SSAD 2026: Synthetic Data Proves Its Worth in AV and Robot Training

At SSAD 2026, synthetic data took center stage for autonomous driving and robotics. Summarizing the sessions, @abursuc reported that synthetic data has moved beyond science fiction: over the past year it has been deployed in AV and robot policy training with visible results.

Confirmed

@abursuc noted that automotive players are improving on AV mileage and disengagement rate metrics, while long-tail scenarios dominate risk and are hard to capture with real data — synthetic data is a key way to fill that gap.
@abursuc surveyed multiple strategies for using synthetic data in AV, with the core insight that real logs and simulation teach different things: real data teaches "how to drive," simulation focuses on safety scenarios.
German Ros, original author of the CARLA simulator and now at Toyota Research Institute, shared practices on accelerating AV development with synthetic data workflows.
Per @abursuc, the conference consensus: the more correct synthetic data you add, the better — but figuring out "what is the right thing to generate" is the real open problem.

Why it matters

The bottleneck for synthetic data is selection and targeting rather than scale, suggesting the industry should prioritize "what to generate" over blindly scaling up.
The division of labor between real data (driving behavior) and simulation (safety scenarios) offers a clear blueprint for hybrid training pipelines.

2026-09-16 ~ 2026-09-16 · 5 related posts

Primary sources

Synthetic data actually works: a year of wins in training driving and robot policies — abursuc · 2026-09-16 18:43 (UTC+08:00)
SSAD 2026: strategies for leveraging synthetic data in autonomous driving — abursuc · 2026-09-16 18:51 (UTC+08:00)
Synthetic data pays when you add more — the hard question is generating the right thing — abursuc · 2026-09-16 18:46 (UTC+08:00)

CARLA veteran Ros shares synthetic data workflows to accelerate AV development — abursuc · 2026-09-16
CARLA Creator German Ros on Synthetic Data Workflows to Accelerate Autonomous Driving — abursuc · 2026-09-16
[source] Synthetic data actually works: a year of wins in training driving and robot policies — abursuc · 2026-09-16
[source] Synthetic data pays when you add more — the hard question is generating the right thing — abursuc · 2026-09-16
[source] SSAD 2026: strategies for leveraging synthetic data in autonomous driving — abursuc · 2026-09-16