Simulation as Essential Infrastructure During Lockdown
With lab access limited, simulation becomes even more critical - accelerating our investment in virtual engineering.
Silver lining of the pandemic, at least for my corner of engineering: simulation work that used to get filed under nice-to-have is now the only way to make progress. We can't access the labs, but we can access simulated ones. I argued for a dedicated simulation team in 2016 on the grounds that it would set our iteration speed. It took a pandemic to make that argument for me.
The Lockdown Acceleration
Before March 2020, simulation supplemented physical testing. Real data collection was primary, and hardware prototyping happened in labs, where it belongs. Since March the relationship has inverted. Simulation has to substitute for physical testing, synthetic data is often the only data we can get, and hardware simulation is what lets design work continue while nobody can touch a bench. Last month I said shipped dev kits and lab hands would keep most development going. They keep the software side going. Anything that touches a sensor design or a bench has moved into simulation.
Virtual Engineering Expansion
We've pushed simulation well beyond training data.
Sensor Prototyping
The old loop was build a prototype sensor, test it in the evaluation lab we built last summer, iterate. The new loop is simulate the sensor characteristics, predict performance, and iterate entirely in software. We model the optical path (ray tracing through the lens systems), the sensor response (photon-to-electron conversion, noise models), and the processing pipeline (depth estimation, feature extraction). The payoff is that we can evaluate a sensor design change in hours instead of months.
System Integration
Same inversion here. We used to integrate components on a bench, find the issues, and debug. Now we simulate the integrated system and try to catch the issues before the hardware exists: timing interactions between sensors, data flow bottlenecks, how failure modes propagate.
Performance Prediction
Profiling used to mean running on target hardware. Now we run cycle-accurate simulation of the target processors. Not perfect, but good enough to catch major issues before silicon.
Infrastructure Investments
We pulled several investments forward: a 3x increase in cloud GPU allocation for the render farm, a license expansion for our optical simulation tools, a bigger asset library (more 3D environments, objects, materials), and better automation around job scheduling, monitoring, and analysis.
Where does simulation break down?
Simulation is not magic. There's a fidelity ceiling - no matter how good the model, reality surprises you. There's a validation gap, which is the awkward one right now: how do you validate a simulation without access to the real thing? Some debugging intuition is tacit knowledge you only build by handling physical hardware. And nothing in a render farm tells you what it's like to actually wear the device.
Hybrid Workflow
When we do get lab access - limited, scheduled - we spend it deliberately. Validation comes first, using physical testing to check what the simulation predicted. Then calibration data, real measurements that feed back into simulation fidelity. Whatever time is left goes to edge cases, the things simulation couldn't anticipate.
Future State
Post-pandemic, I don't expect us to go back to purely physical development. Simulation-first for early design phases, physical validation at maturity milestones, and continuous refinement of simulation accuracy against real measurements. The pandemic forced an evolution we should have made anyway.