Scalix: Uncertainty-Aware Monocular SLAM Recovers Metric Scale in Real Time
zhenjun_zhao · x · 2026-08-19
arXiv paper Scalix: Uncertainty-Aware Scale-Consistent Monocular SLAM from Stefan Leutenegger's group and collaborators.
- Problem: Monocular SLAM suffers inherent scale ambiguity; visual-inertial fixes fail under constant-velocity motion, and learned depth maps are noisy and scale-inconsistent across frames.
- Method: Scalix integrates learned depth cues into a probabilistic factor graph, augmenting monocular depth models with per-pixel depth uncertainty and per-frame scale uncertainty, treating scale predictions as independent measurements improved by multi-view associations.
- Results: State-of-the-art on both metric and up-to-scale benchmarks in large-scale indoor/outdoor environments, in real time, with strong generalization.
More from Embodied
- DIY hardware milestone: power, display, and joystick sorted — pramodk73 · 2026-08-20
- New robots shift the 'humanoid' definition, leaving Cobot's Proxie in question — BradPorter_ · 2026-08-20
- Robotics Form Factor Debate: Humanoids vs. Specialized Robots Can Coexist — lukas_m_ziegler · 2026-08-20
- What Local LLMs Are RTX 5080 Owners Running? Qwen and Gemma Quants Shared — Techniboy · 2026-08-20
- HUMAIN and Qualcomm build AI PC with agentic OS designed around intent, not apps — aziz4ai · 2026-08-20
- Agents Should Learn to Actuate Anything in the Physical World — cephaloform · 2026-08-20