Percy Liuresearch notes

He/Him · @TheCYPER

Che (Percy) Liu

Student researcher and builder working on AI agents, game technology, and developer tools.

MBZUAI undergraduate Abu Dhabi, UAE / Beijing, China

Abstract screenprint collage of motion paths, connected nodes, and layered research windows in indigo, yellow, teal, and violet.
Research field / 2026Agents · motion · systems
Portrait of Che (Percy) Liu working at a computer in a study space.

I'm an undergraduate at MBZUAI and a builder across AI agents, games, robotics, and full-stack systems. I like turning uncertain technical ideas into working tools, then documenting what worked, what failed, and what remains unproven.

Open to research collaborations at MBZUAI during Fall 2026 and AI systems internships for Summer 2027; based in Abu Dhabi and Beijing, and open to remote opportunities or roles in Japan and the US.

Questions under active investigation

Current questions

  1. Hardware Agent Harness

    How can natural-language tasks become constrained, testable, and auditable robot skills that move from simulation through safety gates to real hardware?

    active

  2. KIMODO Martial-Arts Adapter

    Which data, training, and evaluation loops can extend a motion foundation model toward martial arts while exposing failure as clearly as success?

    active

  3. Agent-native / Zero-people Company

    If autonomous agents become economic participants, how might labor, ownership, productivity, and the distribution of wealth change?

    active

Evidence, not a publication list

Selected work

  1. Case Study

    KIMODO capability evaluation: prompts, constraints, and boundaries

    A case-based evaluation of where a motion foundation model helps, where constraints help, and what the evidence still does not establish.

    KIMODO prompt and constraint motion fieldAbstract articulated poses follow a curved trajectory between sparse anchors, showing a prompt guiding a motion draft while constraints shape selected moments.
    1. Prompt
    2. Sparse constraints
    3. Motion draft
    A prompt selects a motion prior; sparse constraints shape selected moments without supplying persistent scene or object state.
    Read

  2. Case Study

    AnimGen: from 1,312 valid clips to runtime pose generation

    A bounded UE 5.8 experiment that connected data filtering, model training, and a single-character Anim BP runtime proof of concept.

    AnimGen latent motion loopAnimation frames enter a compact latent loop and return as a continuous runtime pose sequence, with ghosted poses indicating transitions between frames.
    1. Clips
    2. Latent loop
    3. Runtime pose
    Filtered animation clips are encoded into a latent representation, evolved frame by frame, and decoded into a runtime pose sequence.
    Read

Education

  1. Expected May 2029

    Mohamed bin Zayed University of Artificial Intelligence

    B.Sc. in Artificial Intelligence

    Undergraduate · GPA 3.82/4.00

Beyond the build log

A personal note

Outside research and engineering, I keep returning to music and improvisation, philosophy, history. I am especially interested in conversations that connect technical systems with how people learn, create, and organize.

The easiest way to reach me is che.liu@mbzuai.ac.ae. You can also find my public work on GitHub and connect on LinkedIn.