À propos de ce poste Research Engineer/Senior Research Engineer, 3D Auto-Labeling and Data Flywheel chez Plus
Improving the models that drive our trucks depends on a data flywheel: among petabytes of fleet recordings, we curate the moments that matter, label them, train and evaluate onboard models, deploy, and collect again. Labeling is where this loop usually stalls, because human labeling does not scale with the fleet. Auto-labeling closes that gap: offline models and compute, free of the vehicle's real-time limits, draft the labels, and people verify and correct only where it matters. This role focuses on 3D obstacle auto-labeling: producing accurate 3D object labels from curated multi-sensor recordings, at a quality that lets reviewers verify rather than draw and at a cost that scales. The underlying problems are general: fusing cameras and lidar into consistent 3D estimates, keeping objects consistent through time and occlusion, and knowing when a machine label is good enough. The role starts with 3D obstacles and can expand into other parts of the data flywheel over time.
We are open to candidates at either the Research Engineer or Senior Research Engineer level. Level will be determined by experience, technical depth, scope of responsibility and demonstrated impact. You do not need experience with every technology in our stack; we value strong fundamentals, a sense of responsibility and the ability to learn.
Responsibilities
Develop and evaluate methods that turn multi-sensor recordings into accurate 3D object labels offline, using models and compute the vehicle cannot afford in real time, and measure them against reviewed ground truth
Raise label quality and reduce the human review it needs: find where machine labels fail and fix the causes in models, sensor fusion or data; bring state-of-the-art perception and foundation models into offline use and adapt them to our vehicles and sensors
Build and operate the labeling pipelines at fleet scale: batch GPU inference, data extraction and alignment, selection of what goes to human review, annotation-tool integration, monitoring, throughput and cost, and regression tests that catch quality drops before data reaches training
Close the loop with the modeling teams: deliver training-ready labels, feed model failures and long-tail scenarios back into what gets curated and labeled next; at the senior level, lead designs and mentor
Ensure that your work is performed in accordance with the company's Quality Management System (QMS) requirements and contribute to continuous improvement efforts
Required Skills
BS, MS, or PhD in Computer Science, Robotics, Electrical Engineering, or a related technical field, or equivalent practical experience
Hands-on experience developing and evaluating 3D perception methods (detection, tracking, sensor fusion or related) with a framework such as PyTorch, including experimentation, error analysis, principled metrics, built on solid fundamentals in 3D geometry, coordinate frames, and camera and lidar calibration
Strong Python (NumPy, Pandas, SQL) and working C++ where performance matters; experience running large offline workloads with GPU batch inference
Foundations for data-centric machine learning: how training-data composition shapes model behavior, how to evaluate against ground truth, and how to measure and reduce annotation cost
Self-driven with a strong sense of responsibility: a quick learner who takes problems from ambiguous to shipped; for the senior level, a system you were responsible for end to end, design documents others built on, and mentoring
Preferred Skills
Auto-labeling, offline perception, or ground-truth generation for autonomous driving or robotics, including integrations with annotation tools and vendors
Experience improving models through data: active learning and data flywheels, hard-example mining, uncertainty or disagreement signals, dataset curation
Lidar SLAM, structure-from-motion, factor-graph optimization (GTSAM, Ceres), or stereo and monocular depth estimation
Foundation and multimodal models: open-vocabulary detectors, promptable segmentation, vision-language models, embeddings and vector search
Heavy-truck specifics (articulated vehicles, trailer state estimation, multi-platform calibration) or lane and road-geometry labeling pipelines
