XYXintong YangPhysical Learning · Resilient Robotics · Robotic Science

Research programme

Physical Learning, Resilience & Robotic Science

My research asks how robots can learn from limited physical experience, understand and adapt to their own changing embodiment, and use those capabilities reliably in scientific and other consequential systems. The programme is organised around three connected levels: world, self, and system.

Back to profile · Research demos · Google Scholar

WORLDLearn what is physically true.
SELFKnow what the robot can currently do.
SYSTEMAct safely with people toward consequential goals.

Three research directions

01

WORLD

Experience-Efficient Physical Learning

How can robots learn actionable physical knowledge from limited continual experience, recognise what remains uncertain, and choose interactions that resolve the uncertainty that matters?

Real-world learningPhysical identificationInformative interactionDifferentiable physics

Explore this direction →

02

SELF

Self-Aware Resilient Robotics

How can a robot establish operational knowledge of its own body, sensors, actuators, interfaces and capabilities, then detect and adapt when that embodiment changes or degrades?

Embodiment discoveryCapability reasoningDamage adaptationBounded autonomy

Explore this direction →

03

SYSTEM

Trustworthy Robotic Science

How can robots, AI, mechanistic models and scientists conduct auditable physical experiments that remain safe, resilient and scientifically useful when evidence or equipment deviates from plan?

AI for scienceHuman in the loopLaboratory roboticsResilient experimentation

Explore this direction →

Cross-cutting research principles

These are not separate directions; they are the methodological commitments that connect the programme.

A

Learn from scarce real-world experience

Physical data are costly, slow and consequential. I focus on extracting useful knowledge from the interactions a real robot can realistically acquire.

B

Make physical beliefs testable

Interpretability should connect internal knowledge to physical causes, uncertainty and actions that can confirm or falsify a hypothesis.

C

Design for bounded, resilient autonomy

Autonomy should validate assumptions, monitor failures, recover where justified, and return authority to a human when evidence no longer supports autonomous action.

Research activity

Grants and projects

  • 2026-12 to 2027-08 — UKRI Network+ | DisCouRSE — Developing dRTP Leadership for Reliable AI-Driven Autonomous Laboratories
    Role: Co-I, Lead Researcher | Total value: GBP 14,693.09
  • 2024-07 to 2024-08 — Cardiff-Xiamen Collaboration Fund
    Role: Co-I | Total value: GBP 4,910
  • 2023-2025 — PHYDL: Physics-Guided Differentiable Learning for Robotic Manipulation of Deformable and Granular Materials
    Role: Research Associate | PI: Ze Ji | Funder: EPSRC New Horizon
  • 2025-now — Almond: Autonomous Laboratory for Molecular Discovery
    Role: Research Associate | PI: Oliver Castell; Co-I: Ze Ji, Jin Li | Funder: BBSRC

Supervision

  • Minglun Wei — Co-supervised PhD student, Cardiff University (ongoing) — deformable object manipulation
  • Boliang Cai — Former co-supervised PhD student, now Research Associate at Loughborough University — autonomous vehicles

Selected collaborators

Prof Ze Ji · Honorary Professor, Cardiff University

Robotics and autonomous systems, AI for robotics, computer vision

Prof Yu-Kun Lai · Professor, Cardiff University

3D geometry, computer graphics, AI

Dr Oliver Castell · Reader, Cardiff University

Synthetic biology, single molecule science, artificial cells

Dr Jin Li · Senior Lecturer, Cardiff University

Microfluidics and nanofluidics, synthetic biology

Dr Jing Wu · Lecturer, Cardiff University

Computer vision, AI