Research Scientist, SLAM & VIO

Mecka•New York, NY

About The Position

Mecka AI is building the data infrastructure layer for robotics and embodied AI. We partner with leading AI labs and robotics companies to deliver high-quality, real-world datasets used to train, evaluate, and deploy robotic systems — where model performance is dictated by data quality. The Role We are hiring a Research Scientist (SLAM & Visual-Inertial Odometry) to build and validate state estimation systems that work in the real world, on messy sensors, under tight compute and reliability constraints. This role is research-heavy but production-minded. You will ship algorithms that survive scale, long runtimes, and operational edge cases. A core part of the role is expertise in Structure-from-Motion (SfM) and scene reconstruction, spanning both feed-forward and optimization-based approaches to produce high-quality 3D representations from real-world capture data.

Requirements

Strong experience in SLAM, VO, or VIO research and development
Demonstrated history of shipped systems and/or publishable research
Deep understanding of nonlinear least squares, factor graphs, filtering and smoothing, and uncertainty estimation
Strong SfM experience, including feed-forward pointmap regression approaches (FAST3R, VGGT, DA3) and per-scene differentiable optimization methods (ACE0, FlowMap, DROID-W)
Practical experience with dense reconstruction systems: NeRF, Gaussian Splatting
Strong C++ skills
Comfortable using Python for research and evaluation workflows
Built systems that run reliably for hours or days in production environments
Deep understanding of real-world sensor failure modes: calibration drift, synchronization failures, rolling shutter artifacts, motion blur, low-light conditions
Experience with GTSAM, Ceres, or similar optimization toolchains
Strong intuition for optimization, numerical methods, and system stability
Experience deploying NeRF or Gaussian Splatting systems at scale

Nice To Haves

Experience with learned front-ends or back-ends: learned features, learned depth estimation, learned relocalization, hybrid classical + ML systems
Experience building offline mapping and large-scale batch optimization systems
Familiarity with embedded or edge deployment constraints
Contributions to or deep familiarity with open-source projects such as MASt3R, gsplat, nerfstudio

Responsibilities

Develop robust monocular VO/VIO pipelines (feature-based and/or learned) with strong failure detection and recovery
Address scale ambiguity with inertial fusion, motion priors, and consistency constraints
Optimize for low latency, bounded memory usage, and stable tracking across challenging lighting conditions, motion blur, rolling shutter effects, and dynamic environments
Build offline reconstruction pipelines for long trajectories
Implement global bundle adjustment (BA) and loop closure at scale
Optimize maps for high-quality trajectories and sparse/dense maps for downstream data products
Design evaluation tooling, including drift decomposition, per-segment error analysis, and systematic bias detection
Implement stereo VO/VIO systems with robust calibration handling (intrinsics, extrinsics, temporal synchronization)
Improve depth reliability across challenging scenes (low texture, repetitive patterns, specular surfaces)
Optimize for stability and long-duration operation
Build relocalization and graceful degradation mechanisms
Develop large-scale mapping and trajectory refinement pipelines using stereo constraints
Implement loop closure and global pose graph optimization
Perform uncertainty-aware optimization
Produce maps that are consistent, repeatable, and operationally useful, accompanied by meaningful quality metrics
Apply and extend state-of-the-art SfM methods across feed-forward pointmap regression and per-scene differentiable optimization paradigms
Focus areas for SfM include fast reconstruction, generalizable scene geometry, multi-view image collections, no per-scene optimization requirements, scene-specific reconstruction, differentiable optimization, and iterative refinement pipelines
Produce high-quality dense reconstructions using NeRF and Gaussian Splatting
Build photorealistic scene representations
Integrate reconstruction outputs into downstream data products such as annotated frames, spatial QA systems, and training signals for embodied AI models
Benchmark reconstruction quality across scenes, sequences, and sensor configurations
Define and enforce reconstruction release criteria
Perform sensor modeling and calibration, including rolling shutter correction, time offset estimation, IMU noise and scale-factor modeling, and temperature-driven drift compensation
Implement robustness engineering techniques such as automatic recovery and reset systems, outlier rejection, and failure diagnostics and debugging workflows
Design evaluation suites, curate failure-case datasets, and define quantitative release gates