Research Internship - 3D Computer Vision - PhD Track

Our missionA robot that doesn't know where it is can't reliably navigate, plan, or act. Localization is the foundation for everything else, yet it remains an open problem: ports and tunnels with no usable GNSS, warehouse aisles that look identical in every direction, construction sites whose geometry changes daily, scenes saturated with moving objects that corrupt the very map you're building from them.At Exwayz, we build the LiDAR perception stack that makes reliable autonomy possible: real-time SLAM and localization at sensor rate, centimeter-level accuracy, robustness to geometric degeneracy and dynamic scenes, and sensor-agnostic performance across LiDAR brands and scan patterns. On top of that foundation, we're building the perception layer that turns raw point clouds into something a robot can act on: detection, segmentation, mapping, and change detection.What we care about is generality: methods that remain sensor-agnostic and keep working on real-world data beyond the distribution of public datasets. That's the bar we set for our own work.We're a team of 8, already in production with clients across Europe and the US.Your roleYou'll work on the frontier of LiDAR perception: semantic and panoptic segmentation, object detection, and the questions that sit around them.The internship is structured in two phases. You'll start on the classical single-scan setting, building a solid understanding of the state of the art and of where it actually breaks on our data. From there, you'll move to what we think is the more interesting question: exploiting the temporal dimension of LiDAR data. A robot doesn't see one point cloud, it sees a continuous stream, and it already knows how it moved between scans because our SLAM tells it. Almost no perception method makes real use of that. Aggregating geometry over time, propagating labels across frames, separating what moved from what the sensor simply saw differently, using ego-motion as a free supervision signal: this is where we want to push.The internship is research-first. We expect it to produce a publishable contribution, and if the work and the fit are there, it can open onto a CIFRE PhD position.ResponsibilitiesSurvey the state of the art in your research area and build a clear picture of what works, what doesn't, and whyFormulate research hypotheses and design experiments that can actually falsify themImplement and train models on our internal and public LiDAR datasets, with access to HPC resourcesBenchmark rigorously: ablations, failure-case analysis, honest baselinesMove promising results from notebook to real-time-capable code, in collaboration with the engineering teamPresent your work internally, and write it up for publication at a top-tier venueStackPython, PyTorch, C++, ROS, Git, Linux.Candidate requirementsRequiredFinal-year engineering school or Master's student, looking for an end-of-study internshipSolid grounding in deep learning, and comfortable implementing and debugging non-trivial PyTorch codeHands-on experience with 3D perception and / or LiDAR point cloudsAble to read, reproduce and critique state-of-the-art papersScientific rigor: you form hypotheses, you test them, and you change your mind when the numbers say soWorking French and EnglishNice to haveA prior publication or a serious attempt at oneExperience with SLAM, registration or sensor fusionC++ and real-time or embedded constraintsExperience with HPC / SLURM environmentsConditions6-month end-of-study internship, start date flexibleBased at our Saint-Lazare office in Paris€1,808.80 gross per monthMeal vouchers50% of your Navigo passCompute resources sized for real training runs, including HPC accessSupervised directly by the Head of AIProcess30' intro call with the Head of AITechnical case study with the Head of AI and the CTOLunch and team meeting at our Saint-Lazare officeKeywordsDeep Learning · Computer Vision · Machine Learning · Artificial Intelligence · Neural Networks · Robotics · Autonomous Vehicles · 3D Vision · Spatio-Temporal · Self-Supervised Learning