2 months ago
London, United KingdomSenior
Responsibilities
- Design and train multimodal world models covering video, state, action, and language.
- Build physically consistent action-conditioned video prediction and dynamics models for contact-rich manipulation and VLA policy pretraining.
- Develop learned-simulator evaluation systems to assess policies, predict deployment success, and expose future behavior for safety and planning.
- Generate synthetic rollouts and counterfactual experience for rare events, cross-platform transfer, sim-to-real transfer, and policy training.
- Establish fidelity metrics and calibration protocols for measuring world-model reliability.
- Build training-data pipelines from fleet telemetry, teleoperation logs, and internet-scale video.
- Run scaling and ablation studies and communicate experimental findings.
- Collaborate with pretraining, reinforcement learning, and manipulation teams to integrate world models into policy training and evaluation.
Requirements
- Track record of training large generative models for video, world modeling, or multimodal applications, with shipped models or published artifacts.
- Deep hands-on experience with diffusion models, autoregressive transformers, latent-variable models, or video prediction.
- Experience with large-scale distributed training, streaming datasets, checkpointing, state management, numerical debugging, and training-instability resolution.
- Strong Python and PyTorch or JAX skills, including kernel profiling, data-loader optimization, and maintainable research-code development.
- Ability to design rigorous evaluations, run honest baselines, and document experiments clearly.
- Experience with robotics or autonomous-driving world models, learned simulators, model-based reinforcement learning, or action-conditioned video models is preferred.
- Familiarity with Isaac Sim, MuJoCo, and sim-to-real considerations is preferred.
- Experience using world models for policy evaluation or synthetic-data generation at scale is preferred.
- Publications at NeurIPS, ICML, ICLR, CoRL, or CVPR, or equivalent open-source contributions, are preferred.
- Experience optimizing generative models for fast inference is preferred.
Benefits
- Competitive equity through stock options.
- 30+ paid days off, including annual leave, UK bank holidays, and company closure days.
- Private healthcare with virtual and in-person care.
- Pension scheme with an 8% total contribution.
- Free daily breakfast, catered lunch, and in-office snacks.
- Opportunity to work with engineers, researchers, and product experts on humanoid robotics and AI.
- Significant ownership, direct access to founding leadership, and influence over product direction from day one.
- The posting states that the role is in-office.
Categories
ML EngineeringRobotics
