Low-light RGB SLAM holds up only with inertial fusion and global optimization
ucu-autonomous-ugv · hf · 2026-07-23
Low-light RGB SLAM still breaks down unless inertial fusion and global optimization are both present
This project report benchmarks six SLAM/VIO systems — ORB-SLAM3, DSO, Kimera-VIO, OpenVINS, DPVO, and DPV-SLAM — on five LaMARia sequences under varying illumination.
Main findings
- Kimera-VIO is the only system that tracks all five sequences to completion.
- It also has the lowest relative pose error, but its absolute error keeps growing because it lacks loop closure.
- DPVO and DPV-SLAM never lose tracking, yet their absolute drift reaches roughly 100 m in low light.
- Classical monocular pipelines (ORB-SLAM3, DSO) and the filter-based OpenVINS often fail or diverge on harder low-light sequences.
Takeaway
The report argues that RGB-only SLAM remains stable in the dark only when inertial fusion and global optimization are both in the stack. To close the gap further, the authors expect either low-light-specific learned front ends or additional sensing such as LiDAR, depth, or thermal inputs.
More from Embodied
- A wearable tactile controller lets voice-first AI use push-to-talk and review gestures — Real-Command1804 · 2026-07-23
- Robot hands matter less than the intelligence driving them, says GeneralistAI — E0M · 2026-07-23
- NVIDIA Isaac Sim demo moves a forklift mast 2 meters with Codex and MCP — MonaJalal_ · 2026-07-23
- AI glasses may fail socially unless bystanders can trust the camera signal — IronCuk · 2026-07-23
- GEN-1 now supports everything from five-finger hands to specialized robot tools — k7agar · 2026-07-23
- YC-backed Tornyol is testing a 40-gram drone that hunts insects in mid-air — Olivier__OG · 2026-07-23