Open-source SLAM Dunk becomes first VIO to beat Meta's Aria on 3 of 5 LaMAria categories

KyleMorgenstein · x · 2026-11-15

Jacob Van Meter and Ayush Paul released SLAM Dunk, an open-source pipeline for long-horizon egocentric SLAM and the first VIO to outperform Meta's closed-source Aria SLAM on LaMAria, winning the medium, long, and low-light categories; it also took first place on the UZH-FPV leaderboard at submission.

Key findings: a conventional offline bundle adjustment pipeline closes much of the gap between academic and industry systems but plateaus below Aria. The remaining gap comes from standard VIO models omitting two physical phenomena — camera calibration drift and gyroscope intrinsic calibration — which SLAM Dunk explicitly models. Code is available on GitHub (jacobvm04/slam-dunk).

Related event: Open-Source SLAM Dunk Beats Meta's Closed Aria on LaMAria Benchmark(2 posts)→

Original post →

More from Embodied

Embodied channel →