Computer Vision Researcher – VLM

Responsibilities

• Architect Semantic Grounding:
  Lead research into cross-modal grounding that connects 3D spatial features with language embeddings, enabling the LGM to “understand” object relationships and environmental context.
• Scale “Understand” Capabilities:
  Deploy algorithms for continuous semantics, allowing our 3D maps to evolve and improve situational awareness as new ground-level and aerial data is ingested.
• Agentic Frameworks:
  Build the “spatial brain” for Embodied AI, enabling robots, drones and other machines to move beyond simple navigation to mission-level reasoning.
• Multimodal Benchmarking:
  Define standards for measuring “spatial common sense” in LLMs, creating evaluations that test a model’s ability to interpret and operate within complex 3D scenes.
• Technical Mentorship:
  Serve as the technical anchor for the London R&D hub, resolving architectural disagreements and mentoring researchers in the fusion of 3D CV and NLP.
• Collaborative Innovation:
  Partner with product leads to ensure the “Understand” API delivers high business value for enterprise customers in robotics, logistics, and field operations.

• PhD (or equivalent) in Computer Vision, Machine Learning, or Robotics with a focus on Multimodal/Semantic understanding.
• 4+ years of experience in ML research, with a proven track record of shipping models that bridge 3D Vision and Language.
• Expert knowledge of 3D Geometry (SfM, SLAM, VPS) and Transformer-based architectures (VLMs).
• Multiple first-author publications at top-tier venues (CVPR, NeurIPS, ICLR) focusing on VLMs, scene understanding or semantic segmentation.
• Ability to write production-quality research code in PyTorch or JAX and manage large-scale data pipelines.
• Required In-Office Days: 3 days per week
• Experience with Gaussian Splatting or NeRFs for semantic scene representation.
• Background in robotics (ROS) or building agentic systems that interact with physical environments.
• Experience with “open-set” recognition and Zero-Shot learning.