Yaozeng Huang · 黄耀增

Yaozeng Huang 黄耀增

School of Computer Science & Technology, Xinjiang University

I am a third-year undergraduate at Xinjiang University. My research interests are at the intersection of embodied AI and reinforcement learning — vision-language-action (VLA) models for contact-rich manipulation, and policy learning for legged locomotion and dynamic-obstacle navigation.

I am fortunate to be working with researchers at Tsinghua SIGS on force-aware manipulation, and with Eastern Institute of Technology, Ningbo on learning-based local navigation. I want robots to acquire the geometric and physical common sense needed to act competently in the real world.

Email13325905201[at]163.com

GitHub@hyyyyyyz

CV中文 CV ↗

LocationÜrümqi, Xinjiang · Hometown: Zhoushan, Zhejiang

01News

Jan 2026

Started a research collaboration with Eastern Institute of Technology, Ningbo on RL-based local planning under dynamic obstacles. Targeting T-RO / TMECH.

Oct 2025

Joined Tsinghua SIGS as a research intern; working on force-aware VLA models (pi0.5 / TA-VLA) and an RGB-D privileged-geometry distillation method for contact-rich manipulation.

Aug 2025

First-author paper accepted to ICIC 2026 as Oral Presentation — RL locomotion & LiDAR navigation for quadruped robots.

Aug 2025

Awarded National Second Prize at the National Undergraduate Electronic Design Contest (NEDC).

Jul 2025

Won two National First Prizes at the 24th ROBOCON — Bionic Quadruped Obstacle Race & Off-road Race.

Jun 2025

Started research at the XJU Innovation Lab on RL-based legged locomotion and quadruped LiDAR navigation.

02Research Interests

I want robots to develop the geometric and physical common sense needed for contact-rich, dynamic, real-world tasks. The questions I find most exciting sit at the seam of perception, policy, and embodiment — and the evidence that matters most to me is real hardware, not just simulation curves.

Vision-Language-Action ModelsForce-Aware ManipulationReinforcement LearningLegged LocomotionDynamic Obstacle NavigationSim-to-Real TransferRGB-D Geometric Reasoning

03Publications & Preprints

First authors are bolded. ★ denotes oral presentation.

ICIC 2026Oral ★

Yaozeng Huang, et al. Reinforcement-Learning Locomotion and LiDAR Navigation for Quadruped Robots across Heterogeneous Terrains.

21st International Conference on Intelligent Computing · CCF-C · Accepted

In Prep.

Yaozeng Huang, et al. Action-Conditioned 2.5D Interaction Maps: RGB-D Privileged-Geometry Distillation for Contact-Rich Manipulation.

First-author · Tsinghua SIGS collaboration · targeting top-tier robotics venue

In Prep.

— et al. Curriculum Reinforcement Learning with Temporal Dynamic Encoding for Local Navigation among Moving Obstacles.

Co-author · EIT Ningbo collaboration · targeting T-RO / TMECH

Working titles for in-preparation manuscripts; final wording will be updated upon submission.

04Selected Research Projects

Force-Aware VLA Models for Dexterous Manipulation

Tsinghua SIGS · Oct 2025 – Present · Research Intern

Built a bimanual real-robot platform (松灵 7-DoF arms + grippers, Jetson AGX Orin, dual D405 wrist cameras + D435 head camera) collecting joint-position / torque proprioception and depth streams for contact-rich tasks such as optical-module insertion and block placement.
Reproduced and extended pi0.5 and TA-VLA; surveyed diffusion-policy, flow-matching, autoregressive, and hybrid VLA architectures, contrasting their fitness for high-frequency action generation under contact.
Ongoing: investigating geometric failure modes of VLA policies and an RGB-D privileged-geometry distillation method that uses action-conditioned 2.5D interaction maps to repair contact / placement failures under occlusion and depth ambiguity.

Reinforcement Learning for Dynamic-Obstacle Local Navigation

Eastern Institute of Technology, Ningbo · Jan 2026 – Present

Extending the DRL-DCLP local-planning line of work from static to dynamic obstacle settings — formulating the avoidance MDP for robots of varying footprint and designing matching simulation tasks, state space, and rewards.
Designed an 8-stage curriculum on StageRos: each stage scales the map by ~80%, increases static density, and introduces a hyperparameterised count of dynamic obstacles.
Implementing a temporal dynamic encoding module for multi-obstacle state and fusing it into both policy and value networks; aim is robust generalisation in cluttered, time-varying environments. Targeting T-RO / TMECH.

RL Locomotion & LiDAR Navigation for Quadrupeds

XJU Innovation Lab · Jun 2025 – Present

Trained PPO and N-P3O policies in Isaac Gym / Isaac Lab for multi-terrain quadruped locomotion, generalising across ≥ 3 terrain types.
Self-designed quadruped hardware; deployed HIMLoco MuJoCo sim-to-sim and finally sim-to-real, sustaining > 10 N pushes with stable gait control — walking 0.8 m/s, jogging 2 m/s, crawling — across 4 indoor / outdoor terrain classes and slopes.
SLAM with fast_lio and point_lio; ROS 2 Nav2 stack with A* / Dijkstra global planning and AMCL relocalisation for downstream task scenarios.
Outcome: 1 first-author paper accepted at ICIC 2026 (Oral).

Deep-Learning Strawberry Disease & Pest Monitoring

Xinjiang University · Jun 2024 – Jan 2026

YOLOv8 detection pipeline with attention modules and an improved feature-fusion neck — +15 % recall on small targets and +25 % precision; pruned and quantised for embedded deployment.
Inverse-kinematics solver for an arm mounted on a tracked rover; coordinated chassis-arm motion to acquire clear, comprehensive plant imagery from preset observation viewpoints.
Outcome: National Innovation & Entrepreneurship grant — Outstanding close-out · 1 design patent (1st author) · 1 software copyright (3rd author).

05Selected Honors & Awards

2025.08National Second Prize · National Undergraduate Electronic Design Contest (NEDC)

2025.07National First Prize · 24th ROBOCON — Bionic Quadruped Obstacle Race

2025.07National First Prize · 24th ROBOCON — Bionic Quadruped Off-road Race

2024–259 national-level + 3 provincial-level competition awards (in total)

2025National Innovation & Entrepreneurship Programme — Outstanding close-out

2024–Top 14.2 % of CS cohort (rank 17 / 120) · GPA 3.9 / 5.0