ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Qianggang Ding, Xingyao Wang, Rui Feng, Zhibin Wang, Feixiang Wang, Kelong Mao, Hao Sun, Zhiyao Luo, Jiankai Tang, Lei Li, Jiadong Guo, Minheng Ni, Weicong Lin, Chenxi Yang, Hongxiang Gao, Zhenghua Chen, Yang Bai, Min Wu, Jun Cheng, Huazhu Fu, Dacheng Tao, Bang Liu
cs.AI
2026-08-11
Argues digital and embodied agents both ignore a person's evolving state, and proposes a human-centric closed loop of perception, longitudinal memory, a personal world model, and a consent-bounded intervention policy, plus agency-preservation metrics and a CombodiedBench proposal.
When an older adult misses a medication dose, a software agent can fire another reminder and an embodied robot can bring the pill. Neither models the person: did they forget, are they confused, suffering side effects, or deliberately refusing? The paper calls this a structural gap. Digital agents transform software states, embodied agents transform physical states, and neither makes a person's evolving state the primary object of modeling, intervention, and evaluation. Personal assistants, health agents, AI companions, and adaptive systems each cover a fragment, with shallow, episodic models of the person and metrics that capture single-task success but not whether the person grew or eroded over time.
Combodied Agents fold those fragments into a human-centric closed loop:
The deliberate design choice is to not require an exhaustive human digital twin. The authors state plainly that a high-fidelity replica of a whole person remains aspirational, so the framework substitutes purpose-bounded, uncertainty-aware, user-correctable representations. The chain keeps a strict order, from observation to event to inferred state to predicted trajectory to authorized intervention, with safety constraints embedded at each layer.
There are no experiments. This is a framework paper, and the parts that correspond to results are proposed evaluation schemes, not measured numbers:
This deserves to be stated plainly: ranked by SOTA numbers, this paper does not qualify. Its output is a research agenda and a set of evaluation dimensions.
For practitioners building health, eldercare, companion, or adaptive-learning agents, the most useful contribution is the agency-preservation metric set. It operationalizes whether someone ends up more or less capable after long exposure, instead of fixating on task completion and satisfaction. The PWM loop also gives a more structured answer to whether and how hard to intervene than pure rules.
A position paper with no implementation or validation. The authors concede that a high-fidelity personal twin is still aspirational, and that sensing, model validation, synchronization, compute, privacy, and governance remain open. The PWM reads as a technical wish list, and the edge-native personal-model section is speculative. The eight metric families name dimensions but their measurement protocols are coarse.