Loading interactive logo…
AI training data · Robotics simulation · Policy task authoring · MuJoCo · 2026–present
I build MuJoCo training problems in which AI systems write policies that turn live observations into physical actions. I create the environment, policy interface, reference policy, diagnostics, and scoring that make each problem useful for training.
01 / Simulation work
These diagnostic renders show reference policies operating in the same simulation environments used to evaluate AI-written policies.
The drone surveys a simulated refinery, uses changing gas readings to locate a plume source, and adjusts its search as the wind moves that signal through the environment.
Two cable-linked rovers coordinate so one can anchor while the other climbs difficult terrain, then continue moving as one coupled physical system.
The policy tilts a maze to guide a free-rolling marble through obstacles and settle it inside a small target without controlling the marble directly.
A released card follows a controlled S-curve through three offset gates before settling at the landing checkpoint under simulated crosswind.
02 / How the policies work
Each policy has to interpret the current simulated state, choose bounded actions, and remain stable through disturbance, contact, and recovery. The physical problem changes, but the control loop stays inspectable.
The aerial policy repeatedly turns uncertain gas measurements into a better estimate of the plume source instead of following a fixed search path.
The rovers behave as one cable-coupled system: one creates support while the other moves, then their roles change as the pair advances.
The maze controller accounts for where the marble is headed, not only where it is, and changes behavior as the motion enters a new phase.
The task asks a policy to shape a card's flight through three offset gates in sequence, then complete a controlled landing under simulated crosswind.
03 / What I built
My role covers the complete authoring loop. I design the simulation, define how a policy interacts with it, build a strong reference solution, and create the evaluation evidence used to understand performance.
A simulation world designed around a clear physical objective.
A defined observation-and-action contract for every submitted policy.
A capable baseline used to establish feasibility and a meaningful comparison.
Deterministic scoring and reviewer-facing evidence that make results inspectable.
04 / Reference-policy comparison
Result
I build capable reference policies that establish feasibility and provide a meaningful same-information baseline for evaluating model-generated policies.
This work has moved my AI training from written responses toward complete systems where reasoning has to produce an observable physical result.
These examples are intentionally presented at a high level; implementation and evaluation details remain private.
Next case study
GraphDat