Hi, my name is

Haoyang Weng.

Nice to meet you!

I’m a third-year undergraduate at Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University, Beijing, China.

I am working on research projects under the supervision of Prof. Yang Gao. Currently, I am serving as a student intern in LeCAR Lab at CMU, where I am advised by Prof. Guanya Shi.

I am intersted in the intersection between Machine Learning and Robotics. My goal is to build scalable and efficient robot learning algorithms, by incorporating both priors from intuitive physics and large scale real-world data. I believe that the combination of clean implementation of minimal algorithmic priors and large scale multi-source data will lead to generalizable robot learning algorithms.

Email CV GitHub

Publication

Robotics Impedance Control Reinforcement Learning

FACET: Force-Adaptive Control via Impedance Reference Tracking for Legged Robots

Force-adaptive policy conditioned on desired stiffness as command input, enabling variable impedance control.

Arxiv Twitter Website Live Demo

3D Reconstruction Distributed Training ICLR 2025

On Scaling Up 3D Gaussian Splatting

Developed √batch-size learning rate scaling enabling near-linear speedup in distributed 3DGS training.

Arxiv Twitter Website

Reinforcement Learning Foundation Models Robotics CoRL 2024

Reinforcement Learning with Foundation Priors: Let Embodied Agent Efficiently Learn on Its Own

Developed RLFP framework using foundation models for policy/value guidance, enabling sample-efficient real-world RL finetuning.

Arxiv

Course Projects

Model-based RL Robotics

Accelerating Model-based Learning of Locomotion with Pretrained Demonstrations

PPO-trained A1 robot locomotion in IsaacGym with Dreamer-based online fine-tuning in MuJoCo.

Reinforcement Learning C++/Pybind11 Sim2Real Robotics

Cyber Dribble: Flexible Framework for Agile Dribbling with Cyberdog2

Trains a RL policy for Cyberdog2 to perform agile dribbling tasks. Full-stack framework for real-world deployment, built on pybind11.

Code

Gaussian Splatting 3D Reconstruction

Gauss Fusion: Seamless 3D Object Integration using Gaussian Splatting

Gaussian Splatting preprocessing and merging pipeline with interactive 3D visualization.

Code

Embedded Systems C++/libpulse Raspberry Pi

Audio Reactive Controls: Real Time Visualization of Music on Customised Hardware

A custom music player hardware controlling ferrofluids and LED light strips for audio visualization.
Low-latency audio processing (C++/libpulse) for real-time hardware visualization on Raspberry Pi.

Code