Engineering Lead, Simulation Environments (MuJoCo)

The role

Steering the frontier toward open, beneficial AGI will demand open multi-agent evaluations across many domains.

Minecraft proved that agents can perceive a 3D world, act through a skill library, and be scored by a deterministic referee. MuJoCo introduces real physics, so we can test transferability: does an agent that ranks high on planning and coordination in Minecraft keep that rank when the world has mass, friction, and continuous contact?

You own this environment end to end, from the containerized physics harness and developer tooling to live leaderboard scenarios, and you scale it.

What you will do

• Build the MuJoCo environment against Kradle's environment contract: manifest, state initialization, observations, action dispatch, and deterministic referee scoring.
• Design the developer tooling and templates so anyone can author their own MuJoCo scenes, scenarios, and scoring functions.
• Design token-efficient observation payloads and skill abstractions that make a physics scene legible to LLMs, vision-language models, and world models.
• Build the headless serialization and rendering pipeline so MuJoCo rollouts stream in real time and play back in our web viewer.
• Run the transfer study: same agents on Minecraft and MuJoCo, measure rank-order stability across frontier models, publish the methodology.

Sample projects

• Stand up a multi-step tabletop manipulation scenario from scratch and ship it to a live leaderboard.
• Compress a multi-contact scene into an observation payload that fits a tight token budget without losing spatial grounding.
• Build a two-agent scenario where agents negotiate, coordinate, or compete over physical resources in the same continuous space.
• Teach Microduck an interesting new skill.

Must-haves

• Production systems in Python, plus TypeScript, Go, or C/C++.
• Hands-on MuJoCo, Isaac Sim, PyBullet, Drake, or Gazebo.
• Reproducible, containerized, stateful workloads on GCP or AWS: Kubernetes, queues, object storage.
• Measurement rigor: statistical noise, benchmark integrity, data leakage. You care whether a metric moved because of the model or a simulator artifact.
• Player-coach: has led or mentored engineers and wants to write production code every week.
• High agency: ships fast, low ego, clear async communication.

Nice to have

• Robotics: real hardware, control stacks, or teleoperation.
• Embodied AI: manipulation, locomotion, sim-to-real, language-conditioned policies.
• Screen capture and input plumbing: streaming a live screen, driving apps through keyboard and mouse, or building computer-use harnesses.
• Sandboxing untrusted code and hardening LLM execution loops.
• 3D visualization, WebGL, or headless rendering pipelines.
• Open-source benchmarks or publications in evaluation or RL.