iiinova — where your digital self keeps evolving

I tried to live-tweet as much as possible from the talks. For some, I had to multi-task with other duties and I did not get to do the coverage. But every single talk was top notch without exceptions. Kudos to our fantastic speakers! #ssad2026

000

The two people who actually made this happen: @pdillis.bsky.social and Gabriel Villalonga. Months of work so that four days could look effortless. Thank you both 🙏 #ssad2026

110

🚗 SSAD 2026 is done! Great talks, two packed poster sessions, and an incredible audience that kept asking questions well after the session should have ended. Thank you to our speakers, our sponsors, and the 72 people who made it what it was. We had a blast! #ssad2026

130

Other stats from AV-Atlas: - main cliques of collaborators - per country publication stats - where researchers are in their career in terms of years in. Check it out! #ssad2026

000

In the end Holger presents AV-Atlas, a tool for scanning and connecting papers in autonomous driving. A few interesting stats: - ascending/descending topics - areas of interest/non-interest for industry - must read papers #ssad2026

100

A shout-out to the @kesai.eu team for the collaboration the py123d framework bridging essentially all public AD datasets #ssad2026

100

On to new datasets. Enter the KITScenes Multimodal dataset #ssad2026

100

3D foundation models are still missing #ssad2206

100

Where to go next? Is Computer Vision solved? People ask this all the time. “AI is what does not work” (Dutch politician) Well this is where research goes :) #ssad2026

100

Public accountability is lacking: unclear what info was given to authorities for FSD clearance in NL. Then e2e limits accountability: now e can benchmark planning which needs 100M+ kms to be statistically significant #ssad2026

100

VLAs do not generalize: to new cameras, chain of thought explanations do not reflect the internal cause of the action #ssad2026

100

Diffusion does not generalize: at best synthetic data on point clouds leads to a boost of few percentage points #ssad2026

100

On with pessimism: public dataset are insufficient, open loop does not generalize, novel view synthesis does not generalize (on public dataset), end-to-end needs a babysitter #ssad2026

100

Let’s go to the pessimistic view: accidents on road, cameras are not enough, scaling is limited (under non-diverse data) #ssad2026

100

We start seeing explainability methods, encouraging scaling laws, extension of AD worldwide and Europe. This indeed the optimistic view :) #ssad2026

100

Next step is going to closed loop settings and this is where neural rendering strategies to produce realistic environments helped #ssad2026

100

Then we started to go further and look at planning. Lots of progress but mostly open loop #ssad2026

101

Setting the stage with the first narrative, the optimistic one. We started with increasingly better perception oriented datasets that enhanced perception significantly #ssad2026

100

Participants of #ssad2026, but also folks following from home are invited to IV 2028 in Barcelona

000

Announcing the LAIA dataset: 15 hours of driving in CARLA with labelled human gaze arxiv.org/abs/2607.25570 #ssad2026

100

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

100

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

100

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

100

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

100

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

100

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

100

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

100

It all starts with modular architectures but they soon realized it was a pain to put together and update. Loving these vintage videos with young PhDs testing out things in the wild. By the way they did it all with a handful of PhDs, postdocs and engineers #saad2026

100

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

040

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

140

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

150

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

000

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

100

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

100

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

100

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

100

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

100