Research
The problems we work on.
Aurora sits on a small set of hard technical questions at the intersection of simulation, rendering, and machine learning. We have no publications to list yet — this page describes the problem space honestly, as we see it.
Sim-to-real transfer
When does synthetic experience improve real-world performance?
Transfer is the only metric that matters for synthetic data. We study how rendering fidelity, sensor modeling, and distribution design each contribute to downstream task performance — and where cheap approximations are good enough.
Domain randomization design
Which variations actually matter for a given task?
Randomizing everything wastes compute; randomizing too little overfits models to the simulator. We work on principled, task-aware randomization — treating the variation space itself as something to optimize.
Sensor modeling
How accurate must a virtual sensor be?
Real cameras have noise, rolling shutter, and lens artifacts. LiDAR has dropout, multi-path returns, and intensity falloff. We study which sensor imperfections must be simulated for perception models to transfer.
Failure-driven generation
Can evaluation failures automatically define the next dataset?
A model failure is a coordinate in scenario space. We work on turning that coordinate into a generative neighborhood — thousands of controlled variants around the failure condition — closing the loop between evaluation and training.
Distributed simulation
How do you schedule a million scenario renders efficiently?
Synthetic data generation is an embarrassingly parallel but resource-intensive workload. We work on sharding, caching shared world state, and scheduling heterogeneous GPU workers to keep cost per labeled frame falling.
Working on adjacent problems? We would like to hear from you.