Shi (Billy) Chen

I'm a final-year undergraduate at Fudan University studying Computer Science and Technology, and an exchange student at UC Berkeley. I'm the Co-founder & CTO of Intuition Core Inc., an AI & Robotics startup developing infrastructure solutions to accelerate real-world robot deployment.

My research interests span computer vision, 3D generation, robotics, and embodied AI. I've been fortunate to work with amazing researchers at UC Berkeley, University of Chicago, University of Cambridge, and Johns Hopkins University. I'm honored to be funded by the National Science Foundation of China under its Youth Fund (one of the highest honors for undergraduates in China; ~120 students nationwide annually).

My mission: I'm driven by the vision of bringing intelligent robots from research labs into the real world, creating transformative solutions that meaningfully improve human lives and unlock new possibilities for society.

Email  /  CV  /  GitHub

News

• [May 2025] Co-founded Intuition Core Inc. and secured $950K in pre-seed funding from Founders Fund and Fen Venture at $24M post-money valuation!
• [Jan 2025] Started exchange program at UC Berkeley, taking courses on Programming Languages and Compilers, Robotic Manipulation and Interaction, and Advanced Large Language Model Agents.
• [Sept 2024] Awarded funding by the National Science Foundation of China under its Youth Fund.
• [Aug 2024] Completed summer research at University of Cambridge, graduating with First with Distinction.

Entrepreneurial Experience

Co-founded an AI & Robotics startup developing infrastructure solutions to accelerate real-world robot deployment. Designed a novel Masked Autoencoders (MAE) based world model pretraining architecture achieving 15%+ downstream policy training efficiency. Built ultra-low latency (<100ms) teleoperation systems over long distances (Hawaii-SF, Shanghai-SF). Secured $950K in pre-seed funding from Founders Fund and Fen Venture at $24M post-money valuation cap.

Demo video of our Vision-Language-Action (VLA) model trained with our customized data collection pipeline and optimized with our world model pre-training method:

Research

I'm interested in computer vision, 3D generation, robotics, and embodied AI. My research focuses on developing methods that enable machines to perceive, understand, and interact with the physical world. Representative works are highlighted.

Proposed a 3D latent diffusion model with multi-stage data augmentation for sparse LiDAR datasets. Designed a multinomial discrete diffusion model achieving 39 mIoU, surpassing previous SOTA (37.9 mIoU). Funded by the National Science Foundation of China Youth Fund. Currently deploying on Unitree G1 humanoid and Go1 robot dog for outdoor navigation.

Extended the Continuous Remeshing pipeline to generate high-quality meshes from single-frame images. Introduced vertex color optimization with novel loss terms. Demonstrated improved mesh quality over baseline Marching Cube algorithm with >70% preference in crowdsourced user study.

Built a custom quadrotor under the Agilicious framework with specialized hardware. Enabled real-time state estimation using SVO Pro algorithm for GPS-free navigation. Fine-tuned Qwen-2.5 3B VLM for onboard deployment on Jetson Orin Nano. Verified feasibility of onboard VLM deployment for real-time quadrotor navigation, reducing hardware costs by several hundred dollars.

Proposed a multimodal diffusion pipeline integrating ControlNet into Stable Diffusion, increasing MediaPipe Confidence scores from 0.248 to 0.556 (2.24x improvement). Created a synthetic dataset of 170,000+ images across 809 gesture types with 83.5% quality rating. Selected for Cambridge Summer Research Programme, graduating with First with Distinction.

Developed a diffusion model to generate 3D bounding boxes and BEV perception data on NuScenes dataset. Combined synthetic input for MagicDrive to create temporally consistent RGB videos. Achieved 2-4% accuracy improvement on downstream tasks including BEV perception and 3D bounding box detection on BEVFusion. Used NeRF-based Nerfacto for controllable 3D scene rendering.

Education

Other Selected Projects

Developed a stateful LLM agent with persistent memory for academic research management. Built Chrome extension to auto-fill graduate school applications. Integrated automated paper discovery using LangChain + GPT-4o. Achieved 98% success rate in automated posting tasks.

Skills & Technical Expertise

Programming Languages: C, C++, Python, OCaml, Lean (Theorem Proving), Verilog HDL, Pascal, MATLAB
Frameworks: PyTorch, PyTorch3d, JAX, OpenCV, NumPy, LangChain
Robotics & Simulation: ROS / ROS2, MuJoCo, OpenAI Gym
Modeling & Design: Blender (3D Modeling), Unity
Tools: Git, Linux, Docker, LaTeX
Languages: English (TOEFL 106, Speaking 25; GRE 336, Quantitative 170)

Thanks to Jon Barron for the template.