iiinova — where your digital self keeps evolving

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 17/09/2026

KE:SAI is an open science lab! We talk about our research without secrets. Visit our new website to find some of our recent talks: kesai.eu/talks/

064

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

👥 Work done by Haofei Xu, Daniel Barath, @andreasgeiger.bsky.social, and Marc Pollefeys from ETH Zurich, University of Tübingen, Tübingen AI Center, KE:SAI and Microsoft.

110

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

However, the number of reconstructed scenes today is only a tiny portion of the total amount of available data due to the cost of 3DGS optimization. Methods like ReSplat speed this up, reducing the cost of reconstruction, and can be used in the future to massively scale closed-loop evaluation.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

🚗 3DGS is an essential technique for autonomous driving simulation and evaluation. We, for example, use it in our current AlphaSim challenge, where we evaluate full self-driving stacks on reconstructions of real scenarios in closed loop.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

The recurrent adaptation of ReSplat instead allows it to generalize to novel datasets without retraining from scratch.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

Existing feedforward 3DGS methods were limited to the data domains they were trained on and struggled to generalize to new scenes, a major limitation for using these networks in practice.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

ReSplat works by predicting a depth map at 16x lower resolution, which is then converted to a 3DGS representation via kNN and global attention. This produces an initial representation that can then be iteratively refined by applying the global attention and kNN attention iteratively.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

Compared to standard 3DGS, ReSplat reduces optimization time by a factor of 86x to under a second while improving the visual quality of the result.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

Feedforward 3DGS methods like ReSplat instead predict the Gaussian representation from input images in a single neural network forward pass, eliminating this expensive optimization step.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

⚡ ReSplat is a feedforward reconstruction method targeting few-step novel view synthesis from sparse input images. The standard method, 3D Gaussian Splatting (3DGS), can render target scenes very fast but depends on a slow initial optimization process to produce the Gaussian representation.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 02/09/2026

🌏 KE:SAI is going to ECCV, Europe's premier computer vision conference, and we will contribute 2 main conference papers and 7 invited workshop talks. I will cover some of the contributions over the coming days.

171

Reposted by Bernhard Jaeger

Thomas Dietterich @tdietterich.bsky.social · 23/08/2026

4. An increase in the amount of good work. This is a great problem to have, and the solution is better tools for discovery and recommendation. I rely on Google Scholar and scholar-inbox myself.

1362

Reposted by Bernhard Jaeger

Daphne Cornelisse @daphne-cornelisse.bsky.social · 21/08/2026

What are the biggest bottlenecks to applying reinforcement learning and high-performance simulation in biology? In a new series, I’ll explore this question one domain at a time, starting with weakly electric fish. daphnecornelisse.substack.com/p/training-a...

0304

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

🇬 The paper was written as part of @haofeixu.bsky.social 's internship at Google in collaboration with ETH Zurich, University of Tübingen, Tübingen AI Center, Microsoft, KE:SAI, and the Technical University of Munich.

010

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

I think this general paradigm, learning representations with real data and self-supervised losses and then training task-specific models in simulation, is a key to solving the data bottlenecks that we have in many robotics applications today.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

🤖 Notably, this work also solves a computer vision problem by training with entirely synthetic datasets. Conditioning the model on DinoV3 features is sufficient to close the sim2Real gap.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

As compute becomes more abundant, perhaps we will see more and more fields that have predictive tasks replace deterministic models with generative ones due to their robustness to task ambiguity and label noise.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

However, this method also produces strong results at single-step inference, so the cost is more like a scaling axis at inference to spend more compute to get better results.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

In autonomous driving, diffusion policies have recently become popular for a similar reason: modeling uncertainty in labels. The cost of diffusion policies is, unfortunately, that they are more expensive at inference.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

By modeling geometry estimation as a generative diffusion task (instead of deterministic regression), the model can handle uncertain regions like transparent objects better and predict sharper geometric details.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

👁 Instead of predicting a depth map, the work predicts for each pixel an x,y,z coordinate of where the surface is in the scene, representing geometry as point clouds.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

This work shows that such complicated architectures and loss functions are unnecessary. It proposes to model geometry estimation with a simple diffusion transformer built on a plain ViT.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

3D reconstruction methods often rely on complex hybrid architectures, loss functions, or compressing geometry to latent spaces to leverage diffusion.

100

Bernhard Jaeger @bernhard-jaeger.bsky.social · 18/08/2026

🔬This week's KE:SAI research highlight is PointDit, our ICML paper on monocular geometry estimation. 🧵 📜 arxiv.org/abs/2607.02515

131

Bernhard Jaeger @bernhard-jaeger.bsky.social · 04/08/2026

Additionally, KE:SAI is building its own custom Blackwell cluster in collaboration with Amber and JH-Computers GmbH. This new cluster will be purpose-engineered for the needs of Physical AI research and is scheduled to go live in Q1 2027.

020

Bernhard Jaeger @bernhard-jaeger.bsky.social · 04/08/2026

To build a self-driving car, you need lots of GPUs. KE:SAI is renting a dedicated H100 cluster from @scaleway.com to perform our research and train foundation models. The pictures show one of our many racks.

161

Reposted by Bernhard Jaeger

ELLIS Institute Tübingen @ellisinsttue.bsky.social · 03/08/2026

🚀 We are thrilled to welcome Dr. Kashyap Chitta to the ELLIS Institute Tübingen as our newest Principal Investigator and Hector-Endowed Fellow! Alongside this academic role, he will continue his work with @kesai.eu, a non-profit open-science AI lab he co-founded in May 2026.

institute-tue.ellis.eu

Dr. Kashyap Chitta Joins ELLIS Institute Tübingen as Principal Investigator and Hector-Endowed Fellow

1123

Reposted by Bernhard Jaeger

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 23/07/2026

I am mildly jealous of whoever winds up with this position. Amazing team, amazing goal

1141

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 23/07/2026

KE:SAI is hiring! You are super excited about world models and autonomy? Come and join our team! Let's have impact together! kesai.eu/join/

0164

Reposted by Bernhard Jaeger

Tübingen AI Center @tuebingen-ai.bsky.social · 21/07/2026

Startup news from Tübingen! SOO happy for 🎉 Ontic Labs and Feyer, who have been selected for @SPRIND's Next Frontier AI Challenge. As two of just ten teams across Europe, they'll each receive €3 million to develop their companies. Congratulations to everyone involved! 🚀 tuebingen.ai/news/breakth...

tuebingen.ai

Breakthrough for two University of Tübingen AI startups

Ontic Labs and Feyer each to receive three million euros as they reach second round of Europe-wide competition

1195

Bernhard Jaeger @bernhard-jaeger.bsky.social · 20/07/2026

🚗 KE:SAI got their own car for data collection and, once we have a permit, policy testing! Read more at: kesai.eu/blog/2026-07...

0171

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 17/07/2026

It's getting real! We have been working on self-driving for more than 15 years, and most of that work has happened in simulation. Today that changes: our research vehicle arrived, and we can start testing our ideas in the physical world. kesai.eu/blog/2026-07...

0216

Reposted by Bernhard Jaeger

Alex Turner @turntrout.bsky.social · 15/07/2026

I resigned from Google DeepMind bc it broke its founding promise by selling AI to the military without restrictions against killer robots or mass spying. For months, I worked to stop this but watched powerful ethicists and institutions choose silence. Here's what happened. 🧵

3455118

Reposted by Bernhard Jaeger

Andreas Kirsch @blackhc.bsky.social · 14/07/2026

I work at Google DeepMind. This won't make me popular. But it's all public reporting: 2014: DeepMind reportedly sold to Google on conditions: no military use, independent oversight 2026: a Pentagon contract for "any lawful government purpose" Not one safeguard survived intact

Collage titled "Trust is not Governance — an essay from inside Google DeepMind, written in personal capacity."

A 2014 memorandum, "Conditions of the Acquisition," lists: military applications of DeepMind technology banned; deployment decisions before an independent ethics board (as reported in Mallaby's The Infinity Machine).

Red threads lead to a 2026 U.S. Department of Defense agreement for classified networks reading "any lawful government purpose," with safety settings and filters adjusted at the government's request and no contractor veto (reported terms, The Information, Apr. 2026).

Below: a 2018 AI Principles strip ("no weapons, no surveillance") stamped DROPPED 2025, and a Project Mario 2016–2021 tag stamped ABANDONED.

171272433

Bernhard Jaeger @bernhard-jaeger.bsky.social · 14/07/2026

🗣️ I will give a talk at the Emerging Behaviors for Achieving Robust Autonomy workshop at ECCV 2026 this year in Malmö. emerging-ad.github.io 📜 The workshop is also accepting paper submissions. The deadline is next week: 20.07.2026

010

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 10/07/2026

In this talk from the GenAV Workshop at ICRA 2026, our CTO Kashyap Chitta highlights how open tools, shared benchmarks, and collaborative research can help accelerate progress in autonomous driving. Watch here: www.youtube.com/watch?v=_CYp...

youtube.com

Keshyap Chitta - KE:SAI - Workshop on Generalization in Autonomous Driving at ICRA 2026

What does it take to make autonomous driving research more open, accessible, and scalable? In this talk from the 1st GenAV Workshop at ICRA 2026, Kashyap Chitta explores Democratizing Autonomous…

072

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 08/07/2026

The AlpaSim E2E Closed Loop Challenge 2026 is open for registration and leaderboard access in a soft-open period. Build AV policies for realistic closed-loop simulation, where each policy's own decisions shape future scenes, observations, and interactions.

182

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 03/07/2026

Our team placed 3rd in the KITScenes LongTail Challenge at the CVPR 2026 Workshop on Autonomous Driving! We used zero-shot evaluation with the NVIDIA Alpamayo 1.5 vision-language-action model, with zero fine-tuning or task-specific training.

1103

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 03/07/2026

Our CTO Kashyap gave a very nice talk at ICRA 2026 on "World Models: The Next Frontier of Motion Prediction". The talk sketches the history and also lays out what we plan to do at KE:SAI. Watch here: www.youtube.com/watch?v=sxlh...

youtube.com

Kashyap Chitta: World Models: The Next Frontier of Motion Prediction

Talk given on the 8th Workshop on Long-term Human Motion Prediction (LHMP) at ICRA 2026. Link to the Workshop website: https://motionpredictionicra2026.github.io Talk Abstract: Predicting how the…

063

Reposted by Bernhard Jaeger

Kashyap Chitta @kashyap7x.bsky.social · 27/06/2026

Target points (from a system like Google maps) are how autonomous vehicles navigate long routes. However, make them too precise, and the policy exploits shortcuts. To deploy these systems in real-world conditions with low-res maps and noisy GPS, @kesai.eu is studying how to mitigate this bias.

282

Reposted by Bernhard Jaeger

KE:SAI - Kyutai ELLIS Scalable Autonomous Intelligence @kesai.eu · 28/06/2026

World Engine is a RL post-training simulator for end-to-end driving. In collaboration with industry (and industry-scale data), we present an end-to-end policy post-trained within World Engine that achieves 200 km real driving without intervention. github.com/OpenDriveLab...

062

Reposted by Bernhard Jaeger

Nicklas Hansen @ncklashansen.bsky.social · 26/06/2026

Super excited to share the last paper of my PhD: "Hallucination in World Models is Predictable and Preventable" ✨ We train a 350M-parameter generative world model on a large dataset spanning 210 tasks and show that we can predict *when* hallucination will happen and use that info to fix it! 🧵1/n

15514

Bernhard Jaeger @bernhard-jaeger.bsky.social · 25/06/2026

Tianyu Li, LI CHEN, Caojun Wang, Haochen Liu, Kashyap Chitta, Zhenjie Yang, Yuhang Lu, Naisheng Ye, Yihang Qiu, Yufei Wang, Luoxi Zou, Jiaxin Peng, Jin Pan, Zhaoyu Su, Andrei Bursuc, Shengbo Eben Li, Andreas Geiger, Peng Su, and Hongyang Li

010

Bernhard Jaeger @bernhard-jaeger.bsky.social · 25/06/2026

The project was a large collaboration between 10 different organizations and was executed by:

110

Bernhard Jaeger @bernhard-jaeger.bsky.social · 25/06/2026

One of KE:SAI's research goals is to improve the sample efficiency of the technology such that this level of performance can be reached purely with open-source data (of which we have ~2.000 h currently).

110

Bernhard Jaeger @bernhard-jaeger.bsky.social · 25/06/2026

🏭 The model and data are, unfortunately, not released. Achieving this level of performance still requires using a large amount (80.000 h) of proprietary data.

110