Kaitian Chao

I am a second-year Master student majoring in Robotics at the University of Pennsylvania affiliated to GRASP Laboratory, specializing in robot learning, control, and 3D computer vision. I have extensive experience in deep learning and robotics research, with projects spanning VLM and VLA for robot arm fine-manipulation, robotic fish modeling and control, F1tenth autonomous driving, UAV control and navigation, diffusion models, and large language models, all aimed at advancing AI-driven capabilities. My passion lies in leveraging cutting-edge AI and vision technologies to develop super-intelligent robotics and software that can significantly enhance productivity. I am fortunate to work with Junyao Shi, Jason Ma and Prof. Dinesh Jayaraman at PAL lab.

Previously, I received my B.S. in Electrical Engineering from ShanghaiTech University, where I worked with Xiaozhu Lin and Prof. Yang Wang on robotics fish modeling and control. I also spent my junior year as an exchange student at the University of California, Berkeley Electrical Engineering & Computer Sciences.

Email  /  CV  /  Scholar  /  Twitter  /  Github /  LinkedIn /  GRASP Lab Profile /  Location: Philadelphia, PA

Research

I'm interested in robot learning, control, and 3D computer vision. My research focuses on developing AI-driven capabilities for robotics applications. Some papers are highlighted.

Maestro is a general-purpose “code-as-policy” framework that uses coding VLMs to compose perception, planning, and control tools for diverse robot manipulation on both tabletop and mobile platforms. By self-evolving its generated programs to gradually improve task capabilities and integrating state-of-the-art VLA models, it achieves strong zero-shot performance, supports active perception can be used for data collection/finetuning, and in some settings surpasses VLA-only systems.

We present a novel data-driven dynamic modeling framework capable of characterizing the swimming motions of the robotic fish under various background flow conditions without the necessity for explicit flow information. The model is synthesized by an internal model with an adaptive residual acceleration model to effectively isolate and address external flow effects. Notably, the residual model employs the innovative Domain Adversarially Invariant Meta-Learning (DAIML) approach, allowing the framework to adapt to fluctuating and previously unseen background flow scenarios, enhancing its robustness and scalability. Validation through high-fidelity Computational Fluid Dynamics (CFD) simulations demonstrates the framework’s effectiveness in improving the performance of robotic fish across diverse real-world aquatic environments.

Inspired by the natural lateral line system(LLS), a flowresponsive organ in fish that plays a crucial role in behaviors such as rheotaxis, this paper introduces the first Artificial Lateral Line System (ALLS)-based ambient flow classifier for robotic fish that allows robotic fish to perceive flow fields while swimming freely. To be specific, using just 5 pressure sensors and 3.5 minutes of swimming data, we trained a Long Short-Term Memory (LSTM) network, achieving a classification accuracy of 81.25% across 8 flow speed categories, ranging from 0.08 m/s to 0.18 m/s. A key innovation of this work is the formulation of ambient flow perception as a classification task, which not only enables the robotic fish to extract meaningful information but also enhances the robustness and generalizability of the perception framework.

Projects

Here are some of my personal and course projects showcasing my technical skills in robotics, machine learning, and software development.

We pushed the limits of autonomous racing at the 24th Roboracer Grand Prix at ICRA 2025, a high-stakes, head-to-head competition for 1/10th scale F1 racing cars. Our approach combined meticulous hardware tuning with a hybrid software stack, switching between Pure Pursuit and Model Predictive Control (MPC) to optimize for raw speed and agile control. This strategy proved highly effective, leading us to secure a high ranking against a field of international teams.

This project demonstrates a complete pipeline for autonomous quadrotor navigation in dense, unknown environments. We developed a system that uses onboard vision to build a map, plans the optimal collision-free path using A*, and executes agile maneuvers using a geometric controller based on differential flatness. This integration of perception, planning, and advanced control allows the drone to fly aggressive, time-optimal trajectories with high precision.

This project showcases a high-performance autonomous driving system for a 1/10th scale F1 racing car, leveraging a Model Predictive Path Integral (MPPI) controller for agile navigation. Using a LiDAR-built SLAM map within a ROS2 framework, the system demonstrates robust, real-time obstacle avoidance in both simulation and on the physical F1TENTH vehicle.

This project showcases a complete robotics pipeline for a 7-DOF Franka Emika Panda arm to autonomously perform a dynamic pick-and-place task. By integrating ROS2, computer vision, and advanced motion planning, our system ranked high in the course competition by demonstrating high-precision manipulation in stacking both static cubes and dynamic cubes on a rotating platform.

Website template adapted from Jon Barron. Last updated: