Lessons learned from their work on 3D and motion (loving these honest takeaways) #ssad2026

Point trackers for frame to frame motion can easily drift on longer temporal horizons. We can imbue them with memory: enter Track-On #ssad2026

We already have good perception: better geometry with 3DGS, accurate optical flow, tracking on longer horizons. Why not building upon them? #ssad2026

Bringing up a question to which a part of the ECCV crowd this year had strong opinions: do we need intermediate representations? Not clear that the answer should be No #ssad2026

Previously we relied on spatial representations, eg, BEV, that could evolve and predict future stats: see StretchBEV (ECCV 2022) #ssad2026

A frame is rarely enough to get context for driving. Adding the entire history is obviously not feasible.
Compact states for decision making can be a way out. We’ll find out in this talk #ssad2026

A handy list of advice for when you have lots of real data but rare cases are missing #ssad2026

For common cases you do imitation, for rare cases you reward #ssad2026

Different strategies to leverage synthetic data for autonomous driving these days #ssad2026

Real logs and simulation teach different things: real data teaches driving, while simulation teaches safety #ssad2026

We know it pays as you add more, if you generate the right thing. What is the right thing is a though question #ssad2026

This is not SciFi, synthetic data actually works to enhance your training protocols for driving and for robot policies. A list of successful examples from the last year #ssad2026

Automotive players are starting to do well on autonomous miles and disengagement rates.
Long tail dominates the risk and is not obvious. Synthetic data to the rescue? #ssad2026

The Carla OG German Ros talks about Synthetic data workflows to accelerate AV #ssad2026

Can we take inspiration from multi-domain benchmarking in LLMs and come-up with smarter evals for AD across routes of different difficulties? Item Response Theory is a potential solution and used in Alpasim challenge #ssad2026

If you didn’t do it yet, please check the AlpaSim challenge: bigger and more diverse data, closed loop evaluation on SoTA renderers, evaluation run on organizer servers under reasonable compute use, and much more.
The leaderboard and final rules are now up #ssad2026

Enter the NAVSIM suite of benchmarks using lightweight simulation.
NAVSIM has been a super-enabler for progress of e2e driving models in the last 2 years #ssad2026

NAVSIM may not be perfect, but you should check the issues on open-loop evaluation on nuPlan that researchers user before. The paper on “is ego status all you need for open-loop autonomous driving” was eye opening #ssad2026

Common task frameworks then and now:
- 1990s no papers had benchmarks
- 2020s no papers without benchmarks #ssad2026

Getting into a common benchmark culture was not easy, but years after the progress turned out impressive #ssad2026