Carnegie Mellon University · School of Computer Science
Ray Muxin Liu
I am an undergraduate at Carnegie Mellon University studying Artificial Intelligence in the School of Computer Science. I am a fourth-year student expected to graduate in December 2026, and I am interested in research at the intersection of robotics, embodied AI, and learning systems.
I am preparing to apply for PhD programs and am especially interested in dexterous manipulation, behavior generalization, and post-training generalist robot policies with real-world feedback. I want to build systems that can refine their behaviors through experience and, over time, support increasingly autonomous exploration and self-improvement.
Email:
- Program
- B.S. in Artificial Intelligence
- School
- CMU School of Computer Science
- Year
- Fourth Year · Expected December 2026
- Focus
- Robotics, Embodied AI, Manipulation
Research philosophy
Intelligence needs a body
I see robotics as the work of giving intelligence a body. Learning systems can reason over language and vision, but an intelligence that genuinely explores the world must also act in it, observe the consequences, and revise its understanding through physical experience.
I am motivated by the long-term possibility of embodied intelligence becoming a scientific partner: a system that helps extend how humanity investigates, understands, and learns from the physical world.
A working view
Discovery through interaction
I am interested in the intelligence that emerges when perception, reasoning, and action are treated as one connected process. A robot should not only describe its surroundings; it should be able to test hypotheses, use tools, and learn from the consequences of contact with the world. My current interests include post-training generalist vision-language-action policies with demonstrations, DAgger, and reinforcement learning so that useful behaviors can become more robust and eventually support continual self-improvement.
I write about the more technical side of this view—behavior generalization, simulation, real-world learning, and control—in the blog.
Papers
Selected papers
COWBOY: A Scalable Sim-to-Real Framework for Learning Contextual Whole-Body Manipulation
CoRL 2026 · Accepted (unreleased)
A scalable sim-to-real framework for contextual whole-body mobile manipulation, using local reinforcement-learning experts, student-teacher distillation, whole-body control, and simulation evaluation.
Bimanual Dexterity for Complex Tasks
Conference on Robot Learning (CoRL) 2024
A low-cost, highly dexterous bimanual teleoperation system for collecting data and training policies for complex multi-hand manipulation tasks.