Robotics Researcher · Zhejiang University

Li Hongyi 李弘毅

Advancing embodied intelligence through perception, learning, and action in the physical world.

Ph.D. student at Zhejiang University, working on learning-based locomotion, whole-body control, and sim-to-real transfer for legged and humanoid robots.

View research About me

Black-and-white cat and brain illustration

Selected work

All publications

PGMT2026 · Preprint

PGMT: Perceptive General Motion Tracking for Humanoid Robots

Flat-ground trackers struggle when the terrain changes. PGMT adapts diverse whole-body motions to stairs and uneven ground, with locomotion, teleoperation, and fall recovery in one policy.

PGMT humanoid motion tracking across stairs and outdoor terrain.

A single policy deployed on Unitree G1 supports terrain-adaptive locomotion, whole-body motion tracking, teleoperation, and fall recovery.

Paper Project YouTube GitHub

TAGA2026 · Preprint

TAGA: Terrain-aware Active Gaze Learning for Generalizable Agile Humanoid Locomotion

Narrow beams, gaps, and sparse footholds leave little room for error. TAGA learns to focus on the terrain that matters next, enabling agile humanoid locomotion with onboard sensing.

TAGA humanoid traversing gaps, beams, and sparse footholds.

An active gaze module uses vision, proprioception, and motion commands to select terrain patches, which a cross-attention encoder fuses into the locomotion policy.

Paper Project YouTube

KiViIROS 2026 · Accepted

KiVi: Kinesthetic-Visuospatial Integration for Dynamic and Safe Egocentric Legged Locomotion

Glare and occlusion can undermine visual locomotion. KiVi combines vision with body sensing so a quadruped can navigate uneven terrain and stay stable when visual information degrades.

KiVi quadruped maintaining locomotion across varied terrain and visual disturbances.

A dual-branch estimator keeps proprioception as a stable base and uses memory-enhanced visual attention to add terrain information to the locomotion policy.

Paper Project YouTube

Collaborative work

SWAP2026 · Preprint

SWAP: Symmetric Equivariant World-Model for Agile Robot Parkour

Using symmetry in learned world models and control policies for agile quadruped parkour.

Paper

SWAP Figure 1: quadruped gap jumps, platform climbs, and outdoor parkour

STARTRA-L 2026

START: Traversing Sparse Footholds With Terrain Reconstruction

Reconstructing local terrain from onboard vision for precise foot placement across sparse footholds.

Paper

START Figure 1: sparse-foothold locomotion in simulation and the real world, with reconstructed terrain

HeLoM2025 · Preprint

HeLoM: Hierarchical Learning for Whole-Body Loco-Manipulation by a Hexapod Robot

Coordinating hexapod locomotion and foreleg manipulation to move objects toward desired poses.

Paper

HeLoM Figure 1: a hexapod pushing objects with its forelegs in simulation and real-world experiments

SATARSS 2025

SATA: Safe and Adaptive Torque-Based Locomotion Policies Inspired by Animal Learning

Learning torque-based locomotion inspired by animal movement for compliant, adaptive control and zero-shot transfer to real robots.

Paper

SATA Figure 1: compliant quadruped interaction with a person and simulated locomotion during training

Full publication list

2026
PGMT: Perceptive General Motion Tracking for Humanoid RobotsarXiv preprint
Paper ↗
2026
TAGA: Terrain-aware Active Gaze Learning for Generalizable Agile Humanoid LocomotionarXiv preprint
Paper ↗
2026
SWAP: Symmetric Equivariant World-Model for Agile Robot ParkourarXiv preprint
Paper ↗
2026
START: Traversing Sparse Footholds With Terrain ReconstructionIEEE Robotics and Automation Letters · 11(2): 2194–2201
Paper ↗
2026
KiVi: Kinesthetic-Visuospatial Integration for Dynamic and Safe Egocentric Legged LocomotionIROS 2026 · Accepted
Paper ↗
2025
HeLoM: Hierarchical Learning for Whole-Body Loco-Manipulation by a Hexapod RobotarXiv preprint · Updated in 2026
Paper ↗
2025
SATA: Safe and Adaptive Torque-Based Locomotion Policies Inspired by Animal LearningRobotics: Science and Systems
Paper ↗

My path in robotics.

My work has grown from automation and robot systems to policy learning, perceptive locomotion, and whole-body control. Throughout, I focus on making learning work on real robots.

2025 — Present: Zhejiang University
Ph.D. student · Robotics
2023 — 2025: National University of Singapore
M.Sc. in Robotics
2018 — 2022: Beijing Institute of Technology
B.Sc. in Automation

Earlier engineering experience

Tsinghua University · AIR

Robotics Engineering Intern · Sep 2022 – May 2023

Worked on robotic arm control and robot system integration, including C++ / ROS development, simulation, and coordinated multi-arm demonstrations.

Institute of Process Engineering, CAS

Embedded Systems Intern · May – Jul 2023

Contributed to STM32 development and sensing and control circuit design, integrating hardware and software for practical systems.

Get in touch

For conversations about robot learning and control.

hongyi.li@zju.edu.cn

Li Hongyi 李弘毅

Selected work

PGMT: Perceptive General Motion Tracking for Humanoid Robots

TAGA: Terrain-aware Active Gaze Learning for Generalizable Agile Humanoid Locomotion

KiVi: Kinesthetic-Visuospatial Integration for Dynamic and Safe Egocentric Legged Locomotion

Collaborative work

SWAP: Symmetric Equivariant World-Model for Agile Robot Parkour

START: Traversing Sparse Footholds With Terrain Reconstruction

HeLoM: Hierarchical Learning for Whole-Body Loco-Manipulation by a Hexapod Robot

SATA: Safe and Adaptive Torque-Based Locomotion Policies Inspired by Animal Learning

My path in robotics.

Zhejiang University

National University of Singapore

Beijing Institute of Technology

Tsinghua University · AIR

Institute of Process Engineering, CAS

Get in touch