good morning

I've been a Tamiya RC fan since owning and loving a Grasshopper in the 80s, and more recently built one or their RC kits with each of my kids (Hornet, LunchBox and Midnight Pumpkin!)

Live Demo of MASt3R-SLAM this afternoon in the demo session at #CVPR2025.

Full draft of the SLAM Handbook now released --- available as a free PDF, with a printed version coming soon. Now including Part 3, "From SLAM to Spatial AI" (I knew it would catch on eventually), with contributions from @hidematsu82.bsky.social and me. #SpatialAI github.com/SLAM-Handboo...

Great diagram.

It's 3PM. Exactly 40 years ago the first ARM processor was powered up. It worked. To mark the occasion, a bottle of champagne was opened and the time and date was written on the bottle - by Sophie Wilson IIRC. I hope today another bottle is opened to celebrate the anniversary.

Great look at back at 2D scrolling methods in games.

Slow motion Galton Board. Love that Gaussian!

We were fortunate to have Jianyuan visit us at Imperial yesterday, and he gave a brilliant and thoughtful talk about 3D geometry and learning. Maybe every 6--12 months or so there is a paper in my area which really seems like something new and important, and VGGT is one of those!

Should be good!

Remember: the route to Spatial AI is through SLAM. Key research topics are representations for real-time fusion of measurements and priors into persistent semantic/metric scene models; and algorithm/hardware co-design for 1000X efficiency. #SpatialAI arxiv.org/abs/1803.11288

On the tram station video, VGGT fails entirely, whereas MASt3R-SLAM had done a more reasonable (albeit also not stellar) job. bsky.app/profile/chri...

Visual Geometry Grounded Transformer (VGGT) predicts cameras, point maps, depth maps, and point tracks for up to hundreds of images in less than a second on a H100 GPU. github.com/facebookrese...

Cool to see this.

If you like word puzzles, Bracket City is really cool: bracket.city The clues are cleverly nested in a way that looks like Lisp code! (Via the @recomendo.bsky.social weekly newsletter.)

Many interesting tests of MASt3R-SLAM on prerecorded videos.

Aria Gen 2 is very impressive, with fully onboard SLAM and various other perception all within a 75g device with an hours-long battery life. Processing on custom silicon. Congrats to the Reality Labs team.

Shawshank Prison reconstructed with MASt3R-SLAM... from one of the best films ever.

I remember seeing this drone video a few years ago and thinking "we'll never run SLAM on that".... but here it is, complete with dense reconstruction (single camera, unknown calibration, no IMU). MASt3R-SLAM is absurdly robust.

Open source code now available MASt3R-SLAM: the best dense visual SLAM system I've ever seen. Real-time and monocular, and easy to run with a live camera or on videos without needing to know the camera calibration. Brilliant work from Eric and Riku.

Wow I'd like to try to understand this. Something for @tinytapeout.com ?

Looks like a great workshop, happening in Malaga this July.

Gradually unfollowing all of the namesakes I picked up from that automatic X to Bluesky converter tool!

Forest Wanderings. 🌳

Not one VC would ever fund a startup to do the kind of hardcore optimization work that DeepSeek did. Every VC firm should be asking themselves why.

In other news, IQ is on Bsky now and you should follow him.

tree house

Gaussian Belief Propagation really going mainstream now... www.rollingstone.com/music/music-...

I haven't tried one yet but this is a very thoughtful long review from someone who's been using Waymos a lot.

Yes, software for robotics (and vision) comes and goes and in the end you'll have to build it yourself if you want to really push forwards.

I'm recruiting a new PhD student on Gaussian Belief Propagation for Spatial AI. For this particular opportunity: 1. The funding I have is only available to a student who qualifies for "home fees" in England; 2. I'm looking for someone who can start soon, ideally April 2025.

Nice analogy. I'm wondering what the future of scientific publishing should be, given that LLMs are increasingly good at automatically going from full paper to summary, or vice versa. Surely we could be publishing new ideas in the "sketch" form only, but I can't quite imagine exactly what that is.

I was surprised that on Strava my best average speed for 10km was a bit faster than for 5 miles (which is shorter, about 8km). Strava automatically finds the fastest segments from any ride you've recorded. Normally of course you see a faster average speed for shorter distances. 1/n

Excellent post about the recent OpenAI o3 results on ARC (& other benchmarks). I don't know how @natolambert.bsky.social manages to write these so quickly! I highly recommend his newsletter. www.interconnects.ai/p/openais-o3... I am (more slowly) writing my own take on all this, coming soon.

One of the landmark games of all time for me that I spent hours playing on my friend's BBC micro. 3D vector graphics, trading, upgrades and thousands of procedurally generated planets all in 32k, incredible.

Very nice computer vision nostalgia!

MASt3R-SLAM: easily the most robust dense monocular SLAM system I've ever seen, and also very accurate. Real-time, RGB only (no IMU), handles unknown focal length and zoom. More new videos coming soon @ericdexheimer.bsky.social @rmurai0610.bsky.social. Paper and more: edexheim.github.io/mast3r-slam

This is a must-watch talk from the guy that made so much of deep learning progress actually happen.

I'd love to see all of this talk (public link?) No we are not running out of "data", but human-written text is unique data which has been supervised by intelligence (it's ground truth for general intelligence, or at least part of it), and very different therefore from passively captured video.

It's weird and interesting to play a bit with this.

Nice explanation of information entropy.

This looks great!

This is so cool!

Fit-NGP: millimetre accurate 3D object pose estimation from RGB images only via Instant NGP. What can we do in manipulation with this level of accuracy? From Marwan and Ignacio from the Dyson Robotics Lab, Imperial College, ICRA24. marwan99.github.io/Fit-NGP/ youtu.be/KQ7yH_em3Qg?...

Reposting this post of mine from 1.5 years ago... I still find this fascinating vitalik.eth.limo/general/2023...

Nice article. Building LLMs is probably like running an airline, a tough business to be in.

For my first post on Bluesky, this recent talk I did at the recent BMVA one day meeting on World Models is a good summary of my work on Computer Vision, Robotics and SLAM, and my thoughts on a bigger picture of #SpatialAI. youtu.be/NLnPG95vNhQ?...