I am a master's student at the University of Michigan School of Information and a former humanoid robot industrial designer, now working on whole-body control for humanoid robots and human-to-humanoid motion retargeting. At DeepCybo, I worked on whole-body control for Unitree G1 soccer, including human-and-ball motion capture, motion retargeting, reinforcement-learning-based motion tracking, sim-to-real deployment, and physical-robot shooting tests.
I am also a co-author of Human-as-Humanoid. I contributed to its motion-retargeting pipeline, which converts recovered human motion into controller-aligned 60-DoF action labels for the PrimeU upper-body humanoid robot. Together, these projects connect motion data, whole-body control, and physical deployment across two humanoid robot platforms.
Goal: Build whole-body control and embodied-AI systems that enable general-purpose humanoid robots to perceive, adapt, and act in the physical world.
Although my current work focuses on control algorithms for humanoid robots, my undergraduate training was in industrial design at the Academy of Arts & Design, Tsinghua University. I first worked on automotive interiors at Pan Asia Technical Automotive Center (PATAC), a 50-50 joint venture between General Motors and SAIC Motor that specializes in automotive design and engineering.
In early 2025, I moved into humanoid robot design at Noetix Robotics, where I worked on the industrial design of the Xiaonuo bionic humanoid robot. I later led the industrial design of the Prime full-size humanoid robot while working with Zhongguancun Academy and Zhongguancun Institute of Artificial Intelligence. When the project became DeepCybo, I was among its earliest team members and joined the founders in early investor meetings. Those conversations repeatedly returned to the same question: what intelligence would make the robot genuinely useful? That question convinced me to move from designing the robot's body to developing the algorithms that control it.
I entered this transition with undergraduate foundations in mathematics and C programming, a score of 145/150 on the mathematics section of China's national college entrance examination, and confidence in my ability to learn unfamiliar subjects. I studied machine learning through Professor Hung-yi Lee's Machine Learning course at National Taiwan University, then took EECS 545: Machine Learning at the University of Michigan. My final project, SpaRe-lite, explored spatial-representation rewards for sparse-reward reinforcement learning in the LIBERO simulation benchmark. On RedNote, I continue this process by explaining recent robotics and embodied-AI papers and tracing unfamiliar ideas back through their prerequisite literature. When I returned to DeepCybo in May 2026, my experience with reinforcement learning and my long-standing love of soccer led naturally to the Unitree G1 soccer project presented above.
An 8B physical foundation model that unifies embodied understanding, end-effector action generation, and future-state prediction within a shared autoregressive framework.
A supervision framework for humanoid robots that converts synchronized ego-exo human videos into controller-aligned 60-DoF action labels for PrimeU and high-DoF VLA policy training.
DeepCybo internship · 2026 - present
I am developing a soccer-motion pipeline for the Unitree G1, spanning synchronized human-and-ball motion capture, motion retargeting, physics review, reinforcement-learning tracking, ONNX export, and physical-robot validation. The current corpus contains 5.20 hours of raw capture. After quality control and independent segmentation, the V48 training set contains 104 recordings and 238 continuous segments (4 h 09 min, or 8 h 18 min with left-right mirroring).
Latest tracking result. A 30.26-second single-motion BeyondMimic acceptance run completed a full rollout, with 99.22% of final training episodes reaching the time limit. This stage evaluates the whole-body motion tracker without ball rigid-body dynamics.
Next step. Couple the learned whole-body controller with ball contact dynamics and egocentric perception so the robot can localize the ball and goal, adjust its support and approach, and execute skills in closed loop.
Goal-directed shooting replay with the retargeted G1 motion, ball trajectory, and contact diagnostics.
A G1 kinematic replay retaining the ball rigid-body trajectory, used to inspect the approach, support-foot transition, and shot timing.
A 50 Hz physics-review replay of the robot and ball reference. The inspected sequence contains no foot-ball penetration; it remains a reference trajectory rather than a closed-loop policy.
A BeyondMimic rollout with an independently simulated rigid-body ball. This is the strongest stage result from that training branch, although repeated foot-ball contacts are not yet stable across fixed-seed evaluations.
Six clips share this description: real-robot validation of the ONNX-deployed soccer-shooting policy from several camera views. The first two clips document the first shooting test, and the final vertical clip also includes the corresponding human reference. Drag or swipe horizontally to view all clips.
This work was conducted during an industry internship. Videos and high-level results are shared here; source code and implementation details are not publicly released.
Led an EECS 545: Machine Learning final project on offline reinforcement learning for VLA sparse-reward tasks. Built 409 simulation rollout transitions with failure samples and improved LIBERO-10 success rate from 16.13% after supervised fine-tuning (SFT) to 18.40%.
Led the industrial design of Prime, from its form language and exterior surfaces to packaging integration and the final realized prototype. This experience became the bridge from my design background to embodied intelligence and whole-body control.
Author of Chinese-language robotics research briefs for the Tsinghua MBA club focused on embodied intelligence, and writer for EID research reports, covering paper selection, summaries, and technical interpretation.
University of Michigan School of Information, Master of Science in Information.
DeepCybo, whole-body control for humanoid robots.
DeepCybo, Zhongguancun Academy, and Zhongguancun Institute of Artificial Intelligence, industrial design lead for the Prime full-size humanoid robot.
Noetix Robotics, industrial design for the Xiaonuo bionic humanoid robot.
Pan Asia Technical Automotive Center (PATAC), automotive interior design.
Academy of Arts & Design, Tsinghua University, B.A. in Industrial Design.
Design origins · 2025
Before moving into algorithms, I led the industrial design of Prime at DeepCybo's earliest stage. The work connected concept development, full-body form language, mechanical packaging, and the final physical prototype. It also made the direction of my next chapter clear: understanding and building the intelligence that gives a humanoid robot useful motion.
Complete design archive
89 distinct images · click to expand
89 distinct images in 6 groups · swipe horizontally; click an image to open the original
Noetix Robotics · 2025
At Noetix Robotics, I contributed to the industrial design of the Xiaonuo bionic humanoid robot. My work included the upper-body product identity, torso and neck surfacing, face-and-screen integration, and manufacturable exterior geometry. This was my first direct step from automotive design into humanoid robots.
Design gallery
12 images · click to expand
12 images · swipe horizontally; click an image to open the original
Official product portrait · open original ↗
Official profile concept · open original ↗
Official livestream demonstration · open original ↗
Official product image · source ↗
Official side-profile product image · source ↗
Official front product image · source ↗
Official facial-detail image · source ↗
Development-stage prototype in the studio · open original ↗
Final product and upper-body design · open original ↗
Public stage demonstration · open original ↗
Product display at iROS Hangzhou 2025 · open original ↗
Three-quarter surface development · open original ↗