ML Engineer – Embodied Agents

Our client is building AI that helps anyone create playable, editable interactive worlds with real-time feedback, persistence, and controllable behaviour. The embodied agents work is central to making those worlds genuinely interactive and intelligent.

The role:
You will design and train general embodied agents that perceive (vision, depth, language), reason (memory, planning), and act (continuous and discrete control) in interactive environments. This spans both the agent systems themselves and the infrastructure that makes training and evaluation possible at scale.
What you will build:
Agent systems:

Action abstraction and control: keyboard and mouse primitives, macro-actions, hierarchical action spaces
Latency-aware control loops: debouncing, key timing, cancellation, safe interruption, and recovery from stuck states
Perception to action pipelines: screen state modelling from pixels, OCR, and UI structure

Temporal state aggregation: frame stacking, event deltas, change detection, robustness to lag, resolution changes, and UI noise
World modelling for UI environments: latent dynamics over screen states, predictive models for action consequences, counterfactual rollouts for planning and error recovery

Infrastructure:
High-throughput interaction pipelines with parallel rollouts and synchronised input/output capture
Deterministic replay for debugging agent behaviour and regression testing
Rich telemetry: action traces, screen diffs, attention maps, failure tagging
Sandboxed OS/VM environments with input gating and strict policy/host separation
Data engines: trace mining from human and agent play, automatic dataset generation, versioned datasets tied to behavioural deltas

What you will need:
1 to 10 years in relevant technical domains with strong evidence of ability and speed
Demonstrated track record building scalable products, not just research prototypes
Hands-on experience designing and training embodied agents across perception, memory, planning, and action
Strong fundamentals in agent systems design and building systems and infrastructure from scratch
Breadth across disciplines, ideally computer vision and interactive systems rather than narrow specialisation
Genuinely excited to do cutting-edge research and build products at scale

For more information please reach out to [email protected]