ADMIRE Research Group

Group photo ·

ADMIRE (Adaptive Distributed Multi-agent Intelligence via Recursive Evolution) is a research group at the State Key Laboratory of Complex and Critical Software Environment at Beihang University. The group is led by Prof. Wenjun Wu.

We study adaptive multi-agent systems in complex environments, focusing on how agents remember, learn, reason, collaborate and evolve. Our research covers software infrastructure for agent coordination and controlled recursive self-improvement, embodied swarm intelligence for drones and robots, and scientific discovery through collaboration among theory, computation and experiment agents.

News

–

–

–

–

–

Video Highlights

Multi-agent Cooperative Control Platform
UAV Navigation & Swarm Experiments
All demos

About the Lab

ADMIRE is based at Beihang University’s State Key Laboratory of Complex and Critical Software Environment.

Research Areas

Built on the RSI foundation and governance of the software space, the group connects the physical/embodied and scientific-discovery spaces, developing key technologies and system frameworks for adaptive agent collectives in complex environments.

Software Foundation and governance for agent collectives

Multi-agent service composition architecture

Coordinate agents through Agentic Mesh and support controlled recursive self-improvement.

Research details

A unified runtime and governance foundation (Agentic Mesh) for agents and services across the software, embodied and scientific-discovery spaces, with controlled recursive self-improvement (RSI) through isolated experiments, independent evaluation and progressive release.

  • Agentic Mesh governance: runtime, lifecycle, gateway, telemetry, memory and orchestration for agent and service collectives.
  • Controlled self-improvement: candidates are generated and tested in isolated sandboxes, filtered by constraints, independently re-evaluated, then released progressively with rollback.
  • Domain agents: service composition, microservice root-cause analysis and AIOps, evolving as typical applications on the foundation.
Agentic Mesh RSI foundation RCA Agents

Physical Systems Embodied collective intelligence

Multi-agent Cooperative Control Platform

Study coordinated control, path planning and validation on physical drones and robots.

Research details

For UAVs, robots and unmanned systems, we combine high-level LLM cognition with multi-agent reinforcement learning, physical priors, safety constraints and real-world feedback.

  • Stable control, team consensus, path planning and spatial reasoning enable deployable multi-robot collaboration.
  • Symmetry, potential fields and Lyapunov stability improve training efficiency and safety.
  • LLM agents generate rewards, organize training, package models and support sim-to-real transfer.
UAV Swarm MARL Sim-to-Real

Scientific Discovery Agentic Scientist

RGP framework for symbolic regression

Bring theory, computing and experiment agents together to support research automation.

Research details

Scientific problems become verifiable agent workflows, connecting theory, computation, experiments and feedback for scientific computing, AI4Sci and laboratory automation.

  • Theory, computing and experiment agents collaborate on literature, hypotheses, simulations and experimental design.
  • For PDE solvers and fluid simulation, we connect retrieval, planning, code generation, execution and correction.
  • ScienceOS, intelligent computing and robotic/multimodal laboratory equipment support reproducible research automation.
ScienceOS AI4Sci PDE Agents

State Key Lab of Complex and Critical Software Environment

Addressing major national needs, the laboratory studies critical software for large complex equipment and space–air–ground real-time systems, focusing on development efficiency, quality assurance and continual evolution.

Research spans complex-system mechanisms and software theory, collective intelligence and collective software engineering, software trustworthiness and quality assessment, and runtime support and continual evolution.

People

Prof. Wenjun Wu

Professor and doctoral supervisor at Beihang University’s Institute of Artificial Intelligence, leading research in foundation models, multi-agent systems, swarm intelligence and embodied AI.

  • Deputy head, National AI Standardization General Group
  • Head, Model and Algorithm Standards Research Group, AI Subcommittee of the National IT Standardization Committee
  • Expert, National Science and Technology Innovation 2030 New-Generation AI Major Project
  • Expert, Intelligent Science and Technology Discipline Panel, Academic Degrees Committee of the State Council

Group photos

Research Outputs

Research outputs by group members, including work co-authored by Prof. Wenjun Wu. Wenjun Wu’s DBLP bibliography

Papers

Software space: foundation and governance

On the governance foundation for agents and services, we study service composition, workflow scheduling, software quality and AIOps, covering the full loop of runtime sensing, fault localization, collaborative repair and experience reuse.

Embodied Collective Intelligence

Structural priors, reinforcement learning, neural algorithmic reasoning and language models form deployable perception–decision loops for UAVs and robots.

Agentic Scientist for Scientific Discovery

Knowledge augmentation, graph learning and reasoning training help scientific agents understand problems, organize computation and produce verifiable reasoning.

Publication Index

2026

REMIX: Reusable Experience Memory for Adaptive Web Service Composition

Reusable experience memory adapts service-composition agents to changing tasks and environments.

WISE 2026 · Accepted
2026

FLARE: Causal Fault Localization and Minimal Repair for Web Service Composition with LLM Agents

LLM agents locate causal faults in service composition and generate minimal repairs.

WISE 2026 · Accepted
2024

ELAKT: Enhancing Locality for Attentive Knowledge Tracing

Yanjun Pu, Fang Liu, Rongye Shi, Haitao Yuan, Ruibo Chen, Tianhao Peng, Wenjun Wu

Self-attention and causal convolutions aggregate and smooth student knowledge states; prediction correction reduces behavioral noise for knowledge tracing.

ACM Transactions on Information Systems
2020

Learning to Handle Exceptions

Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, Yanjun Pu, Xudong Liu

Neural networks learn Java exception-handling patterns, predict try-block locations and generate catch code.

ASE 2020
2017

Crowd intelligence in AI 2.0 era

Wei Li, Wenjun Wu, Huaimin Wang, Xueqi Cheng, Huajun Chen, Zhi-Hua Zhou, Rong Ding

Frontiers of Information Technology & Electronic Engineering

Selected Papers & Architectures

Books & Chapters

Chapters

2003

Peer‐To‐Peer Grids

Geoffrey Fox, Dennis B. Gannon, Sunghoon Ko, Sangmi Lee Pallickara, Shrideep Pallickara, Marlon Pierce, Xiaohong Qiu, Xi Rao, Ahmet Uyar, Minjun Wang, Wenjun Wu

Wiley · Book chapter

Patents

Fault Localization in Microservice Systems

一种故障定位方法、装置、设备及可读存储介质

Patent grant · CN114579346B

吕海、吴文峻、郭洋、徐敏、于卓、许畅、吴跃、于笑明、赵煜阳、郝艳亚

Assignee at grant · 北京航空航天大学、北京中电普华信息技术有限公司、国网安徽省电力有限公司信息通信分公司

Power System Work Ticket Generation

电力工作票构建方法和装置

Patent grant · CN114841519B

吕海、吴文峻、欧阳红、王思宁、余沫、郝庆利、于卓、于笑明、郝艳亚、宋亚琼、廖星创

Assignee at grant · 北京中电普华信息技术有限公司、北京航空航天大学、重庆广汇供电服务有限责任公司信息通信分公司

Edge Data Processing for Power IoT Systems

基于边缘计算的电力物联网数据处理方法和装置

Patent grant · CN114567568B

吕海、吴文峻、刘晗、雷晓萍、王心妍、于卓、闫丽景、李帅、于笑明、郝艳亚、廖星创

Assignee at grant · 北京中电普华信息技术有限公司、北京航空航天大学、国网河南省电力公司信息通信公司

Faulty Node Identification

目标故障节点的确定方法、装置、设备及存储介质

Published patent application · CN117749607A

李君婷、纪鑫、张奎、谭晶、项楠、张文雅、吴文峻、郝文科、陈睿博、董林啸

Applicant at publication · 国家电网有限公司大数据中心、北京航空航天大学、国网江苏省电力有限公司南京供电分公司、国家电网有限公司、国网江苏省电力有限公司

Anomaly Detection for Microservices

一种微服务异常检测方法、装置、设备及存储介质

Published patent application · CN117707899A

陈屹婷、谭晶、项楠、赵天成、李君婷、吴文峻、马永、张靖、王青亮、张文雅、陈睿博

Applicant at publication · 国家电网有限公司大数据中心、国网江苏省电力有限公司南京供电分公司、国网江苏省电力有限公司、北京航空航天大学、国家电网有限公司、国网安徽省电力有限公司信息通信分公司

Active Learning for Deep Answer Recommendation

一种面向深度答案推荐模型的主动学习方法

Patent grant · CN112487172B

吴文峻、汪群博、赵永驰、辛治旻

Assignee at grant · Beihang University

BPMN-Based Microservice Workflow Import

一种基于BPMN的微服务工作流程导入方法

Patent grant · CN112486998B

吴文峻、于笑明、廖星创

Assignee at grant · Beihang University

Attention-Based API Misuse Detection

一种基于注意力机制的API误用检测方法及系统

Patent grant · CN114741322B
ZL 2022 1 0566644.1

吴文峻、杨京波

Assignee at grant · Beihang University

Selection of Workflow Management Nodes for Microservices

基于微服务的工作流管理节点的确定方法和装置

Published patent application · CN114218048A

吕海、吴文峻、李海涛、杨竞奋、于卓、马军伟、于笑明、张琪、廖星创、张波

Applicant at publication · 北京中电普华信息技术有限公司、北京航空航天大学、国网山西省电力公司信息通信分公司、山西益通电网保护自动化有限责任公司

Business Process Recommendation for BPM Platforms

基于BPM平台进行业务流程推荐的方法及BPM平台

Published patent application · CN114004505A

吕海、吴文峻、于笑明、于卓、张德海、王大帝、郝庆利、廖星创

Applicant at publication · 北京中电普华信息技术有限公司、北京航空航天大学

BPMN-Based Microservice Workflow Deployment

一种基于BPMN的微服务工作流部署方法

Patent grant · CN112463141B
ZL 2020 1 1338284.7

吴文峻、于笑明、王洋洲

Assignee at grant · Beihang University

Scenario-Based Intelligent Service Adaptation

一种基于场景的智能适配服务的方法及系统

Published patent application · CN112950147A

吴文峻、杨京波

Applicant at publication · Beihang University

Intelligent Code Detection and Repair System

一种智能检测及修复代码的方法与系统

Published patent application · CN112947985A

吴文峻、梁堉、杨京波

Applicant at publication · Beihang University

Kinect Gesture Control for Tiled Displays

一种基于Kinect手势控制显示墙的方法

Patent grant · CN103713741B

楼奕华、张海阔、吴文峻

Assignee at grant · Beihang University

Energy-Aware Data Backup for Data Centers

一种面向数据中心基于能耗感知的数据备份方法

Published patent application · CN105607967A

Inventors include Wenjun Wu

Applicant at publication · Beihang University

Software Copyrights

These registrations are listed on ADMIRE alumnus Qunbo Wang’s Beijing Jiaotong University faculty page.

Facilities

Explore experimental spaces, robots, specialized equipment and simulation environments, with Beihang and collaborating-team resources identified separately.

Facilities Overview

01Experimental Spaces

The Beihang Beijing space covers 100 m²; collaborating-team space is approximately 200 m². Indoor sandbox and coordination experiments are supported by motion capture and documented field trials.

Beihang urban sandbox100 m²
Collaborating Team Experimental Space~200 m²

Three motion capture systems provide position and orientation measurements for trajectory observation, coordinated control and simulation-to-real coordinate mapping.

More Space Photos
Experimental Spaces
Motion capture equipment

Field Trial Records

Heterogeneous Air–Ground Field Platform

UAV deployment, ground vehicles and swarm return

02Robotic Platforms

UAVs, ground and quadruped robots, mobile manipulators and dual-arm platforms support different experimental tasks; model and quantity details are in the inventory.

Multi-agent Cooperative Control Platform
Dual-arm manipulation: grasp and place
Models, quantities & resource scope

Equipment inventory (including collaborating resources)

The source inventory totals 176 units/systems across Beihang platforms and collaborating teams; it is not a count of resources independently owned by the group.

UAVs · 140 aircraft
AMOVLAB P230
Bitcraze Crazyflie
Ryze Tello

Product illustrations above; physical experiments are shown in the related demonstrations.

PlatformCountExperimental use
AMOVLAB P23020Onboard visual navigation, trajectory tracking and multi-UAV control
Bitcraze Crazyflie50Indoor formations, trajectories and swarm policies
Ryze Tello20Lightweight vision and flight programming
Collaborating team UAVs (multiple models)50Collaborating team resources
Ground Mobility & Inspection · 21 robots
SCOUT MINI R&D KIT
LIMO
Unitree Go2 EDU

Product illustrations above; physical experiments are shown in the related demonstrations.

PlatformCountExperimental use
SCOUT3Navigation and multi-vehicle coordination
SCOUT MINI R&D KIT1Indoor/outdoor navigation and path planning
LIMO9Multi-robot navigation and sandbox tasks
YUHESEN FW-mid3Lateral motion and path tracking
Unitree Go2 EDU5Quadruped navigation, inspection and vehicle–quadruped coordination; 3 at Beihang and 2 with the collaborating team
Mobile Manipulation · 15 systems
TRACER + PiPER
LIMO COBOT
COBOT MAGIC

Product illustrations above; physical experiments are shown in the related demonstrations.

PlatformCountExperimental use
SCOUT + Manipulator2Mobile grasping and transport
TRACER + PiPER2Mobile grasping and visually guided navigation/manipulation
LIMO COBOT10Pick-and-place, transfer and visual demonstrations; 5 at Beihang and 5 with the collaborating team
COBOT MAGIC1Bimanual teleoperation, demonstration collection and imitation learning
More Platform Photos
UAVs
Robotic Arms
Quadruped Robots
Humanoid Robots
Platforms & Demos

Multi-Agent Coordination

Experiments cover joint simulation and physical validation, indoor UAV formations, and air–ground tasks in an urban sandbox.

Air–Ground Sim-to-Real Platform

Three-phase mission (target search, coordinated strike, air-ground escort) verified on simulated and physical swarms in sync

Multi-agent Cooperative Control Platform

Multi-agent simulation and indoor UAV formation experiments

Urban Sandbox Air-Ground Testbed

UAV and ground-vehicle experiments among model roads and buildings

Navigation & Agentic UAV Systems

AERIS, VoLN and AeroWeaver study runtime orchestration, vision-only long-horizon navigation and coordinated aerial skills, respectively.

VoLN · Vision-Only Long-Horizon Navigation

Studies vision-only long-horizon UAV navigation using target images and visible cues, with simulation and real-flight demonstrations.

View method figure
VoLN vision-only long-horizon navigation framework
AeroWeaver console interface
AeroWeaver

Maps natural-language tasks to aerial skills, with distributed agents coordinating execution and feedback.

Robot Manipulation & Autonomous Motion

Four demonstrations cover dual-arm manipulation, cooperative navigation, autonomous exploration and quadruped control; select each in the playlist.

Embodied AI & Robotics Platform

Dual-arm manipulation, cooperative navigation, exploration and quadruped control

03Specialized Equipment

Intelligent Cockpit & Multimodal Human Factors

BJUT platforms at the Beijing Key Laboratory of Embodied Interaction Intelligence. This equipment is presented separately and is not included in the robot inventory above.

CARLA three-screen cockpit

Built around CARLA, three screens display the road, surround and rear views, dashboard and task interface. A steering wheel supports simulated driving experiments.

Synchronized sensing
Cameras, microphones and EEG sensors synchronously record visual behavior, speech interaction and physiological signals.
Comfort research
Facial expressions and EEG features support research into personalized driving style matching.
Safety research
Driving state and regional brain activity are analyzed to study distraction and fatigue alerts.

EEG Acquisition & OpenBCI Analysis

EEG headset and acquisition setup

An electrode cap connects to an OpenBCI Cyton biosignal amplifier. A USB receiver relays digitized signals wirelessly to OpenBCI GUI for live display and session recording.

Acquisition & integration
CSV and BDF+ export and LSL, UDP and OSC streaming support further analysis and platform integration.
State visualization
Focus and relaxation estimates, with topographic maps of brain activity.

Tri-Optical Gimbal & Power Line Inspection

Tri-optical gimbal and controls

The tri-optical gimbal and sensors support airborne perception and stable-imaging research. Algorithms, the inspection platform and demonstrations are grouped in the system case study.

Intelligent Inspection

Intelligent Inspection

Tri-Optical Gimbal & Power Line Inspection

A State Grid headquarters research project supports an intelligent gimbal for a dual-flying-wing VTOL UAV, combining onboard AI edge recognition with stable imaging during high-speed cruise.

Tri-optical gimbal and controls
Integrated UAV exhibit, 2025 Changchun Air Show
Fusion recognition
Visible-light and infrared feature matching combines four-stage data augmentation with an edge recognition algorithm using a recursive attention head.
Stable imaging
Onboard AI edge processing and AI servo control support power line imaging during high-speed cruise.
Detection & analysis
A digital twin platform displays defects, statistics and spatial locations. The iFLYTEK Spark language model generates inspection reports and supports conversational queries.
Algorithms & Engineering Results

The gimbal was exhibited on a collaborating organization’s dual-flying-wing UAV at the 2025 Changchun Air Show.

Four-stage augmentation and fusion recognition
Visible-light and infrared feature matching
Gimbal stabilization

Animated demonstration of the gimbal hardware and stabilization.

Animated demo

Visible-light and infrared inspection

Visible-light and infrared footage with object detection during flight.

Animated demo

Power line defect detection

Inspection task configuration and detection results.

Animated demo

Detection statistics & visualization

Detection summaries, charts and visual analysis.

Animated demo

AI inspection assistant

Detection results, map views and conversational interaction.

Animated demo

AI inspection report

Power line inspection reports generated from detection data.

04Computing & Simulation

3D simulation, swarm monitoring and physical testing

Simulation & Robot Software

Existing capabilities

The Tianji platform supports simulations with hundreds of agents using Unreal scenes and containerized services. Existing experiments cover MARL pursuit, robot navigation, manipulation demonstrations, policy training and real-robot testing.

In development

Harness is being developed to unify high-level tasks, commands and status across heterogeneous platforms through chassis and quadruped SDK adapters. The initial target is collaborative language-guided navigation with one vehicle and one quadruped, followed by UAV integration.

The Isaac Sim / Isaac Lab pipeline is under development, spanning modeling, unified observations, collaborative policy training, robustness validation, ROS 2 deployment and evaluation feedback, while reusing established low-level controllers.

Scenarios & Collaborative Tasks

Existing methods include knowledge-enhanced visual question answering, 3D object detection, language-guided navigation and multi-agent learning. Future collaboration targets collaborative perception, active exploration, cross-region inspection, mobile manipulation and long-horizon coordination, with jointly defined tasks and evaluation.

Academic Exchanges

Conferences, academic visits and research collaboration

Domestic & International Partnerships

The group maintains long-term international collaborations with the National University of Singapore, Nanyang Technological University and Politecnico di Milano, and conducts joint research with the Shanghai Artificial Intelligence Laboratory, Zhongguancun Laboratory, Qianyuan Laboratory and Qiyuan Laboratory (all four are national laboratories), advancing research collaboration and joint doctoral training.

National University of Singapore · Nanyang Technological University

Collaboration spans agents, multi-robot cooperation, swarm intelligence and large language models; several group members have visited Singapore for academic exchange and collaborative research.

Singapore

National Laboratory Partnerships

Long-term collaboration with the Shanghai Artificial Intelligence Laboratory, Zhongguancun Laboratory, Qianyuan Laboratory and Qiyuan Laboratory (all four are national laboratories) focuses on software agents, LLM safety, embodied AI and complex intelligent systems, alongside joint doctoral training.

China

Politecnico di Milano

Sustained collaboration on complex systems, intelligent decisions and interdisciplinary science, including visits by PhD students and postdoctoral researchers.

Italy

Experimental Platforms