Visual SLAM Algorithm Engineer

Overview

We are seeking a Visual SLAM Algorithm Engineer to develop next‑generation visual SLAM systems with online self‑calibration capabilities. The core objective is to simultaneously solve online localization and sensor calibration (intrinsic/extrinsic parameters, IMU‑Camera time offset, IMU bias) during autonomous truck operation. The role blends traditional geometric SLAM (ORB‑SLAM3, VINS‑Mono) with cutting‑edge techniques such as 3D Gaussian Splatting (3DGS) and vision foundation models (MASt3R / VGGT) to build robust multi‑sensor fusion maps, all accelerated on NVIDIA GPUs using CUDA.

Key Responsibilities

Online Localization & Online Calibration System
- Develop tightly‑coupled visual‑inertial‑odometry (VIO) fusion to achieve real‑time localization in unknown environments.
- Design and implement an “online calibration” module that estimates camera‑IMU extrinsics, time offset (td), IMU bias, and scale factor on‑the‑fly, eliminating the need for frequent offline calibrations.
- Leverage MASt3R/VGGT depth and geometry priors as constraints to improve observability in degenerate scenarios (e.g., static or constant‑velocity motion).
Multi‑Sensor Fusion Mapping
- Build globally consistent sparse or semi‑dense maps primarily from visual input, with optional LiDAR assistance.
- Integrate a Gaussian‑SLAM framework that combines real‑time radiance‑field mapping with pose‑graph optimization to generate high‑fidelity renderable maps.
Algorithm Deployment & Acceleration (NVIDIA GPU + CUDA)
- Port dense matrix operations from the SLAM front‑end (feature extraction, matching, optical‑flow tracking) and back‑end optimization to CUDA for speed‑up.
- Accelerate MASt3R/VGGT inference with TensorRT to achieve real‑time dense depth/point‑cloud prediction.
- Optimize differentiable Gaussian rasterisation kernels for 3DGS, targeting ≥30 Hz frame rates.
- Address coupling stability between online calibration and localization, including robustness to sudden parameter changes and parameter‑freezing strategies when observability is insufficient.

Essential Qualifications

Master’s degree or higher in Robotics, Computer Vision, Automation, or a related field.
≥3 years of hands‑on experience in SLAM or VIO with real‑world product deployment.
Deep understanding of SLAM/VIO core concepts (must satisfy both A and B below).

Technical Requirements – Must Satisfy Both A and B

A. SLAM/VIO Core (choose at least one)
- VINS‑Mono: Mastery of source‑level principles, especially initialization, online extrinsic calibration, and marginalisation strategies.
- ORB‑SLAM3: Mastery of IMU fusion, map reuse mechanisms, and multi‑map merging.
B. Acceleration & Deployment (must have)
- Proficient in CUDA C/C++ with solid knowledge of memory models, warp divergence, and bank conflicts.
- Demonstrated CUDA optimisation of at least one of the following: sparse matrix operations, graph optimisation solvers, feature matching, optical flow, or 3DGS rasterisation kernels.

Preferred Technical Skills (the more the better)

Online Calibration: Expertise in observability analysis (e.g., Fisher information matrix) and hands‑on experience with IMU‑Camera extrinsic/time‑delay calibration.
NeRF / 3DGS: Familiarity with Gaussian‑Splatting; experience with Gaussian‑SLAM or similar real‑time radiance‑field SLAM.
Foundation Models: Experience with MASt3R, VGGT and ability to inject their depth/geometry outputs as priors into traditional VIO optimisation.
Programming & Tools: Proficient in C++ (≥C++17) and Python; experienced with an optimisation library such as Ceres, g2o, or GTSAM; familiar with ROS/ROS2 and real‑hardware deployment (drones, robots, AR glasses).

Strong Pluses

Publications in ICRA, IROS, CVPR, or related venues on online calibration, observability analysis, or 3DGS acceleration.
Successful engineering solutions that mitigated calibration drift in long‑duration robot operations.
Experience with multi‑IMU redundant systems for online extrinsic calibration.
Open‑source contributions on GitHub related to CUDA‑accelerated SLAM or 3DGS.
Familiarity with cuSOLVER / cuSPARSE for back‑end optimisation acceleration.