About the role
Help robots understand what happens next.
Dynamic Creatures builds interactive character robots for entertainment and hospitality. We are looking for a hands-on researcher or engineer to develop visual world models that help robots understand their surroundings, anticipate how situations evolve, and use those predictions to guide behaviour.
The role centres on learning from images, video and robot interaction data. You will investigate representations of the physical world, predictive and action-conditioned models, and the connection between visual understanding and robot policies. Experience with vision-language-action models is particularly relevant.
What you will do
- Build and evaluate visual world models that capture scene dynamics, object interactions and the effects of robot actions.
- Develop models and representations from video and robot data, including self-supervised learning and action-conditioned prediction where appropriate.
- Connect learned representations and predictions to planning, policy learning and robot behaviour in collaboration with robotics and controls engineers.
- Own experiments from data collection and curation through model training, evaluation and validation on physical systems.
- Design evaluations that test predictive accuracy, generalisation, robustness and usefulness for downstream robot tasks.
- Investigate failures, improve data and models, and address inference latency and compute constraints.
What we are looking for
- Substantial hands-on experience in world models, video representation learning, predictive modelling or closely related robot-learning research.
- Strong Python and PyTorch skills, with personal ownership of model training and evaluation.
- A strong foundation in computer vision, deep learning and temporal or sequential modelling.
- Evidence of rigorous experimentation: clear baselines, meaningful evaluations, and an ability to explain what worked, what failed and why.
- Interest in translating research into capabilities on real robots, working closely with engineers across perception, autonomy and control.
Particularly useful experience
- Action-conditioned video or latent dynamics models, self-supervised visual representations, or model-based planning.
- Vision-language-action models, imitation learning or reinforcement learning.
- Training on robot demonstrations and interaction data, and deploying learned systems on physical robots.
- Efficient inference, distributed training or large-scale video and robotics datasets.
We welcome relevant experience from industry, academic research or substantial independent work. A PhD is welcome but is not a requirement where equivalent research and engineering depth is demonstrated.
Please apply with your resume and, where available, a paper, repository, project or short demonstration. Tell us what you personally built, how you evaluated it and whether it has run on physical hardware.