Overview of the project and takeaways:
- 87.5% time spent in simulation
- imitation learning was much more efficient than center lane estimation #ssad2026

If you go in the mountains in November you are likely to encounter quite a few adverse weather events #ssad2026

They train the driving policy in imitation and then deployed on the car as it arrived #ssad2026

On the architecture-side they proposed a hybrid-vision approach producing depth and semantic segmentation predictions and waypoints using camera, gps, ego state and driving actions #ssad2026

They designed a driving expert with privileged map information to collect data in Carla and generation imitation learning data #ssad2026

The project had a tight timeline with the car arriving very late in the project.
They used an old version of Carla to train in simulation and plugged in government maps #ssad2026

Can a car drive autonomously in the Catalan Pyrenees using only camera sensing under budget? Aiming high. Let’s find out how it went. #ssad2026

Last talk of #ssad2026: Dirk Hornung on OpenXLA: ready to use end-to-end compiler stack #ssad2026

The one and only Holger Caesar on stage with Autonomous Driving at the crossroads: scaling laws, end to end learning, and the gap between #ssad2026

Antonio Lopez starts day 4 #ssad2026 with Vision-based end-to-end driving: CVC experience.

Finally, a list of further challenges for VLAs down the road: how much language is enough, how to evaluate, how to deploy, how to enforce causal reasoning, how to finetune without crippling, how to fix language collapse #ssad2026

A potential explanation on why LADA works: it breaks down a difficult problem into two simpler ones #ssad2026

Enter LADA - Latent Action Driving Annotations.
This is a low-cost strategy to imitate the function of language from few pairs of observation-language pairs by learning to produce latent actions #ssad2026

Shifting gears to latent action models (LAMs) with a primer on LAMs #ssad2026

Quite impressed that FIVE-VLA does this well with just 640M parameters in total #ssad2026

But there are a few more ingredients to make this work efficiently for driving. Enter FIVE-VLA: high image res, really small vision encoder and LLM, memory encoder and decoder, fewer tokens -> 7.5x more efficient than SimLingo on T4 TPU #ssad2026

4 steps for building a VLA (for driving): model selection, data curation, action-head and multi-stage finetuning #ssad2026

Setting the stage with a brief history of the AD stack and where the future is headed #ssad2026

Puneet Dokania on stage speaking about VLAs for drivong: why, how, and challenges? #ssad2026

Putting all these ideas together into GAIA-4 with impressive usability: interactivity + efficiency + stability #ssad2026