Vincent Lannurien
d822cdaa6a
8.4 KiB
title, venue, year, date, tags, paper_count, draft
| title | venue | year | date | tags | paper_count | draft | |||
|---|---|---|---|---|---|---|---|---|---|
| MobiSys 2025 Digest | MobiSys | 2025 | 2025-06-23 |
|
12 | false |
12 papers selected.
Hopter: a Safe, Robust, and Responsive Embedded Operating System
Zhiyao Ma, Guojun Chen, Zhuo Chen 0011, Lin Zhong 0001
TL;DR — Hopter is a new embedded OS that enforces memory safety and real-time responsiveness through a Rust-based task model with cooperative and preemptive scheduling co-designed from the ground up.
Why notable — Building a ground-up safe embedded OS is a long-standing challenge; Hopter addresses it without sacrificing the determinism that IoT and robotics workloads demand, offering a credible alternative to unsafe C-based RTOSes.
WhisperFlow: speech foundation models in real time
Rongxiang Wang, Zhiming Xu 0001, Felix Xiaozhu Lin
TL;DR — WhisperFlow pipelines and partially overlaps Whisper's encode-decode stages so that large speech foundation models can transcribe audio with latency low enough for interactive mobile use.
Why notable — Running large encoder-decoder speech models in real time on mobile hardware was previously impractical; the paper shows that careful pipeline scheduling—not quantization alone—can close this gap, with implications for on-device voice assistants.
ARIA: Optimizing Vision Foundation Model Inference on Heterogeneous Mobile Processors for Augmented Reality
Chanyoung Jung, Jeho Lee, Gunjoong Kim, Jiwon Kim et al.
TL;DR — ARIA partitions and schedules vision foundation model layers across CPU, GPU, and NPU on a mobile SoC to meet the strict latency budget of augmented-reality pipelines.
Why notable — Foundation models are typically too large for AR frame rates; ARIA's heterogeneous mapping strategy achieves real-time throughput without dedicated server offload, making high-quality AR semantics viable on commodity handsets.
You Only Render Once: Enhancing Energy and Computation Efficiency of Mobile Virtual Reality
Xingyu Chen, Xinmin Fang, Shuting Zhang, Xinyu Zhang 0003 et al.
TL;DR — YORO eliminates redundant per-eye rendering in mobile VR by synthesizing one eye's view from the other using a lightweight neural warp, cutting GPU work nearly in half.
Why notable — Stereo rendering is the dominant energy cost in standalone VR headsets; halving it via a neural single-render approach is a surprising result that could significantly extend battery life on devices like Quest.
AutoDroid-V2: Boosting SLM-based GUI Agents via Code Generation
Hao Wen 0004, Shizuo Tian, Borislav Pavlov, Wenjie Du 0004 et al.
TL;DR — AutoDroid-V2 improves on-device GUI automation agents by having a small language model generate executable action code rather than selecting from a fixed action vocabulary, dramatically improving task success rates.
Why notable — Shifting from action classification to code generation is a paradigm change for mobile agents; the paper demonstrates that even small, phone-resident SLMs can outperform larger cloud models on standard Android benchmarks when given the right output format.
EdgeLoRA: An Efficient Multi-Tenant LLM Serving System on Edge Devices
Zheyu Shen, Yexiao He, Ziyao Wang, Yuning Zhang et al.
TL;DR — EdgeLoRA multiplexes many LoRA-adapted LLM variants on a single edge GPU by sharing the frozen base model weights and swapping only the low-rank adapters, enabling multi-tenant LLM inference at the edge.
Why notable — Multi-tenant serving of personalized LLMs on a single edge node is an open systems problem; EdgeLoRA's adapter-swap architecture achieves near-dedicated throughput per tenant while keeping memory footprint proportional to the number of adapters rather than full model copies.
Never Start from Scratch: Expediting On-Device LLM Personalization via Explainable Model Selection
Haoming Wang 0002, Boyuan Yang 0001, Xiangyu Yin 0002, Wei Gao 0006
TL;DR — Rather than fine-tuning from a generic base, this system selects the best pre-existing task-specific model checkpoint as the personalization starting point, guided by an interpretable feature-matching score.
Why notable — The finding that checkpoint selection dominates fine-tuning cost savings—and that an explainable selector can match exhaustive search—challenges the assumption that on-device personalization must always begin from a single canonical base model.
Non-Line-of-Sight 3D Object Reconstruction via mmWave Surface Normal Estimation
Laura Dodds, Tara Boroushaki, Kaichen Zhou, Fadel Adib
TL;DR — By estimating surface normals from mmWave reflections, this system reconstructs the 3D shape of objects hidden around corners without requiring a line-of-sight path.
Why notable — NLOS 3D reconstruction with commodity mmWave hardware is a significant sensing advance; using surface normals rather than time-of-flight alone yields object reconstructions detailed enough to identify object categories, with clear implications for autonomous driving and search-and-rescue.
Toward Spoofing-Resilient and Communication-Integrated MmWave Radar Sensing
Kun Qian 0004, Parth Pathak 0001
TL;DR — This work integrates communication waveforms into mmWave radar so that sensing and data transmission share the same spectrum, while a spoofing-resilience mechanism prevents adversarial injection of false radar echoes.
Why notable — Combining ISAC (integrated sensing and communications) with active spoofing defense in a single mmWave system addresses two open problems at once; the result is particularly relevant as mmWave bands are slated for both 5G and automotive radar use.
Are LoRa Logical Channels Really Orthogonal? Practically Orthogonalizing Massive Logical Channels
Shiming Yu, Ziyue Zhang, Xianjin Xia, Yuanqing Zheng et al.
TL;DR — The paper shows that LoRa's supposedly orthogonal spreading-factor channels have measurable inter-channel interference at scale, then proposes a software-only scheduler that restores near-perfect orthogonality for dense deployments.
Why notable — The result that LoRa orthogonality breaks down in realistic dense networks—and that a pure software fix suffices—is a surprising finding that will directly affect how city-scale IoT networks are planned and managed.
Towards End-to-End Latency Guarantee in MEC Live Video Analytics with App-RAN Mutual Awareness
Juheon Yi, Goodsol Lee, Minkyung Jeong, Seokgyeong Shin et al.
TL;DR — By exposing RAN scheduling state to the MEC video analytics application—and letting the app's feedback influence RAN scheduling—this system achieves end-to-end latency guarantees that neither layer can provide alone.
Why notable — Cross-layer co-design between the RAN and MEC application is rarely demonstrated in a working system; the paper shows that even coarse-grained mutual awareness cuts tail latency by over 50% compared to independent operation.
Unraveling the Missing Link in Low-power Communication: An Autodyning Receiver Architecture that Achieves a Long Range
Pramuka Medaranga Sooriya Patabandige, Rajashekar Reddy Chinthalapani, Wenqing Yan, Prabal Dutta et al.
TL;DR — An autodyning receiver design reuses the transmit oscillator for self-mixing, eliminating a separate LO and achieving orders-of-magnitude better sensitivity than prior backscatter receivers without added hardware cost.
Why notable — Long-range backscatter has been a persistent gap between battery-free IoT and practical deployment; this architecture achieves kilometer-scale range on microwatts of harvested energy, a result that could unlock new classes of batteryless sensors.