I am currently pursuing a master's degree at EPFL and joining the Soft Robotics Lab at ETH as a master's student, supervised by Prof. Robert Katzschmann. I received my honours degree (summa cum laude) from Turing class at the School of Computer Science, Peking University, where I was mainly advised by Prof. Hao Dong on robotics learning. I also collaborated closely with Prof. Zongqing Lu on the spatial cognitive abilities of machine learning models.

Outside of academics, I spend much of my free time at the piano, particularly in classical chamber music. During my undergraduate years I formed a piano trio with friends. Currently I am collaborating with local soprano Liudmila Arno, actively performing in Lausanne, Switzerland.

I'm interested in reinforcement learning, robotics and computational neurology.

To me, the essence of robotics lies in agility and dexterity.

Across evolution, animals have refined motor control in a billion-year interplay of body and nervous system — through a vast distributed training process governed by survival and death, extending three orders of magnitude beyond humanity’s brief venture into language, logic, and mathematics, the very emblems of our civilization. The architecture of the brain itself inscribes this history: to motion are entrusted far more neurons than ever to words.

We take pride in having taught machines to mirror our language, yet in doing so we are reminded of our limits: for what we prize as the summit of intellect, evolution regards as a late ornament.

As humble pilgrims we walk, striving to grasp nature’s vast bequest in motor control, to emulate its mastery, and perhaps one day, to venture beyond it.

Arnold is a single generalist muscle control policy trained on 14 different muscle control tasks, with 4 different embodiments. We used parallel on-policy behavior cloning, RL fine-tuning and self-distillation to match or even surpass expert performance on all tasks.

RoboVerse is a unified framework for robotic learning. It provides an API interface to run simulations with multiple back-ends, equipped with large-scale datasets and benchmarks for training and evaluating robotic learning algorithms.

We achieved state-of-the-art manipulation performance by combining reinforcement learning and multi-view pose estimation.

We investigated whether a compositional language for describing 2-D positional relation can emerge from communication under environmental pressure.

We proposed an image-guided depth completion model which is SOTA on NYUv2 dataset and won the first place on KITTI benchmark at the time of submission.

This is the summary publication to the MyoChallenge 2022, a competition on die reorientation and baoding balls.

In this study, we take advantage of visual affordance by using the contact information generated during the RL training process to predict contact maps of interest.

Reconfiguring a die to match desired goal orientations. This task require delicate coordination of various muscles to manipulate the die without dropping it.

We propose a framework that utilizes only RGB inputs and enables 3D articulated object manipulation.

This template is a modification to Jon Barron's website.