Can neural networks reason like pointcloud registration algorithms? Excited to share our latest paper, "NAR-*ICP: Neural Execution of Classical ICP-based Pointcloud Registration Algorithms", with @PetarV_93, @d_d_martini, and @kula78! arxiv.org/abs/2410.11031 More info below 🤖

Having watched ~ an infinite number of @PetarV_93 ‘s tutorials, I can confidently say that this should not be missed. It opens a new door to the way we approach learning. For 🤖: NAR combines algorithms with DL so we don’t have to choose. It just works and it works every time🚀

Amazing!! Dataset coming soon 👀

Our method introduces a more interpretable and efficient robot learning framework. By integrating classical algorithms with deep learning, we combine structured reasoning and logical computations with the adaptability and generalisation capabilities of neural networks. 🚀 4/4

ICP algorithms: ❎require strong initialisation ❎sensitive to large translation errors ❎limited to Cartesian inputs ❎do not generalise NAR-*ICP: ✅superior performance ✅robust to noise/large errors ✅handle abstract input features ✅2-16x faster ✅differentiable 3/4

We extend the Neural Algorithmic Reasoning framework to address a fundamental robotics task. Our results showed that NAR-*ICP consistently outperforms even the algorithms it was trained on, especially in complex, noisy datasets where the traditional methods fail to converge. 2/4

We leverage the intermediate computations of ICP-based algorithms as supervisory signals to iteratively train GNNs to approximate the algorithmic reasoning processes, effectively combining neural networks and classical robotics algorithms for enhanced robot learning. 1/4

Today, @d_d_martini is presenting our OORD paper at the Workshop on Resilient Off-road Autonomy (NT 402). We are currently at the poster session talking about our adventures at the Scottish Highlands. @MRG_Oxford arxiv.org/pdf/2403.02845