A full-stack workshop for the next generation of ubiquitous AI assistants, bridging sensing, decision-making, alignment, and longitudinal evaluation.
A proactive embodied assistant has to do many things at once. Sense the user and environment under uncertainty. Model context well enough to act on. Decide when to interrupt. Choose actions aligned with long-horizon goals, not the immediate task. Evaluate in the wild, without ground truth. Each is a research community of its own.
The layers don't compose. A calibration error in sensing shifts what perception learns, which shifts what decision reasons over, which changes what alignment even means. A misjudged interruption rarely has a single cause. Most work fixes one layer in isolation; the failures that matter cross them.
PEACE 2026 is organized around this seam problem. We want work that bridges layers, work that re-examines one layer in light of the others, and longitudinal studies that surface what the lab cannot.
PEACE is organized at the program level around three axes, with submissions tagged into six finer-grained threads. We particularly welcome work at the interface between two or more threads.
What the assistant perceives, on what hardware, with what uncertainty.
When the assistant acts, how it represents user values, and how those constraints are made legible.
How proactive systems are built, tested, and evaluated in the wild.
Acceptance criteria weight originality, relevance, and discussion potential — submissions that seed productive workshop conversation are preferred over polished work suitable for the main conference. Two awards will be presented at the workshop: a Best Paper Award and a Best Bridging Paper Award. To be eligible for the Best Bridging Paper Award, authors must explicitly articulate the bridging contribution within the body of the paper.
Completed or substantial in-progress work — systems, sensing pipelines, toolkits, datasets, and empirical studies. Excluding references.
Framings, critiques, agendas. Experience reports, negative results, and dataset / toolkit descriptions especially encouraged.
New wearable devices let assistants perceive the world from the user’s point of view. This creates both the need and the opportunity for assistance that does not wait to be asked. An ideal proactive assistant watches the user’s activity as it unfolds, decides when to speak and what to say, and guides users back when they deviate from the expected steps. The user should not have to keep asking “What’s the next step?” or “Did I do it correctly?”. Building such assistants raises hard questions: when to interrupt and when to stay silent, how to track a task in real time, how to extend proactivity beyond vision to sound and other sensors, and how to evaluate behavior that is inherently subjective. This talk reviews our recent work on these questions, from real-time guidance for physical tasks to proactive audio assistance, and shares public datasets and benchmarks for the community. It closes with open challenges and directions for the field.
Seungwhan (Shane) Moon, Meta Reality Labs · 40 min + Q&A
Two papers · 15-min talk + 3-min Q&A each
Four papers · 15-min talk + 3-min Q&A each
Best Paper Award · Best Bridging Paper Award
