浙江大学 REAL Lab，致力于打造能在真实世界中思考、行动并持续成长的人工智能系统
REAL Lab at Zhejiang University builds AI systems that reason reliably, explore the physical world, act autonomously as agents, and learn continuously throughout their lifetime, bridging the gap between virtual intelligence and physical intelligence.
HuggingGPT  An LLM as controller and the Hugging Face hub as executors: plan, pick expert models, run them, and compose the answer. |
HugAgentOS  A self-evolving AgentOS. Memory, skills, and orchestration improve with every task, and each evolution step is attributable, replayable, and reversible. |
Easel  An open-source social-media agent: discover trends, create posts, audio, and video, publish across platforms, and learn what actually works. |
ClawGUI  Train, evaluate, and deploy GUI agents in one framework: online RL across parallel emulators and real Android, HarmonyOS, and iOS devices. |
Models that can reason for as long as a problem requires, spend only the thinking it needs, and know when an answer is right.
| Direction | Focus | Representative work |
|---|---|---|
| Long-horizon reasoning | Break the context-length ceiling so reasoning can continue as long as the problem demands. | InftyThink · InftyThink+ |
| Efficient reasoning | Fit the amount of thinking to the difficulty of the problem and stop overthinking. | Self-Braking Tuning · LAPO · Mind the Gap |
| Verifiable reasoning | Verifiers and rewards that can be trusted, and models that pause instead of fabricate. | VerifyBench · CoVerRL · Pause or Fabricate? |
Taking intelligence off the screen: understanding 3D space, anticipating how the world changes, and acting through a body.
| Direction | Focus | Representative work |
|---|---|---|
| Spatial perception | Measure and teach the 3D understanding that 2D pretraining lacks, including viewpoints other than the camera's. | ViewSpatial-Bench · SpatialLadder · GPD |
| Self-evolving spatial intelligence | Let geometry and physical interaction supply their own supervision, without human labels. | SpatialEvo · Spatial-Interactor |
| Prediction and world models | Reason about what happens next and keep a memory of the world over time. | SpaceCast-Bench · WorldMemBench |
| Embodied action | Turn perception into reliable physical action and coordinated embodied behavior. | VLA-Corrector · OmniEmbodied |
Agents that plan, use tools, and carry long-horizon tasks through to completion across screens, software, and the open web.
| Direction | Focus | Representative work |
|---|---|---|
| GUI grounding | Locate exactly where to act on a screen, with dense rewards, test-time learning, and adaptive zoom. | GUI-G² · GUI-RCPO · Reliable Negatives · UI-Zoomer |
| Computer-use agents | Long-horizon control of real devices through multi-turn RL, GUI plus command line, and online self-improvement. | UI-S1 · EMPO · UI-Copilot · HybridCUA · ComputerSD · Reflect, Revise, Reuse · ClawGUI |
| Deep research agents | Search the open web, synthesize evidence, and carry a project from literature to a reviewed paper. | IterSynth · BrowserForge · EviNote-RAG · PaperGym · Polaris |
| Code agents | Code and tests that co-evolve adversarially through reinforcement learning. | Code-A1 |
| Creation agents | Agents that discover, create, publish across platforms, and learn what actually works. | Easel |
| Tool orchestration | LLMs as controllers that plan, call expert models and tools, and compose the result. | HuggingGPT · TaskBench |
| Agent evaluation | Interactive, personalized benchmarks and a plug-in harness that evaluates any agent in any environment. | KnowU-Bench · ageval |
Deployment is where learning continues: agents turn their own experience into guidance, skills, and better weights.
| Direction | Focus | Representative work |
|---|---|---|
| Guided exploration | Densify the sparse rewards of long-horizon agentic RL with adaptive hints and milestones. | Agent-G² · BEACON |
| Skill internalization | Discover skills in context, then internalize them into the policy across tasks and modalities. | SKILL0 · ViSkill · SkillRise |
| Self-distillation | Agentic RL with on-policy self-distillation, and the family of methods built on it. | SDAR · AgentOPSD · Pass the Baton · RetireOPD |
| Post-deployment steering | Adjust a deployed model's behavior at inference time without retraining. | EasySteer |
| Self-evolving systems | Memory, skills, and orchestration that evolve with use, with every step attributable and reversible. | HugAgentOS |
If our work helps you, a ⭐ on the repository helps others find it too.
