7-Paper Research Program · Architecture → Governance → Benchmark

AEROS Project

A systematic research program on runtime architecture, governance, evolution, contracts, federation, and benchmarking for embodied agent systems.

Paper Series BibTeX GitHub

System Architecture

Seven papers, one coherent architecture — from the ECM runtime core to fleet-scale governance evaluation.

P7 — EmbodiedGovBench · Governance Evaluation Framework P6 — Federated Single-Agent Fleet Robot A · Single-Agent Runtime P3 Governance Policy · Watchers · Recovery · Human Override P2 | P4 Governed Evolution Safe Upgrade · Shadow Mode · Rollback · Validation P5 Contracts Interface · Compatibility · Versioning · Release Discipline P1 AEROS Runtime ECM Navigate v2.3 ECM Grasp v1.7 ECM Inspect v3.0 Audit Trail · Execution Trace · Recovery Log Human Supervisor Hₐ Robot B · Single-Agent Runtime P3 Governance Policy · Watchers · Recovery · Human Override P2 | P4 Governed Evolution Safe Upgrade · Shadow Mode · Rollback · Validation P5 Contracts Interface · Compatibility · Versioning · Release Discipline P1 AEROS Runtime ECM Navigate v2.3 ECM Deliver v1.2 Audit Trail · Execution Trace · Recovery Log Human Supervisor Hₑ Shared Registry Task Delegation Fleet Recovery Fleet-Level: Shared ECM Discovery · Trust-Scoped Delegation · Policy Composition · Layered Recovery Fleet Human Supervisor Hᴇ Evaluates: Authorization · Drift · Recovery · Portability · Upgrade Safety · Human Override · Audit Completeness
P1 Runtime
P5 Contracts
P2|P4 Evolution
P3 Governance
P6 Federation
P7 Benchmark

Research Narrative

Seven papers form a complete research program: from defining what an embodied agent runtime is, to how it evolves, how it is governed, how capabilities are contracted, how robots federate, and finally how the field should measure all of this.

1
Architecture
AEROS — What is an embodied agent runtime?
Defines the single-agent runtime and Embodied Capability Module (ECM) abstraction. Capabilities are installable, removable, and governable at runtime.
2
Evolution
How do capabilities grow without breaking the agent?
Capability evolution through ECM versioning, replacement, and regression-aware upgrading — without altering the agent's identity or existing behaviors.
3
Governance
How does the runtime stay controllable during execution?
Runtime governance layer: admission control, policy enforcement, behavioral watchers, structured recovery, and human override protocols.
4
Governed Evolution
When capabilities evolve, how does governance stay valid?
Bridges evolution and governance: gated deployment, shadow mode testing, rollback conditions, and policy compatibility checking across capability versions.
5
Contracts
How do capability packages compose reliably?
Six-dimensional contract model for ECMs: functional signature, behavioral assumptions, resource requirements, permission boundaries, recovery semantics, and version compatibility.
6
Federation
How do multiple robots coordinate without fragmenting each agent?
Fleet coordination through federation across coherent single-agent runtimes: shared capability registries, trust-scoped delegation, layered recovery, and hierarchical human oversight.
7
Benchmark
How should the field measure all of this?
EmbodiedGovBench: seven governance dimensions, six protocol families, single-robot and fleet evaluation tracks. The measurement layer that completes the research program.

Paper Series

All papers share unified terminology, author information, and cross-references.

P1

AEROS: A Runtime Architecture for Embodied Agents with Installable Capability Modules

Under review · 2026
Defines the single-agent runtime and Embodied Capability Module (ECM) abstraction: what an embodied agent runtime is.
arXiv Code (paper) Runtime
P2

Learning Without Losing Identity: Capability Evolution for Embodied Agents

Under review · 2026
How capabilities evolve—grow, iterate, replace—without altering the agent's identity or breaking existing behaviors.
arXiv Code (paper) Runtime
P3

Harnessing Embodied Agents: Runtime Governance for Policy-Constrained Execution

Under review · 2026
Runtime governance layer: admission control, policy enforcement, watcher, recovery, and human override.
arXiv Code (paper) Runtime
P4

Governed Capability Evolution for Embodied Agents

Under review · 2026
When capabilities evolve, how does governance remain continuously valid? Gated deployment, shadow mode, rollback, and policy compatibility.
arXiv Code (paper) Runtime
P5

ECM Contracts: Contract-Aware, Versioned, and Governable Capability Interfaces for Embodied Agents

Under review · 2026
Six-dimensional contract model for ECMs: functional signature, behavioral assumptions, resource requirements, permission boundaries, recovery semantics, and version compatibility.
arXiv Code (paper) Runtime
P6

Federated Single-Agent Robotics: Multi-Robot Coordination Without Intra-Robot Multi-Agent Fragmentation

Under review · 2026
Multi-robot coordination through federation across coherent single-agent runtimes, rather than fragmenting each robot into multiple agents.
arXiv Code (paper) Runtime
P7

EmbodiedGovBench: A Benchmark for Governance, Recovery, and Upgrade Safety in Embodied Agent Systems

Under review · 2026
Seven governance dimensions, six protocol families, single-robot and fleet tracks. The measurement layer of the research program.
arXiv Code (paper) Runtime

Citation

BibTeX entries for all papers in the series.

@article{qin2026aeros, title={AEROS: A Runtime Architecture for Embodied Agents with Installable Capability Modules}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.07039}, year={2026} } @article{qin2026evolution, title={Learning Without Losing Identity: Capability Evolution for Embodied Agents}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.07799}, year={2026} } @article{qin2026governance, title={Harnessing Embodied Agents: Runtime Governance for Policy-Constrained Execution}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.07833}, year={2026} } @article{qin2026governed, title={Governed Capability Evolution for Embodied Agents}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.08059}, year={2026} } @article{qin2026contracts, title={ECM Contracts: Contract-Aware, Versioned, and Governable Capability Interfaces for Embodied Agents}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.13097}, year={2026} } @article{qin2026fsar, title={Federated Single-Agent Robotics: Multi-Robot Coordination Without Intra-Robot Multi-Agent Fragmentation}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.11028}, year={2026} } @article{qin2026embodiedgovbench, title={EmbodiedGovBench: A Benchmark for Governance, Recovery, and Upgrade Safety in Embodied Agent Systems}, author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun}, journal={arXiv preprint arXiv:2604.11174}, year={2026} }

Authors

Xue Qin
Harbin Institute of Technology
School of Software
Simin Luan
Harbin Institute of Technology
School of Computer Science
John See
Heriot-Watt University
Malaysia Campus
Cong Yang*
Soochow University
School of Future Sci. & Eng.
Corresponding author
Zhijun Li*
Harbin Institute of Technology
School of Computer Science
Corresponding author

Implementation

Two repositories serve different purposes.

Repository Purpose Status
s20sc/aeros Paper P1 research prototype. Minimal viable runtime for reproducing paper experiments. Frozen (paper reference code)
s20sc/aeros-runtime Full production runtime. Includes governance engine, evolution engine, fleet coordination, benchmarking framework, provider SDK, and marketplace. 1200+ tests, mypy strict. Active development

The research prototype demonstrates core concepts from Paper 1. The full runtime (aeros-runtime) is the unified implementation of all seven papers: runtime architecture (P1), capability evolution (P2), governance engine (P3), governed evolution (P4), ECM contracts (P5), federation (P6), and EmbodiedGovBench (P7).