🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Jan @czarnowski' DeepFactors (RA-L 2020) built code ideas into a unified SLAM system where network mono-depth prediction is a factor in a standard probabilistic factor graph. With Tristan Laidlow and Ronnie Clark from the Dyson Robotics Lab. Open source! 46/n x.com/czarnowskij/st…

To me the astounding thing here is that it shows the 2D diffusion model must have learned consistent 3D structure and lighting internally… and this is a very neat way of extracting that 3D info!

Ever wondered how 3D vision can help understand the weather and climate? Check our new work on reconstructing 3D cloud fields. We test our model using data from Raspberry Pi cameras observing a 10km2 area of the sky. Project: cloud-field.github.io w/ Jacob Lin Ed Gryspeerdt

Excited to announce my student’s Sumeet’s @sumeetrn recent work on Multi-Agent LLM Training is now on arXiv! What if a generator, critic, and refinement model could be jointly trained to improve their reasoning performance? ✨🧵 MALT is an early advance leveraging synthetic data

Excited to announce our work on Multi-Agent LLM Training!🚨 MALT is an early advance leveraging synthetic data generation and credit assignment strategies for post-training specialized models solving problems together. Check out the first project from my PhD🧵✨

If you’re interested in LLM reasoning, check out our new paper on multi-agent LLM training (MALT), where we specialise a base model to create a generator (G), verifier (V), and refiner (R) that work together to solve problems. Paper: arxiv.org/pdf/2412.01928 Led by Sumeet Motwani

We trained a computer vision agent that, given a prompt + image, can choose the best specialist model to solve a task or even chain together models for more complex tasks. #CVPR2025 w/ Microsoft AI See more: arxiv.org/pdf/2412.09612

Motion blur typically breaks SLAM/SfM algorithms - but what if blur was actually the key to super-robust motion estimation? In our new work, Image as an IMU, Ronnie Clark and I demonstrate exactly how a single motion-blurred image can be used to our advantage. 🧵1/9

🚨 New preprint on arXiv from OxWearables PIXL @ Oxford ! Can vision-language models (VLMs) help automatically annotate physical activity in large real-world wearable datasets (⌚️+📷, 🇬🇧 + 🇨🇳). 📄 arxiv.org/abs/2505.03374 🧵1/7