Jialei Huang

Jialei Huang_

Ph.D. Candidate · Tsinghua University, IIIS

I build generalizable and robust embodied intelligence. My work spans the full pipeline from perception to physical action—spatial understanding, feature extraction, control grounding, semantic alignment, and precision execution—translating theoretical breakthroughs into real-world robotic systems.

Looking forward, my vision is the physical landing of general-purpose robotics: mastering the entire end-to-end system pipeline from hardware-software co-integration to high-level reasoning.

01

Research

My technical philosophy is "Layered Knowledge Scaffolding across the Data Pyramid"—constructing a unified pipeline through five pillars:

ICRA 2025 2025

SaTA: Spatially-anchored Tactile Awareness for Robust Dexterous Manipulation

Pillar 5 — Physical Landing & Execution

Jialei Huang, et al. · Sharpa & Tsinghua University

Led the end-to-end architecture anchoring tactile features to hand kinematics for millimeter-level geometric reasoning. 88.3% success rate vs. 53.3% baseline on extreme-precision tasks.

Preprint 2024–25

Tactile-VLA: Unlocking VLA Models' Physical Knowledge for Generalization

Pillar 4 — Semantic-Level Alignment

Jialei Huang, et al. · Tsinghua University

Directed a framework grounding internet-scale VLA models through tactile sensing for zero-shot, contact-rich manipulation. Pioneered Tactile Chain-of-Thought reasoning.

ICRA 2023–24

BSR: Decoupled Visuomotor Manipulation via Sim-to-Real Transfer

Pillar 3 — Control-Level Grounding

Jialei Huang, et al. · Tsinghua University & UC Berkeley

Engineered a decoupled visuomotor pipeline harnessing low-level physics control from scalable simulation data. Achieved zero-shot sim-to-real transfer.

ICML 2022–23

Policy Contrastive Imitation Learning (PCIL)

Pillar 2 — Feature-Level Extraction

Jialei Huang, et al. · Tsinghua University & HKUST

Solved fundamental bottlenecks in adversarial imitation learning by designing an interpretable contrastive representation space. New SOTA on DeepMind Control suite.

NeurIPS 2020–21

Generative 3D Part Assembly via Dynamic Graph Learning

Pillar 1 — Spatial Understanding

Jialei Huang, et al. · Peking University & Stanford University

Designed a dynamic graph learning framework for 3D assembly, establishing foundational spatial and geometric reasoning.

02

Education

2021 — Present

Tsinghua University, IIIS

Ph.D. Candidate in Computer Science

Advised by Prof. Yang Gao. Research focuses on advancing embodied intelligence through deep reinforcement learning, dexterous manipulation, tactile sensing, and sim-to-real transfer.

2017 — 2021

Peking University, EECS

B.Sc. in Computer Science

Advised by Prof. Hao Dong. Developed rigorous foundations in machine learning and computer vision.