jérôme susgin

Initially specialized in teleoperation, Kibitech developped an interface to control a robotic fleet from the cloud: - I helped control the robot behavior on the edge side. - I proposed to develop an AI agent using smolagents able to generate a task plan.

I built a pipeline to generate a 3d half-transparent objects (plastic bottles) from a set of images and camera intrinsics data: - I finetuned Detectron2 to find the plastic bottles boxes with their parts (cap, label...) using the Paco dataset. The segmentation is done using SegmentAnything. The alpha estimation is done using VitMatte. The 3d cloud point estimation and object extraction is done using Neus2. - I built a docker image to make the full application easy to access and deploy.

I controlled the behavior of a robotic arm in a industry 4.0 context: the manipulator is in the center of the converyors delivering wings or fuselage of several colors, the camera is placed above, the microphone close to the operator for vocal command: - I built the python and ROS architecture of the full system. - I controlled the robot using registered plans (internal knowledge) which can be updated with vocal commands. The manipulator moves in real-time to pick the given object on the conveyors. It uses two different types of tools to catch it. - I also helped on some side projects of robotic fleet control.

MyFamilyUp wanted to recognize the emotions in the spoken language but they did not have any knowlede in AI. The IRIT lab had this knowledge but not the knowledge in applied psychology. I was the link between both of the entity: - I had to understand the problem as an AI problem and a psychology problem, provide a solution and build it. - I found suitable data with a suitable volume to train or finetune the models. - I built the model from scratch, trained it and validated it using a remote supercalculator.

First experience. I had to build a system able to generate a world file to input in simulators. The goal was to improve the speed of setting up the test benches to realize in a simulator (gazebo). - I wrote a script taking a list a requirements as a json file and generate a .world file following the constraints. - The constraints were things like: there is a parking on the left at 2m from the car with 3 available spaces located at the given p1, p2, p3 poses.