The RoleWhile our existing perception division handles classical state estimation and VIO, this role is dedicated to the next generation of spatial and temporal intelligence. We are hiring a Research Scientist to architect and train proprietary foundation models from scratch.
Your core mandate is twofold: building our in-house equivalents to cutting-edge 3D reconstruction architectures, and developing highly robust optical flow models tailored for the chaotic domain of egocentric vision. Beyond these core pillars, you will serve as a lead problem-solver for emergent perception challenges as our hardware and downstream robotics needs evolve.
To achieve this,
we can provide a massive, continuous stream of high-quality, proprietary ground-truth data captured by our infrastructure. You will use this data advantage to train networks that surpass current public baselines, owning the complete spatial-temporal perception loop for our data engine.
What You'll Work OnArchitecting Proprietary Spatial Models- Zero-to-One Model Development: Design, implement, and train state-of-the-art feed-forward network and per-scene differential optimization architectures for 3D geometry extraction.
- Large-Scale Distributed Training: Scale multi-view ML architectures across multi-GPU clusters to handle massive, multi-modal spatial datasets.
- Loss & Architecture Innovation: Push the boundaries of current paradigms by developing novel loss functions and attention mechanisms tailored to our specific data distributions.
Egocentric Optical Flow & Temporal Dynamics- Egocentric Motion Modeling: Build and train custom optical flow architectures capable of handling the extreme motion blur, rapid rotations, and sudden occlusions inherent in head-mounted or robot-mounted cameras.
- Dynamic Scene Understanding: Use your flow models to segment dynamic actors, track objects through complex manipulations, and provide motion regularization for downstream action-conditioned world models.
Emergent Perception R&D- Rapid Prototyping: Tackle novel, unmapped AI challenges as they arise. You will rapidly prototype and deploy new models for tasks spanning tracking, segmentation, multi-modal sensor fusion, and beyond.
- Agile Problem Solving: Pivot to resolve sudden algorithmic bottlenecks in the data engine, adapting the latest research to unblock new product capabilities or hardware integrations.
Next-Level Dense Reconstruction- Neural Rendering Integration: Connect the outputs of your foundational models into highly optimized, large-scale dense reconstruction pipelines (3D Gaussian Splatting, NeRFs) to generate photorealistic environments.
What You BringRequired Background- Deep expertise in Deep Learning, 3D Computer Vision, and Temporal/Video Modeling.
- Proven experience training large-scale vision models from scratch, not just running inference or fine-tuning.
- Strong theoretical and practical understanding of modern feed-forward 3D networks and dense motion estimation.
- Mastery of PyTorch and deep learning scaling frameworks.
- Experience handling and curating massive, multi-terabyte image and video datasets for training.
- Comfortable operating in a fast-paced environment where priorities can shift rapidly to capitalize on new research or hardware capabilities.
Strong Signals:- First-author publications in top-tier venues (CVPR, ICCV, ECCV, NeurIPS) focusing on 3D deep learning, optical flow, video generation, or spatial transformers.
- Specific experience working with egocentric video datasets (e.g., Ego4D, Ego-Exo4D) and solving the unique optimization challenges they present.
- Experience writing custom CUDA kernels to accelerate 3D operations, ray marching, or correlation volume computations.
Why This Role?- The Data Advantage: You will have access to a scale and quality of proprietary spatial and temporal ground truth that most academic researchers only dream of.
- Pure R&D & Model Ownership: You are not maintaining legacy systems; you are given a blank slate and the compute resources to build the state-of-the-art.
- High Impact: The spatial priors and motion models you architect will directly define how the next generation of embodied AI agents perceive and move through the physical world.
Warning: Research Scientist positions require hyper-specific expertise. Please limit your applications to
one research role. Applying to multiple Research Scientist positions suggests a lack of focus and may result in the rejection of all submissions. You may, however, apply to other non-research roles alongside your research application.
If you don't check every box above but believe you can do the job, we encourage you to apply - we're looking for capability and trajectory, not a perfect checklist match.
We may use AI-assisted tools to help evaluate applications against the requirements of this role. These tools support our recruiting team in identifying qualified candidates but do not replace human judgment or limit which applications may be reviewed or considered.
