At a Glance
- Tasks: Transform static 3D scans into interactive scenes for robots to learn in.
- Company: Cutting-edge spatial AI company with a strong research foundation.
- Benefits: Hybrid work model, competitive salary, and opportunities for professional growth.
- Other info: Collaborate with top researchers and work on groundbreaking technology.
- Why this job: Be a pioneer in dynamic scene research and shape the future of embodied AI.
- Qualifications: PhD in relevant fields and a solid publication record.
The predicted salary is between 59400 - 72600 £ per year.
Research Scientist, Dynamic 3D Scenes for Embodied AI
London, hybrid (three days a week in the office). A founding research hire turning real-world 3D scans into interactive, physically plausible scenes that robots can learn in.
The company
Our client is a spatial AI company building geospatial foundation models of the physical world.
It was spun out of an established consumer technology business with a nine-figure founding investment, and inherited its research team, its 3D mapping and localisation technology, and one of the largest collections of real-world posed imagery anywhere: tens of billions of images captured at street level.
For the past two years the company has operated like a research lab, publishing at the top computer vision venues and building the core models.
It is now standing up new teams to turn that research into products, and embodied AI is one of them.
The research leadership includes senior academics from leading UK universities, and the London team works alongside colleagues in the US.
The role
This is the first research hire dedicated to dynamic scenes in the Embodied AI team.
The team builds the simulated worlds that robot learning depends on.
Today those worlds are static: a scanned kitchen looks right, but its doors, drawers, lifts and appliances do not move.
Your job is to close that gap.
You will own the research direction for turning static 3D captures of real spaces into articulated, interactive scenes, so that software robotics engineers can train and evaluate agents in environments that behave like the real world.
This is a research scientist role, not a robotics engineering role.
You will not be building hardware or writing low-level control.
You will be defining what a dynamic scene representation should be, how to infer it from captured data, and how to make it usable inside simulation.
- What you'll work on
- Inferring articulation from static captures: segmenting movable parts, estimating joint types, axes and motion limits for doors, drawers, cabinets, appliances and similar objects.
- Building scene representations that support motion, on top of Ne RF and 3D Gaussian Splatting reconstructions, so scenes can be both photorealistic and interactive.
- 4D generation and scene dynamics: modelling how objects and scenes change over time, and generating plausible motion where none was captured.
- World models for embodied agents: representations that let a robot predict how a scene will respond when it acts on it.
- Exporting dynamic scenes into simulation, with geometry, articulation and physical properties that a simulator and a learning pipeline can use.
- Using semantics to guide all of the above, for example knowing that a handle implies a hinge or a slider.
- Setting the research agenda for this area, publishing where it makes sense, and working closely with the reconstruction, mapping and robotics teams to ship it.
What we're looking for
The pool of people who have published in this area is small, so the client will consider anyone from a recent Ph D graduate to a technical lead.
What matters is how close your research is to the problem.
- A Ph D in computer vision, graphics, machine learning or robotics, or a near-complete one.
- A publication record at venues such as CVPR, ICCV, ECCV, Neur IPS, SIGGRAPH, ICLR, Co RL or 3DV.
- Research in at least one of: articulated object modelling, dynamic or 4D scene reconstruction and generation, world models, motion representations, or deformable and non-rigid 3D.
- Hands-on experience with neural scene representations, especially Ne RF or 3D Gaussian Splatting.
- Strong Py Torch and the engineering habits to take research code towards production.
- Nice to have
- Semantic or open-vocabulary 3D scene understanding.
- Experience with physics simulators such as Isaac Sim, Mu Jo Co or Habitat, or with sim-to-real pipelines.
- Robotic perception or manipulation, such as pose estimation or hand-object interaction.
- Work with large-scale real-world capture data.
- Experience setting a research direction or mentoring other researchers.
Research Scientist, Dynamic 3D Scenes employer: Trinnovo Group | B Corp™
Trinnovo Group | B Corp™ is an exceptional employer that values creativity and collaboration, making it an ideal place for a Content Adaptation Manager to thrive. With a strong commitment to employee growth, you will have access to professional development opportunities while working in a supportive environment that champions diversity and inclusion. Located in a vibrant market, you'll enjoy the unique advantage of contributing to impactful projects that resonate with local audiences.