blog
July 30, 2026

Analog Devices, Inc. (ADI) was proud to sponsor ACM MobiSys 2026, held at the University of Cambridge, one of the premier events for mobile systems, edge computing, and emerging physical AI research. Beyond the technical sessions, one of the highlights of the event was the opportunity to engage with researchers, students, and industry leaders working at the intersection of AI and the physical world.
The conversations throughout the conference reinforced a trend we have been observing: the next generation of intelligent systems will not be defined solely by larger models, but by how effectively sensing, computation, communication, and real-world action are integrated into a unified system. While the conference covered a broad range of topics, several presentations stood out for their relevance to ADI's interests in intelligent sensing, edge AI, robotics, and physical intelligence. These were our favorite talks from the event.
Paper: https://arxiv.org/pdf/2412.18695
The talk, a best paper runner-up presented by a Yale University team, addressed the challenge of serving embodied agents like robots and drones with LLMs. Models must produce not only correct results but also arrive within time constraints. Existing engines such as vLLM batch requests first-come, first-served, so an urgent "avoid collision" command needing 200 ms can stall behind a non-urgent "find the hammer" request. The presenters’ key idea builds on the observation that for physical agents, execution is slower than generation: TimelyLLM, which is LLM-agnostic, partitions plan generation into executable segments and reallocates GPU resources to urgent requests, cutting waiting time by up to 84%. The approach reframes serving around real-time deadlines rather than throughput, a dimension we should account for when studying inference for embodied agents. The paper also offers a clean serving-layer baseline to build on without retraining models.
Paper: https://arxiv.org/pdf/2604.13959
The other best paper runner-up was presented by a team from Seoul National University and reframed perception for physical AI around the observation that systems often fail before inference: deployment changes what image sensors capture through motion blur, lighting, or occlusion. Such degraded input makes predictions unreliable regardless of model quality. The authors suggest a shift in where intelligence sits rather than pushing every hard case onto an ever larger neural network. Their solution relies on a bio-inspired architecture composed of three layers: a reflex layer that maintains safe operation, a calibration layer that continuously adjusts sensor parameters such as exposure and ISO, and an inference layer that processes the resulting data.
Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Analog Devices, Inc (ADI) or its subsidiaries. The content is provided for informational purposes only and is not meant to be an endorsement or representation by ADI or any other party. This site may also provide links or references to non-ADI sites and resources. ADI makes no representations, warranties, or other commitments whatsoever about any non-ADI sites or third-party resources that may be referenced, accessible from, or linked to this site.
This echoes a broader paradigm shift toward adaptive sensing rather than pure scaling, and it is a direction that resonates with me: signal lost at capture cannot be recovered downstream, so we should treat sensor control, across vision, audio, and tactile, as a first-class part of the model itself.
Aurchestra, presented by a team from the University of Washington, lets people adjust their acoustic environment in real time, setting separate volume levels for speech, traffic, birds, and other sources, rather than just amplifying or cancelling everything at once. Prior systems isolate a single target sound or speaker; this one handles the full mix, using an extraction network that outputs one stream per active class, robust to as many as five overlapping sounds, plus an interface that lists only the classes currently present. What makes it relevant to our work is where the computation lives: the heavy extraction runs on the hearable device itself, down to a low-power accelerator at tens of milliwatts. Only the detection and user interface run on a smartphone. It is a concrete demonstration that source separation tracking a changing environment can run at the sensor under tight latency limits.
Vec-LUT is about running ternary LLMs, where each weight is just -1, 0, or 1. Presented by a team from Tsinghua and Beijing Jiaotong Universities and the University of Science and Technology of China, the authors’ research relied on ordinary edge CPUs, in some cases barely faster than a dedicated NPU. Since no mainstream processor handles ternary arithmetic natively, lookup-table inference precomputes the small set of possible results once and reads them from a table instead of unpacking weights and multiplying. Their contribution addresses a specific gap: existing lookup methods handle one token at a time well but waste memory bandwidth when processing many at once, since each token reloads its own table. Vec-LUT shares one table across all parallel tokens, cutting the work by a factor of N. This matters for field deployments because parallel token generation is exactly the regime edge AI work depends on. Robots, in particular, rely on multi-step reasoning, long context, and vision-language-action policies that decode in batches, so efficiency gains using ommodity CPUs directly lower the hardware bar.
ConCord, presented by a team from the New Jersey Institute of Technology, focuses on human–robot teaming for exploration: the human is no longer a remote operator but a teammate moving through and acting in the same space. The researchers’ motivation is based on well-grounded problem: a study they conducted showed human navigation and mapping of unknown spaces is inconsistent, prone to distraction and error. ConCord rests on two ideas:
1. Turn signals from a wearable headset worn by the human into information the robot's navigation software can use to understand where its human teammate has been and what they paid attention
2. Using that awareness to guide the robot to cover areas the human did not explore or double back to spots they likely overlooked.
The technique yielded a reported 45% improvement in exploration speed. The most relevant piece is the simulation tool they built for studying humans and robots together. Current tooling fails to treat people as first-class agents, posing a barrier to a closed-loop, human-in-the-loop systems we care about. A remaining challenge is coordinating robot/human teams that move at significantly different speeds, particularly when humans can traverse an environment faster than ground robots.