Frontier AI Research.
Empirical Rigor & Benchmarks.
Read peer-reviewed publications and technical reports from our AI research team advancing active learning, 3D sensor fusion, and multimodal ground-truth validation.
Active Human-in-the-Loop Feedback for 3D LiDAR Object Detection in Dense Urban Adverse Weather
Dr. Elena Rostova, Marcus Vance, Sarah Jenkins, PhD, Engai Research Lab
We introduce a closed-loop human-in-the-loop active learning architecture for 3D LiDAR perception. By prioritizing temporal uncertainty across dense 128-beam point cloud streams, our framework reduces required human adjudication frames by 68% while achieving a 4.2% mAP increase on long-tail rain and snow occlusions.
Spatial Anchoring: Reducing Hallucinations in Vision-Language Foundation Models via Precision Coordinate Supervision
Kavita Rao, Julian Thorne, Alexandre Dubois, Engai AI Lab
Current Vision-Language Models frequently hallucinate fine-grained geometric spatial relationships in complex scenes. We propose Spatial Anchoring, a grounding methodology utilizing verified polygon coordinate priors during instruction tuning. On the RefCOCOg benchmark, this eliminates 74% of geometric hallucinations.
Dual-Specialist Adjudication Protocol for Sub-Millimeter Volumetric Lesion Delineation
Dr. Robert Chen, MD, Amina El-Sayed, Engai Clinical AI Group
Accurate segmentation of rare pathology in abdominal and thoracic CT/MRI scans requires expert inter-annotator consensus. We propose a double-blind adjudication protocol with Bayesian confidence weighting that improves ground truth consistency across multi-institution radiology datasets.
Robust 6-DoF Object Pose Estimation Under Severe Occlusion in Automated Logistics
Tatsuya Mori, Claire Dupont, Engai Robotics Team
Robotic warehouse manipulators face significant challenges when picking tightly packed objects. We demonstrate a multi-modal RGB-D data synthesis and annotation pipeline that yields extreme precision for 6-DoF grasp bounding even when 60% of the target object is occluded.
Stay Ahead in Frontier AI Ground Truth.
Bi-weekly technical postmortems, empirical 3D perception benchmarks, and multimodal dataset releases delivered straight to your inbox.