Liu Yue is an Embodied AI researcher at Tongji University, working at the intersection of foundation models, vision-language-action systems, and surgical/medical robotics. He is passionate about building the next generation of embodied intelligent systems and turning research into real-world products. Outside of research, he enjoys track and field, ball sports, and singing.
Electronics and Mechanical Engineering
Tongji University
My research sits at the intersection of embodied AI foundation models and real-world robot deployment. I focus on two complementary threads.
The first is VLA memory systems — specifically, how robots can accumulate and update knowledge across long-horizon tasks without unbounded memory growth. I’m building GDN-Memory, a constant-footprint associative memory integrated into a Slow-Fast video world model via Gated DeltaNet fast weights, targeting ICRA/CoRL 2026.
The second is endovascular surgical robotics — end-to-end intelligence for catheter navigation, from high-fidelity SOFA simulation and synthetic data generation to diffusion-based trajectory planning and real-time DSA guidance.
Alongside the research, I co-founded a seed-stage startup deploying VLA technology into vertical robotics applications, backed by a partner at Qiji Chuangtan (formerly YC China) through a personal investment.
Feel free to reach out for collaboration or to discuss ideas.
A three-layer (physical / object / behavioral) constant-memory system built on Gated DeltaNet fast weights, integrated into Genie Envisioner's Slow-Fast inference pipeline. 256 KB …
High-fidelity SOFA-based surgical simulation generates synthetic training data for a diffusion model that plans endovascular catheter trajectories, overcoming real-data scarcity.
We are a seed-stage startup focused on deploying embodied intelligence into real-world vertical applications, bridging cutting-edge VLA research and production-ready robotics systems in industrial and medical domains.
We have closed a seed round via a personal investment from a partner at Qiji Chuangtan (formerly YC China), with initial intent orders already in place and first deployment engagements underway.
Our core focus spans rapid VLA adaptation for vertical scenes (few-shot fine-tuning + TTT online learning), embodied perception and manipulation for industrial and medical robots, and an end-to-end deployment stack from training to edge inference.
Interested in partnership or investment? Feel free to reach out.
参数高效微调（PEFT）技术深度解析：BitFit、LoRA、Prefix Tuning、QLoRA 的数学原理与实现细节。
