2026 Mobisys Papers From the Living Edge Lab
(aka, the 24th Annual International Conference on Mobile Systems, Applications and Services Workshops) occurred in Cambridge, UK in June '26. The Living Edge Lab was there! See our papers below.
We propose a new abstraction called a Correctable Operation (C-Op) for accurate, real-time mobile AI applications. Based on sensed input, a mobile device initiates concurrent execution of both device-optimized local AI and state-of-the-art remote AI. Speculative action is triggered by the faster local result, but may be corrected by a more accurate remote result. We show that this approach can improve the success rate of a time-critical sensing and actuation application, and discuss future research to explore the C-Op abstraction.
A long-standing concern in drone-based visual mapping is the hours-long 3D reconstruction latency and the thousands of high-quality images required by traditional techniques. Existing solutions have to follow a labor-intensive and time-consuming human-in-the-loop, one-pass mapping workflow. Motivated by recent advances in vision foundation models (VFMs) and their potential for fast mapping, we propose SwiftMap, an AI-in-the-loop, iterative mapping paradigm that moves toward fast, autonomous, and fully automatic drone mapping. To fully unlock VFM's capability within the paradigm, SwiftMap identifies the most informative inputs via content-token hierarchical keyframe selection and enriches VFM's implicit outputs with explicit-implicit geometric map evaluation for reliable quality assessment and next-flight planning. These two plug-in modules work hand-in-hand to improve mapping accuracy, robustness, and completeness through multi-round flights and iterative mapping. Real-world experiments demonstrate that SwiftMap achieves mapping accuracy within 2 meters and 98% map completeness, while reducing mapping time from hours to minutes. SwiftMap is modular by design and can readily incorporate future VGGT variants, enabling further gains in mapping accuracy, efficiency, and scalability as VFM-based mappers continue to evolve.