Yikai Wang

I am a Ph.D. Candidate at Carnegie Mellon University, advised by Prof.Ding Zhao.

Previously, I was a reserch assistant at the ISR Lab, IIIS, Tsinghua University, and a student intern at Carnegie Mellon University, where I worked with Prof.Jianyu Chen and Prof. Guanya Shi.

I also spent a wonderful summer as a research intern at Bosch AI (2025).

Email  /  Google Scholar

Research

APEX: Learning Adaptive High-Platform Traversal for Humanoid Robots

APEX enables adaptive humanoid traversal of high platforms via contact-rich climbing and a unified multi-skill policy. Leveraging a ratchet progress reward and LiDAR-based perception, the system adapts its behaviors to terrain geometry and approach conditions, achieving zero-shot sim-to-real traversal of 0.8 m (114% of leg length) platforms on a 29-DoF humanoid.

LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing

Equipped with a high-density distributed tactile sensor, our quadrupedal robot can transport everyday objects without mounting or strapping. The transport policy achieves zero-shot sim-to-real transfer, featuring two task-agnostic components: high-fidelity tactile simulation and robust, symmetric, frequency-adaptive gaits.

Human2LocoMan: Learning Versatile Quadrupedal Manipulation with Human Pretraining

Human2LocoMan is a unified framework for collecting human demonstrations and teleoperated robot whole-body motions, along with cross-embodiment policy learning for versatile quadrupedal manipulation.

Guardians as You Fall: Active Mode Transition for Safe Falling

We propose Guardians as You Fall (GYF), a safe falling/tumbling and recovery framework that can actively tumble and recover to stable modes to reduce damage in highly dynamic scenarios. The key idea of GYF is to adaptively traverse across different stable modes via active tumbling, before the robot shifts to irrecoverable poses. GYF offers a new perspective on safe falling and recovery in locomotion tasks, which potentially enables much more aggressive explorations of existing agile locomotion skills.

Learning Robust, Agile, Natural Legged Locomotion Skills in the Wild

We propose a new framework for learning robust, agile and natural legged locomotion skills over challenging terrain with only proprioceptive perception. We incorporate an adversarial training branch based on real animal locomotion data upon a teacher-student training pipeline for robust sim-to-real transfer. To the best of our understanding, this is the first learning-based method enabling quadrupedal robots to gallop in the wild.

Projects

Building an autonomous picking car from scratch

Design and built a car that is capable of autonomous navigation, picking and storing certain yellow objects and unloading them in the designated area. Target recognition was accomplished by a monocular camera and image algorithms. Picking and unloading were accomplished with 3 parallel elastic ropes and a steering engine. Velocities were adjusted using PID control. Mechanical parts were designed using Solidworks, and manufactured by 3D-printing/laser cutting.

Design and source code from Jon Barron's website