The scalability of multi-user SLAM is fundamentally limited by the constrained network and computational resources. Existing approaches either focus on SLAM for single-user scenarios or overload networks and servers by streaming dense, uniform camera data and treating all users equally. This results in poor pose estimation accuracy or slow updates to multiple users. Our key insight is that the sparsity and heterogeneity of user activity reveal that not all users or frames contribute equally to the shared map. Building on this, we propose SPARC - Proximity-aware Scheduling of AR Mapping and a blur-aware adaptive Cloud-based GenAI sampling method, which together form a cloud-native framework for efficient multi-user SLAM. On the client side, adaptive, context-aware frame transmission selectively forwards high-value frames. On the server side, generative AI (GenAI)-based upsampling reconstructs dense scene features from sparse inputs, while a proximity-aware scheduler prioritizes updates for users with higher drift or critical interactions. Together, these components reduce redundant transmission, improve resource allocation, and enable fairness without sacrificing accuracy. We show through extensive experimentation that our method reduces the latency by 2× to 4× compared to state-of-the-art while maintaining similar or better tracking accuracy. More broadly, this work reimagines SLAM as a cloud-native service, paving the way for scalable, real-time AR/VR applications where many users seamlessly interact in shared environments.

In Mobile Augmented Reality (MAR) applications, rendering virtual objects accurately on the user’s screen relies on knowing the poses of the user and the virtual objects. Typically, Simultaneous Localization and Mapping (SLAM) is used for pose estimation, but SLAM consumes lots of space and time. Many MAR applications such as gaming and navigation do not need continuous mapping, hence Visual Odometry (VO) is developed that provides localization without map generation. This paper evaluates VO’s accuracy compared to SLAM by specifically considering feature points at varying depths which distinguishes indoor and outdoor environments. Our findings suggest VO performs better with distant feature points (i.e., commonly outdoors), while SLAM excels with closer feature points (i.e., typically indoors). These findings inspire us to design APPEAR, an adaptive approach that switches between VO and SLAM as needed. We implemented and evaluated APPEAR on a computer rather than a smartphone because the chosen VO method was not available on a phone. APPEAR demonstrates memory savings of up to 100 MB and a 1.67 x speed boost on medium-sized datasets without losing accuracy. Additionally, while Absolute Trajectory Error (ATE) has been the primary metric for evaluating pose estimation, we argue for more comprehensive quality metrics for MAR applications: accuracy, precision, and recall. We also define selective ATE where ATE is only calculated in the region where the virtual objects are visible.

During disasters when the communication and power infrastructures are unavailable, we can deploy low-power, low-cost, portable wireless nodes and use them to facilitate multi-hop voice communication between the survivors and the rescue team. Similarly, using multi-hop, multicast, we can facilitate communication between the rescue team and multiple survivors, with whom direct communication is not possible. This paper examines a multi-layer adaptive approach to the problem, the voice data captured from the sender are compressed based on the availability of the bandwidth and contention it might cause in the network. We also perform distributed admission control to ensure that the new streams entering the network do not affect the old ones. To evaluate, we implement the ideas on a testbed of 18 Raspberry Pi equipped with Xbee radios. Our experimental results show that voice data can be multicast to at least 6 destinations with acceptable voice quality in this setup.