AMB3R-SLAM: Kilometer-Scale Real-Time SLAM on One Consumer GPU, Cutting ATE by 70%
rsasaki0109 · x · 2026-09-30
UCL researchers Hengyi Wang and Lourdes Agapito present AMB3R-SLAM (arXiv 2026), a monocular SLAM system that reconstructs kilometer-scale trajectories over 10k frames in real time on a single consumer GPU.
Key points:
- Training-free and model-agnostic: a lightweight front-end handles low-latency online tracking, while a hierarchical backend progressively enforces local, mid-level, and global consistency via pose graph optimization — no bundle adjustment, so dynamic scenes work out of the box
- One framework extends to stereo, RGB-D, and LiDAR inputs; with LiDAR it reaches sub-meter ATE on KITTI and VBR
- Strong tracking across 9 datasets, reducing absolute trajectory error of prior SOTA by over 70% on VBR and Oxford Spires
More from Embodied
- Flexiv arm localizes touch using only joint torque sensors, no skin sensors — chris_j_paxton · 2026-09-30
- CoRL 2026 to stage scientific debate: do robots need world models? — vincesitzmann · 2026-09-30
- Atlas trains at Hyundai's Georgia Metaplant, targeting 2028 production line deployment — CyberRobooo · 2026-09-30
- Models May Learn Who Generated the Data: Robotic Surgery Study Shows AUC Drops 0.89 to 0.57 — bravo_abad · 2026-09-30
- DoorDash to Let Users Order Chipotle via AI, Delivered by Autonomous Drones — Polymarket · 2026-09-30
- Robotics advice: put an IMU on every rigidbody and an encoder on every pivot — sensormaxxing — _Stocko_ · 2026-09-30