Fan Yang

I am a PhD student in robotics at University of Michigan, Ann Arbor advised by Prof. Dmitry Berenson. I am working on dexterous robot manipulation with a multi-fingered hand. Previously, I was a master's student in robotics (MSR) at Robotics Institute of Carnegie Mellon University, where I was advised by Prof. David Held. I was working on safe Reinforcement Learning and manipualtion with Prof. Held.

Prior to Joining CMU, I was an undergraduate student at Tsinghua University majoring in Engineering Mechanics (Qian class). I was also fortunate to have a chance to be advised by Prof. Richard Voyles, Prof. Oussama Khatib, Prof. Masayoshi Tomizuka, and Prof. Huaping Liu.

Email  /  CV  /  Google Scholar  / 

We propose a diffusion-informed A* planning algorithm, that plans both contact modes and trajectories. We evaluate our method on the challenging screwdriver turning task in both sim and real

Keywords: dexterous manipulation, diffusion model, motion planning

We propose a trajectory optimization algorithm, that takes into account both object and robot finger geometry, framing the contact constraints and finger rolling constraints differentiablly. Our metod also considers the extrinsic contact between the object and the environment.

Keywords: dexterous manipulation, trajectory optimization

We propose an algorithm where the RL agent operates in a modified MDP, embedded with a trajectory optimization algorithm to ensure safety.

Keywords: Safe RL, Trajectory Optimization, Markov Decision Process

We propose a diffusion model that generates subgoals dynamically in a coarse-to-fine manner, trained by random play data.

Keywords: Diffusion Model, Reachability, Coarse to Fine

We propose a spatially-grounded and temporally-abstracted action representation with a hybrid discrete-continuous reinforcement learning framework.

Keywords: RL with 3D Vision, Action Representation, Contact-rich manipulation

The method of graduating incurring a harder safety constraints can lead to a better performance in safe RL tasks.

We develop a benchmark and evaluate the state-of-the-art lifelong learning method on reinforcement learning tasks, especially robotic manipulation tasks.

An adversarial training algorithm is used to increase the robustness of robot joint damage.

We develop a method that uses RL agent to select important interactions for trajectory prediction in a highly interactive environment.

We develop a Graph-Neural-Network-based method captures the interactions between different agents for trajectory prediction in a highly interactive environment.