Clement Thorey

I lead the development of an autonomous robotic harvesting platform providing growers with a competitive precision indoor harvesting solution in greenhouses.

Beside the integration of multiple RGB and RGBD cameras, I implemented, trained and productionized state-of-the-art machine-learning models to detect, segment and estimate grasp-pose candidates for all the targets directly from the camera feeds. I implemented the extrinsic camera calibration routine as well as the low latency/high throughput system which fuses the knowledge from each camera feed into a single consistent world map. Additional unique capabilities of the system include the 3D math to replan the optimal sequence of stops to maximize yield/throughput during the harvest every time new information is aggregated.

In addition to the core architecture, I also maintain complex data/training pipelines to ensure our robots are continuously learning from their own field experiences as well as robot performance dashboards which we use to inform our weekly planning meeting.

An inclusive but not exhaustive list of the technology I have worked with includes:

• 3D point cloud instance segmentation (PointNet++/DGCNN/Point Transformer)
• 6DOF grasp pose estimation (PointNet++ together with a custom head)
• 2D instance segmentation (mask-rcnn/SSD300/detectron2/SOLOv2)
• Inverse Kinematics and Trajectory planning for a 6DOF arm manipulation (MoveIt!, Reinforcement learning (PPO))
• Image classification for safety evaluation and improved automation
• 3D object registrations

Highly competitive worksphop helping scientists over five weeks to transition to Data Science through business lectures and a concrete project with a data-driven company.

• I work in a team of 4, referring directly to the head of the Royal Mail data science team, to providea reliable forecast for each mail type arriving in each delivery office in the U.K.
• Using an ensemble of statistical methods, we were able to decrease the forecast error by 35% over 2016 compared to the model currently in use in the company, ultimately leading to a better allocation of resources over the network

Detection of solidified magma chambers in the lunar crust through numerical simulations and data exploration.

• Successfully used machine learning and statistics, in combination with a Python library I have written, to process and interpret gigabytes of data from the lunar surface and deliver a one-year project as part of the NASA’s GRAIL mission science team.
• Produce efficient pipelines, written in python, to process and visualize gigabytes of data resulting from hundreds of numerical simulations of cooling magma flows.
• Develop excellent communication skills, both in writing by publishing 3 papers in major scientific journals, and speaking, presenting my work in 3 oral awarded presentations in leading international conferences.

• Mathematics - Linear algebra, ODE, PDE, Fourier series, Fourier transform.
• Physics - Mechanics, Experimental Physics.
• Programming - Python.