In many sensing and control applications, data are represented in the form of multidimensional arrays with particular geometric properties such as symmetries. Considering the underlying structure and geometry of the data can be beneﬁcial in many robotics applications. Considering the structure and geometry of the data is even more important when a small set of multimodal data is available.
I exploit tensor methods and Riemannian geometry to develop structure and geometry-aware statistical learning techniques. Instead of transforming high-order data into vectors, I aim at considering their potential multimode relations and handling their underlying geometry. I also consider the case where input modality is complex, meaning that different structured data with different geometries are available and should be combined efficiently and meaningfully.
The aim of the TACT-HAND project is to improve the reliability and the stability of the control of prosthetic hands. It proposes to augment the traditional electromyography (EMG) sensors with tactile myography (TMG) in the form of a muscle bulgings measuring bracelet made of pressure sensors.
We treat the control of prosthetics as a regression problem, fusing information from tactile sensing and EMG sensors. We aim at keeping the structure of the data as EMG are often processed as covariance features and TMG is naturally represented by a cylindric grid. Tensor methods and Riemannian geometry allows us to exploit the spatial and temporal patterns that appear in EMG and TMG data and to improve the detection of hand and wrist movements.
Body posture greatly influences human performance when carrying out manipulation tasks. Adopting an appropriate pose helps us regulate our motion and strengthen our capability to achieve a given task. Similarly, in robotic manipulation, the robot joint configuration affects both the ability to move freely in all directions in the workspace and the capability to generate forces along different axes.
In this context, manipulability ellipsoids are posture-dependent measures indicating the preferred directions to perform velocity, force or dynamic control commands. We propose to transfer manipulability-based posture variation between robots: a teacher robot demonstrates how to perform a task with a desired manipulability profile and a learner robot reproduces the task while matching the learned manipulability profile. This approach allows to transfer posture-dependent task requirements such as preferred directions for motion and force exertion. We encode, retrieve and track robot manipulability ellipsoids using geometry-aware learning and tracking frameworks based on Riemannian geometry and tensor formulation.
Jaquier, N. and Calinon, S. Gaussian mixture regression on symmetric positive deﬁnite matrices manifolds: Application to wrist motion estimation with sEMG, in IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), Vancouver, Canada, September 2017. [pdf] [bibtex] [publisher's website]
Rozo, L., Jaquier, N., Calinon, S. and Caldwell, D. G. Learning manipulability ellipsoids for task compatibility in robot manipulation, in IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), Vancouver, Canada, September 2017. [pdf] [bibtex] [publisher's website]
28.11 Swiss Machine Learning Day, EPFL, Switzerland
14.11 Applied Mathematics Seminar, Applied Mathematic Dept., UCL, Belgium
24.09 Workshop on Micro-data: the next frontier in robot learning?, IROS 2017, Vancouver, Canada
23.05 Numerical Analysis Seminar. Mathematic Dept., University of Geneva
31.01 Rapid-fire talk. Applied Machine Learning Days, EPFL, Switzerland
20.01 Operating room technicians class, Upper School of the Health (ES Santé), Lausanne, Switzerland
I developed controllers for the humanoid robot Baxter to reproduce a simple image with pen strokes while exploiting the compliance of the arms of the robot and the visual feedback provided by its cameras.
I was part of the research and development team designing aerial imaging drones for professional applications. My task was to develop flight control algorithms for an intelligent mapping and inspection quadrotor (the albris).
I implemented a shared-control paradigm, already implemented in simulation, on a WAM robot. The neuroprosthesis executed actions that were evaluated by the user as onerous or correct and the brain correlates of this assessment was exploited to learn suitable motor behaviours.
I used model-free techniques for speed estimation on data collected in two different robots; Oncilla and Pleurobot. This allowed to estimate their instantaneous velocity and to know if they get trapped, fall, are deviated or move as expected.
I modelled a multi-modal bio-inspired walking and flying robot on the simulator Webots and used genetic algorithms to evolve the parameters of the motors to optimize the gait of the robot. This allows to reduce the cost of transport of the gait and to optimise the gait depending on the terrain.