Imperial College's NodeSLAM: Object-Level SLAM Lets Robots Plan Directly from a Graph of Objects
AjdDavison · x · 2026-10-07
NodeSLAM from Imperial College London's Dyson Robotic Lab (Andrew Davison's group) flips the usual pipeline: instead of scanning the scene into dense point clouds and fitting objects afterwards, it fits objects directly and uses the object graph itself as the map. Key points: - A class-conditional VAE learns a smooth latent shape space per object category, representing shapes with few parameters; - A novel probabilistic, differentiable rendering engine turns object volumes into depth images with uncertainty, enabling full 3D object reconstruction from one or more RGB-D images; - It's the first object-level SLAM system to jointly optimize object poses, shapes, and camera trajectory. The system rapidly builds a map of learned shape models and immediately uses it for simple but accurate motion planning, with applications in robot grasping, placing, and AR.
More from Embodied
- 2023's ACT + ALOHA was the original physical AI era, says dev — lukas_m_ziegler · 2026-10-07
- EAPN: execution-aligned noise fixes mode switching in asynchronous replanning, 96.7% bimanual success — Di Wu · 2026-10-07
- Two-step flow denoising cuts VLA inference from 61.6ms to 22ms for real-time robot control — Di Wu · 2026-10-07
- CTP: single-pass multimodal robot policy hits 97.25% on LIBERO, cuts inference latency to 75.8ms — Di Wu · 2026-10-07
- Magic-W0: a structured world-action foundation model topping RoboDojo-Sim at 27.10 — Xuhua Chen · 2026-10-07
- RaiSim is back: v2.8.0 brings ultra-fast accurate CPU simulation plus photorealistic rendering — ChongZzZhang · 2026-10-07