High-resolution Vision-Language Models (VLMs) are widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate an excessive number of visual tokens due to the need to encode multiple partitions of a high-resolution image input. Processing such a large number of visual tokens poses significant computational challenges, particularly for resource-constrained commodity GPUs. To address this challenge, we propose High-Resolution Early Dropping (HiRED), a plug-and-play token-dropping method designed to operate within a fixed token budget. HiRED leverages the attention of CLS token in the vision transformer (ViT) to assess the visual content of the image partitions and allocate an optimal token budget for each partition accordingly. The most informative visual tokens from each partition within the allocated budget are then selected and passed to the subsequent Large Language Model (LLM). We showed that HiRED achieves superior accuracy and performance, compared to existing token-dropping methods. Empirically, HiRED-20% (i.e., a 20% token budget) on LLaVA-Next-7B achieves a 4.7x increase in token generation throughput, reduces response latency by 78%, and saves 14% of GPU memory for single inference on an NVIDIA TESLA P40 (24 GB). For larger batch sizes (e.g., 4), HiRED-20% prevents out-of-memory errors by cutting memory usage by 30%, while preserving throughput and latency benefits.

Reliable and timely delivery of data to mobile targets is a challenging problem in mobile ad-hoc networks, because of the opportunistic and unpredictable nature of the problem. Finding a complete series of mobile-to-mobile contacts and forwarding data toward targets within a designated deadline is even more difficult. We leverage unmanned aerial vehicles (UAVs) as message ferries to travel over highly probable rendezvous points for over-the-air delivery to target users. We propose "Rendezvous," an opportunistic yet disciplined data delivery scheme based on trajectory prediction of users and UAV path planning. During the offline learning phase, we identify both temporal and spatial regularities of mobile users from real-world trajectories using sequence-wise clustering, and construct a compact yet well-summarized cluster signature that enables efficient search. In the data delivery phase, given recent movement history of users, UAVs perform their own distributed path planning collaboratively over time. Each UAV finds a chronological sequence of future visiting points where it can make consecutive data delivery efforts to users to overcome the uncertainty in mobility and ensure timely delivery. Real-world trace-driven simulation experiments demonstrate that "Rendezvous" achieves reliable and punctual data delivery using only few UAVs compared with existing algorithms.

In situations where surveillance or communication infrastructure has collapsed, it is important to keep monitoring affected areas. We leverage unmanned aerial vehicles (UAVs) to collect and provide up-to-date on-site information to a data consumer in an efficient way, for later complete yet agile analysis. We propose a distributed dynamic data collection scheme for persistent surveillance and reconnaissance, using a swarm of connected UAVs with two phases of operation: 1) network formation; and 2) UAV traversal of a region of interest. The main task of a UAV is to continuously collect data within its sensing range, while the UAV swarm travels along the calculated paths. When UAVs are newly connected to form a swarm, or disconnected from an already-formed swarm, a formation phase begins. In the formation phase, UAVs become a single group and produce a compact, dynamically alternating formation called DiagonalX to cover broad areas, including boundary parts, in a fair and effective manner. During the traversal phase, each UAV swarm finds a simple yet efficient navigation path based on data freshness to cover sub-areas and continuously obtain up-to-date information evenly throughout the whole region of interest. Simulation experiments confirm that both formation and traversal procedures perform essential tasks in a distributed manner, while maintaining better data freshness than other counterpart algorithms, with a freshness factor of up to 5.77, and reasonable overheads. An additional feature, a dynamically aperiodic formation change, achieves a more stable performance.

We consider the problem of path planning using multiple UAVs as message ferries to deliver delay-sensitive information in a catastrophic disaster scenario. Our main goal is to find the optimal paths of UAVs to maximize the number of nodes that can successfully be serviced within each designated packet deadline. At the same time, we want to reduce total travel time for visiting over a virtual grid topology. We propose a distributed path planning algorithm that determines the next visit grid point based on a weighted sum of travel time and delivery deadline. Together with path planning, we incorporate a task division mechanism that collaboratively distributes the unvisited grid points with other UAVs so that the entire travel time can substantially be reduced. Simulation results demonstrate that our distributed path planning algorithm mixed with task division outperforms all baseline counterpart algorithms in terms of on-time service node rate and total travel time.

In disaster scenarios where communication networks have broken down, it is important to ensure a reliable data delivery from an emergency operations center to the local target nodes within its effective time limit. We propose a hybrid data delivery mechanism that exploits the load-carry-and-delivery by UAVs with a mixture of localized ad-hoc routing over partially connected terrestrial networks. We aim to achieve reliable on-time data delivery to the target nodes, while preserving the low routing cost. Our proposed routing methodology consists of three steps: 1) localized network construction, 2) network probing by UAVs, and 3) localized ad-hoc routing based on a dynamic depth routing tree depending on the data urgency. Here, we present an innovative cost-effective local data sharing structure called localized minimal routing tree that balances with the direct data delivery by UAVs. After the initial network setup and probing procedure, each UAV makes a series of near-optimal decisions of which grid points to visit considering its localized network topology and data urgency. Our time-dependent routing mechanism dynamically decides data recipient nodes to serve more urgent data delivery with a higher priority at a time. The simulation experiments validated our combined path planning and routing approach, achieving 71% higher reliability than the best possible performance by using only the network nodes and consuming 20% lower energy than the UAV-only approach, while maintaining high reliability. Thus, our work makes a strong case for systematically combining the two approaches.