Highly interested students may also propose projects, or continue an existing topic.
Recent years have shown impressive locomotion control of dynamic systems through a variety of methods, for example with optimal control (MPC), machine learning (deep reinforcement learning), and bio-inspired approaches (CPGs). Given a system for which two or more of these methods exist: how should we choose which to use at run time? Should this depend on environmental factors, i.e. the expected value of a given state? Can this help with explainability of what exactly our deep reinforcement learning policy has learned? In this project, the student will use machine learning to answer these questions, as well as integrate CPGs and MPC into the deep reinforcement learning framework. The methods will be validated on systems including quadrupeds and model cars first in simulation, with the goal of transferring the method to hardware.
As robots become more prevalent in human society, the number of interactions will increase and good communication will be critical for successful human-machine collaboration. In this project, the student will develop a framework for human-robot interaction using both visual and audio feedback. Given a set of user-defined "tricks" (i.e. lie down, turn around, move left), how can we instruct the robot to perform a particular task? Can we also teach the robot a new task it currently does not know how to do? Communication will be done using both a camera mounted on the robot, as well as with a microphone. The three important tasks are 1) developing the motion library, 2) developing the visual interface to human activity recognition software to map to the motion library, 3) developing the voice command interface. To apply, please email Guillaume with your motivation, CV, and briefly describe your relevant experience (i.e. with machine learning, software engineering, etc.).
The project aims at building on top of results in markerless furniture perception achieved in previous semesters. The goal is to integrate perception as a feedback signal for real-time furniture control.
Controlling vehicles at their limits of handling has significant implications from both safety and autonomous racing perspectives. For example, in icy conditions, skidding may occur unintentionally, making it desirable to safely control the vehicle back to its nominal working conditions. From a racing perspective, drivers of rally cars drift around turns while maintaining high speeds on loose gravel or dirt tracks. In this project, the student will compare several approaches for high speed, dynamic vehicle maneuvers, including NMPC with a standard dynamic bicycle model, NMPC with a dynamic bicycle model + GP residuals, NMPC with learned dynamics (i.e. a NN), and lastly a pure model-free reinforcement learning approach. All approaches will be tested in both simulation as well as on a scaled vehicle hardware platform.
