Remco Royen

I work as an Applied Scientist II at Amazon RIVR, where I sit at the intersection of modern AI and robotics. My mission is to build intelligent systems that translate complex visual data into real-world robotic action, scaling robust autonomous navigation to redefine last-mile delivery.

Backed by a PhD and Postdoctoral tenure in Computer Vision at Vrije Universiteit Brussel (FWO fellowship), I specialized in 3D scene understanding, novel view synthesis, and scalable deep learning. I turned theoretical algorithms into patented, real-world solutions alongside industry partners like Xenomatix and VoxelSensors.

Today, my core expertise spans embodied AI, Vision-Language Models (VLMs), and Vision-Language-Action (VLA) models, bridging the gap between perception and robust, autonomous decision-making in the physical world.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  GitHub

News

Experience

Education

Publications

*Indicates Equal Contribution

The first generalizable semantic segmentation method employing Gaussian Splatting. Our method is not only superior in segmentation quality, but also achieves real-time performance of 27 FPS, marking an astonishing 901 times speedup compared to the SOTA.

A 3D instance segmentation method which simultaneously learns coefficients and prototypes. The obtained prototypes are visualizable and interpretable. Experiments on S3DIS-blocks and PartNet.

A significant improvement in performance over RESSCAL3D: resolution scalable 3D semantic segmentation + a new dataset VX-S3DIS exhibiting resolution scalable point streams with semantic labels.

A 3D instance segmentation method which simultaneously learns coefficients and prototypes. The obtained prototypes are visualizable and interpretable.

1) The presentation of a new highly realistic synthetic traffic dataset, SynthV6D, for 6D pose estimation. 2) A weakly-supervised domain adaptation approach for 6D pose estimation of vehicles.

A fully parallelizable vector-quantized variational autoencoder model (VQVAE) that generates high-quality point clouds in milliseconds.

Resolution-scalable 3D semantic segmentation of point clouds. By enabling the processing of different resolution scales in parallel, RESSCAL3D is 31-62% faster than the non-scalable baseline and allows early predictions.

A GPU-accelerated simulator that enables the generation of high-quality, perfectly labelled data for any Time-of-Flight sensor, including LiDAR.

An improvement over Mono6D by the introduction of point cloud visibility prediction.

A new block merging algorithm suitable for any block-based 3D instance segmentation technique. It significantly and consistently improves the obtained accuracy for all evaluation metrics employed in literature, regardless of the underlying network architecture.

The first deep-learning-based approach to automatic foot measurement extraction from a single 3D scan.

A monocular approach for vehicle pose estimation using vehicle 3D priors provided by vehicle make-and-model recognition methods to estimate the 6D pose.

A new neural network layer. It can be integrated in any feedforward network, allowing quality scalability by design by creating embedded feature sets.

Additional academic activities

• Machine Learning and Big Data Processing - 2021, 2022, 2023, 2024
• Sensors and Microsystem Electronics - 2020

• Mihnea-Bogdan Jurca - Real-Time Generalizable Semantic Segmentation For 3D Gaussian Representations of Radiance Fields - 2024
• Madina Myrzaliyeva - Investigating The Impact Of Climate Change On Weather Regimes Using Dimensionality Reduction With Deep Autoencoders - 2022
• Jasper De Coninck - View Synthesis For Light Field Camera Arrays - 2022