Christian Wolf
@chriswolfvision
Principal Scientist at Naver Labs Europe, Lead of Spatial AI team. AI for Robotics, Computer Vision, Machine Learning. Austrian in France.
I am starting "The expanse" series. With 10 volumes, that's quite an investment (of time). Let's see where it goes.
Linus Torvalds on LLM usage (in the Linux kernel project). TL;DR: it is useful so let' use it where it helps and let's not make this a religious question.
There are 10km only between these two places, but 5-6°C difference in temperature. Lyon is a city of > 1M people, whereas the other place (where I live) is in hills with vegetation and lots of trees.
If all countries did what Norway does, we would not be in this mess: they have the highest proportion of electric vehicles!! If you don't want them to export their petrol, don't buy it 😉 The burden is with the consumers, they are causing the problems.
To me this is not new and not surprising: (Applied) ML always had a very strong experimental component and finding out what trained models can do is akin to studying biological systems. This has only become more important when data, models and tasks have become complex. Nothing strange about this.
A bot has reposted my post on an (excellent) keynote, commenting on its clarity. Which it obviously hasn't experienced. We live in strange times.
The Lotte World Tower perhaps? I have taken this picture of it .... from the N-Seoul Tower.
I recall this situation more from the Space-Needle in Seattle 🤔 I didn't take a picture from the inside of the N-Seoul Torwer but I only recall wooden floors. Here is a picture from the outside:
Absolutely recommended: walking up to the N-tower at night (or taking the cable car) and taking pictures of the thousands of sky scrapers around you.
Excellent Keynote by @akorba.bsky.social at Cap-Rfiap in Montpellier, France, on machine learning and distances between distributions.
I am sorry that you lost (fan of Canada as well 🇨🇦), but our household supported Morocco yesterday 🇲🇦, for obvious reasons !
git stache Whenever you want to pull new features and don't need your recent changes.
No need to be unhappy in France: it's colder than in Saudi Arabia or than in the Death Valley!
When projected onto a map, the attention maps correlate with scene structure and with affordances related to the scene structure. These maps have been computed on real nav data in a real environment (like ALL results in this paper). More results in the paper: arxiv.org/abs/2606.21216 ECCV 2026 8/8
Restricting the features provided to the policy also decreases the sim2real gap significantly. 7/8
The number of read-out tokens (retained features values per input patch) can be decreased to as low as 1 scalar value per patch. We would like to stress again that in this case the "attention map" (see Figure) is the ONLY information on the scene the policy receives ("hard" interpretability?). 6/8
One of the key features are projection mechanisms (querying the representation) from patch features to the policy input. In our experiments the best and most robust solution are learnable read-out tokens, similar the "Perceiver Resampler" as this is called for VLMs. 5/8
When training these agents taking RGB input it is beneficial to finetune agents having taken privileged information first (Lidar input during RL pre-training). The pre-trained and then finetuned policy seems to learn to extract better visual features from the ViTs. 4/8
The policies take visual (RGB) input, no Lidar. Encoders distilled from multiple heterogenous teachers (DUNE, AMRADIO) seem to perform better. Dino-v3 very surprisingly is failing and we have confirmed this in numerous experiments exploring a large set of hyper-parameters for its integration. 3/8
This is work at @naverlabseurope.bsky.social by - Steeven Janny - Leonid Antsfeld - Christian Wolf arxiv.org/abs/2606.21216 Accepted to ECCV 2026, builds on our real nav agent introduced in CVPR 2024, CVPR 2025. 2/8
#ECCV2026 paper: A scalar per patch from pre-trained ViTs enables fast moving navigation in the real world 966 *REAL* nav episodes (!!) performed by S. Janny with Dino-v3, Dino-v2, DUNE, VC1, AM-RADIO encoders show that patch features can be bottlenecked to 1 value ➡️ affordances emerge. 1/8
This is work at Naver Labs Europe by - Philippe Weinzaepfel (@weinzaepfelp.bsky.social) - Yours truly, - Mert Bulent Sariyildiz (@mbsariyildiz.bsky.social) - Guillaume Bono, - Gianluca Monaci arxiv.org/abs/2606.21562 7/8
We visualize pose estimation accuracy for different sequences lengths (rows) and recentness/age of the query image (evaluated queries are not actual observations but close viewpoints) and provide a big boost compared to recurrent models without distillation. 6/8
We apply this to the robotics task "Mem-RPE" / "Memory conditioned relative pose est.". Similar to map-free localization, the poses of query images need to be determined, but for a MOVING coordinate frame centered on an agent. We beat SOTA recurrent models and are comparable to transformers. 5/8
We distill the fixed sized teacher representation into the recurrent memory. This distills one COMPRESSION mechanism into another one. Teacher memories at different time steps are backpropagated over length-limited sub sequences "segments". 4/8
This is a problem of learning to COMPRESS data: recurrent models need to learn how to retain/throw away information at each time step. A wrong decision is forever. We train a specific bottleneck transformer teacher with access to the obs history but which compresses it into a fixed size repr. 2/8