Interleaved Projected Gradient Descent for Safe Imitation Learning
We propose an imitation-learning design for neural-network control policies under state and input constraints. Training alternates a standard imitation gradient step with a block …
🚗Shengfan Cao
(he/him)
PhD Researcher in Autonomous Driving & Robotics
Professional Summary
UC Berkeley Ph.D. candidate developing safe, machine-learning-based prediction, planning, and control systems for autonomous vehicles. First-author IEEE IROS oral presenter with 4+ years across machine learning, planning, and controls, spanning PyTorch research, CARLA evaluation, embedded-robot deployment, and C++ production planner integration.
Education
Ph.D. in Mechanical Engineering
2022-08-01
2027-05-31
University of California, Berkeley
B.S. in Mechanical Engineering
2018-09-01
2022-06-30
Tsinghua University
B.A. in Japanese Language and Literature
2017-08-23
2022-06-30
Tsinghua University
Interests
Autonomous Driving
Safe Learning
Control Theory
End-to-End Autonomy
Robotics
📚 My Research
I am a PhD researcher at UC Berkeley specializing in safe learning, control theory, and end-to-end autonomy for robotic systems. My work bridges the gap between theoretical guarantees and real-world deployment, focusing on how large-scale data and physical constraints shape learning-based autonomous systems.
My current research highlights include:
- Trusted Autonomy: developing rapidly deployable ground robots that can explore and map unseen terrain with minimal setup.
- Safe Learning: designing constraint-aware imitation learning frameworks that ensure safety even at dynamic handling limits.
- Policy Optimization: formulating sampling-based weight-space projection methods to enforce safety constraints during policy updates.
I leverage tools like PyTorch, ROS 2, and convex optimization to build interpretable and reliable autonomy stacks. I am actively transitioning into industry to apply these principles to large-scale, safety-critical applications.
Please reach out to collaborate! 🚗
Featured Publications
Constrained Policy Optimization via Sampling-Based Weight-Space Projection
Safety-critical learning requires policies that improve performance without leaving the safe operating regime. We study constrained policy learning where model parameters must …
A Simple Approach to Constraint-Aware Imitation Learning with Application to Autonomous Racing
Guaranteeing constraint satisfaction is a challenging problem in imitation learning (IL), especially when the task involves operating at the handling limits of the system. While …
Recent Publications
(2026).
Interleaved Projected Gradient Descent for Safe Imitation Learning.
Preprint.
(2025).
Constrained Policy Optimization via Sampling-Based Weight-Space Projection.
IFAC WC 2026 (Oral).
(2025).
State-Conditional Adversarial Learning: An Off-Policy Visual Domain Transfer Method for End-to-End Imitation Learning.
IFAC WC 2026.
(2025).
A Simple Approach to Constraint-Aware Imitation Learning with Application to Autonomous Racing.
IROS 2025.
(2025).
Real-Time Regulation-Aware Game-Theoretic Motion Planning for Head-to-Head Autonomous Racing.
IEEE TCST (Under Review).
Recent & Upcoming Talks
Recent News
Paper Accepted at IROS 2025
I am excited to share that my paper “A Simple Approach to Constraint-Aware Imitation Learning with Application to Autonomous Racing” has been accepted for oral presentation at IROS …