01
Reuse what is knownCompositional
Compositional
Generalization
Recombine learned skills to solve tasks never seen during training—to learn by analogy, not memorization.
Digital intelligence · Embodied intelligence
Evolving
Machines Lab.
Building intelligence that learns, adapts,
and evolves with the world.
About the lab
Evolving Machines Lab brings together researchers from institutions including Dalian University of Technology, the University of Oxford, KAUST, and UCSD to advance open, continually evolving intelligence.
We build agents that act, gather experience, and become more capable in the worlds they inhabit. Our research connects digital societies with physical intelligence through a shared idea: learning never stops.
5,000+Open-source GitHub stars
Outstanding Paper AwardTop AI Conference
One question, two worlds
How do machines evolve?
01 / SimulationDigital world agentsInteract at scale, shape societies, and learn from influence.
02 / RealityPhysical world agentsPerceive space, compose skills, and improve through correction.
02
Learn from actionOn-policy
On-policy
Learning
Let agents act, receive feedback, and update from experience collected by their own evolving policy.
- Think3D-RLRL teaches agents to use 3D tools
- OASIS Interaction RewardLearn to interact and gain influence · Read blog →
- Real-world Arm DAggerHuman corrections · Coming soon
03
Expand what is possibleOpen-ended
Open-ended
Tool Ecosystem
Build an open-ended multimodal tool ecosystem, connecting perception, reasoning, creation, and action through a growing network of tools.
- SPAgentToward an open-ended multimodal tool ecosystem
- SPAgent-RSIComing soon
Built in the open
Research into reality.
01 / Digital World / Agent Societies
EML / SYSTEM 0102 / Physical World / Spatial Agent
EML / SYSTEM 02Physical World / Spatial Agent
SPAgent
A foundation agent for the physical and spatial world.
A general-purpose multimodal tool-calling framework that connects vision-language models to an open-ended ecosystem of perception, grounding, 3D, and generation tools.
- 26 tools
- Multimodal memory
- Agentic RL
03 / Physical World / Embodied Intelligence
EML / SYSTEM 03Physical World / Embodied Intelligence
Future
An open ecosystem for agentic, open-ended VLA policies.
Future develops embodied intelligence that decomposes goals into atomic skills, composes them on demand, and coordinates multiple effectors to solve tasks beyond its training distribution.
- ECCV 2026
- Multi-arm VLA
- Open ecosystem
Selected publications
Ideas in motion.
OASIS: Open Agent Social Interaction Simulations with One Million Agents
Ziyi Yang*, Zaibin Zhang*, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, et al.
5,000+ GitHub StarsPsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety
Zaibin Zhang*, Yongting Zhang*, Lijun Li, Hongzhi Gao, Lijun Wang, Huchuan Lu, Feng Zhao, Yu Qiao, Jing Shao
Outstanding Paper AwardThink3D: Thinking with Space for Spatial Reasoning
Zaibin Zhang*, Yuhan Wu*, Lianjie Jia*, Yifan Wang, Zhongbo Zhang, Yijiang Li, Binghao Ran, Fuxi Zhang, Zhuohan Sun, Zhenfei Yin, Lijun Wang, Huchuan Lu
200+ GitHub StarsMA-VLA: Multi-Arm Vision-Language-Action Model for Collaboration and Compositional Generalization
Zaibin Zhang*, Junlan Xiao*, Zhongbo Zhang*, Yifan Wang, Li Kang, Yiran Qin, Changxing Xia, Heng Zhou, Talas Fu, Enshen Zhou, Ruimao Zhang, Zhenfei Yin, Huchuan Lu, Lijun Wang
ECCV 2026Our research community
Connected by curiosity.
From the lab
The latest.
Oct 2026
Familiar skills, new compositions: ACG-Bench and AE-VLA
Our new research blog explores arm-wise compositional generalization, with simulation and real-robot demonstrations.
Read ↗Jul 2026MA-VLA and Attention-DP3 accepted to ECCV 2026
Two steps toward compositional multi-arm systems and spatially object-aware manipulation policies.
Read ↗May 2026Think3D is featured in VALSE 2026 Wuhan
Think3D is featured at VALSE 2026 in Wuhan, highlighting our work on spatial reasoning.
Read ↗