World Action Models
Learning predictive models that connect observations, actions, and future physical states for planning and decision making.
World modelsAction-conditioned prediction
About me
Junqi Jing 荆浚淇
Embodied AI Researcher · B.E. Candidate
Harbin Institute of Technology
AI Research Intern at KNOWIN AI
World Action Models, VLA systems, and real-world robot learning.
I’m Junqi Jing (荆浚淇), a B.E. candidate in Software Engineering at Harbin Institute of Technology, expected to graduate in 2027. My research interests center on Embodied AI and world models, with a particular interest in generative approaches to learning and planning for physical interaction.
I am currently an AI Research Intern at KNOWIN AI, where I work on VLA-based dexterous-hand manipulation and exploratory research on world action models. Previously, I spent a semester as an exchange student at POSTECH, conducting research in the MLV Lab with Prof. Kwang In Kim, and later visited Tsinghua University’s LEAP Lab under the guidance of Prof. Gao Huang.
I am currently seeking research-oriented internship opportunities and look forward to connecting with research teams across industry, especially at leading technology companies working on embodied intelligence and robotics.
Embodied AIWorld Action ModelsVision-Language-ActionDexterous Manipulation
Research agenda
From prediction to physical action.
My work connects visual generation, temporal reasoning, and robot control. I am especially interested in representations that help an embodied agent anticipate what comes next—and choose what to do.
VLA & Dexterous Manipulation
Building vision-language-action systems that translate high-level intent into reliable behavior on real dexterous hands.
VLAReal-world robotics
Generative Physical Intelligence
Exploring intermediate representations and generative objectives that make physical reasoning more structured, transferable, and useful.
Generative AIIntermediate representations
Real-robot work
Robots, not just benchmarks.
This reel is built for real hardware footage. Videos play muted when they enter the viewport, pause off-screen, and retain native controls for sound and full-screen viewing.
DEMO 01
Playing on view
01
Dual-arm teleoperation
A real-robot workspace for studying coordinated manipulation, demonstration collection, and human-in-the-loop control.
Real robotTeleoperation
PLAYBACK
READY
DEMO 02
Ready for footage
02
Dexterous manipulation
Reserved for the next real-hardware behavior and its task, method, and experimental context.
VLADexterous hands
PLAYBACK
READY
DEMO 03
Ready for footage
03
Next robot demo
A flexible slot for future experiments—add an MP4 and optional poster without changing the page layout.
Real-world evaluation
MP4, WebM, or direct video URLs can be added without changing this layout.
ICLR 2025
Selected publication
Masked Temporal Interpolation Diffusion for Procedure Planning in Instructional Videos
MTID introduces latent-space temporal interpolation and task-aware masking to improve action-sequence planning between observed start and goal states.
The method supplies visual mid-state supervision directly in latent space, encouraging temporally coherent plans without relying only on text-level supervision.
Research journey
Learning across models, labs, and machines.
Apr 2026 — Present
AI Research Intern
KNOWIN AI
Dexterous-hand manipulation with VLA models and exploratory work on World Action Models.
Dec 2025 — Mar 2026
Visiting Student
Tsinghua University · LEAP Lab
Real-hardware practice, teleoperation, and exploration of world-action modeling.
Sep 2025 — Dec 2025
Exchange Student Researcher
POSTECH · MLV Lab
Embodied AI and generative-model research with Prof. Kwang In Kim.
2023 — 2027
B.E. in Software Engineering
Harbin Institute of Technology
Research foundations in embodied intelligence, generative AI, and temporal reasoning.
Beyond the lab
Selected projects with measurable impact.
Top 1%
FleetEncounter
Award-winning cross-platform multiplayer game and HIT Best Project of the Year.
UnityMultiplayer
10,000+HITCS Study Guide
Open, collaborative computer-science course materials used by thousands of students.
Open educationCommunity
Top 6Travel Planning Assistant
Qwen-based travel agent recognized at the Alibaba Large Model University Tour.
AgentQwen 2.5
Field notes
Recent milestones.
Collaboration
Let’s build systems that can understand—and change—the physical world.
I am open to research internships, technical conversations, and collaborations with industry research teams working on Embodied AI, World Models, and robot learning.
Start a conversation
xingkong8527@gmail.com