Robotics · Learning · Control
Learning to act
in the real world.
I am a master’s student at Southern University of Science and Technology (SUSTech), advised by Prof. Wei Zhang in CLEAR Lab. My interests lie in robotic manipulation, deep learning, and generative models for robotics.
I am particularly interested in manipulation in the wild, world action models (WAM), and how compact models can generalize in robot learning.
01
Research
Manuscript under reviewICRA 2027 submission
Residual-Gated Language Injection in Diffusion Policies for Language-Conditioned Robotic Manipulation
Tianxing Liu et al. · Sole first author
RGLI introduces a residual-gated cross-attention interface for token-level language conditioning in temporal U-Net Diffusion Policies, preserving visual and diffusion-time FiLM conditioning.
- Conditioning-only adaptation on LIBERO-90: 75.5% success versus 58.6% for DP-Cat, with studies of instruction dependence and paraphrase robustness.
- Evaluation across four LIBERO suites and six real-world manipulation tasks: 94.17% mean real-world success versus 83.33% for the matched U-Net baseline.
Submitted to the 2027 IEEE International Conference on Robotics and Automation. Decision pending.
03
Selected projects
Control theory · Open-source notesApr. — Aug. 2026
Control-Note
A Practical Guide to Control Theory
Compiled four years of undergraduate study notes and practical insights in optimization and control into a structured guide, with AI-assisted organization and editing. Chinese and English editions emphasize intuition and robotics engineering applications.
Explore the notes ↗
Imitation learning · SimulationJan. — Mar. 2025
Dexterous manipulation
Institute for AI Industry Research (AIR), Tsinghua University · Winter camp
Built a data collection pipeline integrating an INSPIRE ROBOTS dexterous hand with an AIRBOT Play robotic arm. Constructed a MuJoCo simulation environment and trained imitation learning policies using ACT.
Selected project code ↗
Embedded systems · ControlSep. 2022 — Sep. 2024
RoboMaster autonomous sentry
T-DT Team, Northeastern University
Designed circuits and control algorithms for an autonomous sentry robot and developed the team’s STM32F4/STM32H7 embedded software framework. Improved six-axis IMU attitude estimation based on Madgwick gradient-descent fusion to enhance stability and estimation accuracy.
Contributed to the team’s national runner-up finish in 2023 and third-place finish in 2024.
6-Axis-AHRS repository ↗
04
Skills & tools
- Programming
- C, C++, Python, MATLAB
- Robot learning
- Imitation learning (ACT, Diffusion Policy); reinforcement learning (PPO)
- Frameworks & simulation
- PyTorch, ROS, MuJoCo
- Embedded systems
- STM32F4 / STM32H7 software development, sensor calibration, robot debugging
- Development
- Linux, Git; AI-assisted development with Codex and Claude Code