Robotics · Machine Learning · Cognitive Science
Liang (Sid) Yi
I work on tactile perception, dexterous manipulation, multimodal robot learning, and real-robot adaptation. My focus is on building learning systems that can improve through interaction, tactile feedback, and accumulated experience.
I received my bachelor's degree in Photoelectronic Science and Technology (basically laser stuff :) ) from Southwest University of Science and Technology in China, and my master's degree in Sensor and Cognitive Psychology from Chemnitz University of Technology in Germany, under the supervision of Prof. Dr. Wolfgang Einhäuser-Treyer from the Physics of Cognition group and Prof. Dr. Peer Neubert from the Autonomous Systems group at the University of Koblenz.
I am broadly interested in the intersection of machine learning and human–robot interaction. My long-term goal is to build robots that can learn from human interaction and continue improving through their own experience. I am particularly interested in bio-inspired learning and insights from cognitive science.
Intelligence lives in the loop, not in a single model.
How can robots turn incomplete perception into grounded action—and keep correcting themselves as the world changes?
From models to operating systems
Research that survives contact with reality
I work across algorithms, infrastructure, and teams because reliable robot behavior is a systems problem.
Selected work
One agenda, expressed at different layers
Tactile intelligence at the contact layer, learning systems at the policy layer, and recoverable runtimes at the task layer.
Tactile Dexterous Manipulation & VTLA Systems
Hierarchical visual–tactile–language–action systems for contact-rich manipulation, combining multimodal state understanding with low-level tactile correction and real-robot evaluation.
Read the OmniVTLA paperRARK: Robot Agent Runtime Kernel
A recoverable, preemptible, and auditable runtime for preserving task continuity across planning, execution, interruption, failure, and resumption.
Explore the projectJanus: Embodied AI Learning Framework
An NVIDIA Isaac Sim–based learning framework that unifies simulation, multimodal data, training, evaluation, and sim-to-real workflows for rapid robot-learning experiments.
Alembic / Paper Explorer
A research-intelligence workspace that turns a paper library into a structured contribution map through reliable background analysis, strict data contracts, and a visual Contribution Atlas.
Research agenda
From raw sensation to sustained agency
My projects are different views of the same problem: how physical agents build useful state, anticipate consequences, and remain capable of correction over time.
Ground state in contact
Learn tactile and multimodal representations that expose contact state, affordance, slip, and uncertainty—not merely sensor values.
Learn consequences on real robots
Combine imitation learning, reinforcement learning, world models, and human correction to improve policies under real hardware constraints.
Keep tasks alive through failure
Design runtimes that make long-horizon robot behavior recoverable, explainable, auditable, and safe to interrupt.
Embodied cognition · Active inference · Object-centric representation · Neuro-symbolic systems · Vector symbolic architectures
Journey
A path through disciplines
Each transition added a missing layer: physics for the world, psychology for perception, machine learning for adaptation, and robotics for consequence.
- Foundations
Photoelectronic Science & Experimental Thinking
Trained in optics, electronics, measurement, and the discipline of testing ideas against physical systems.
- 2019–2023
Sensor & Cognitive Psychology · TU Chemnitz
Studied how humans integrate perception, context, and prior knowledge; investigated visual geo-localization from the human perspective.
- 2022–2023
Reinforcement Learning · Real Robot Sim2Real
Led the AI development of a robotic-arm juggling system, including reinforcement learning, perception, and sim-to-real transfer.
- 2024–Present
Tactile Intelligence & Dexterous Manipulation · PaXini Tech
Led algorithm teams spanning tactile perception, multimodal data, dexterous manipulation, real-robot learning, and shared embodied-AI infrastructure.
- Next
A New Chapter, Taking Shape
Moving closer to where embodied intelligence, product direction, and team-building converge. More soon.
Field notes
Selected milestones
A short record of capability shifts—from a single real-robot policy to multimodal systems, shared infrastructure, and industry-scale data.
-
Led multiple industrial embodied-AI programs—including a 15-SKU picking-and-delivery system for Shinwa and Toyota-related manufacturing environments, a P&G proof of concept, and a food-service robot with PaXini × 一号农场—while leading a roughly 12-person dexterous-manipulation and multimodal algorithm team.
-
Spent 2025 building the multimodal collection, replay, and quality-assurance infrastructure behind Super EID Factory and OmniSharing DB, scaling real-robot data validation to thousands of object types before the work entered public view in 2026.
-
Contributed technical recommendations to multiple national and international standards drafts, translating real embodied-AI data-pipeline experience into dataset and evaluation requirements.
-
Supported the OmniVTLA research program behind the scenes as its vision–tactile–language–action architecture and semantic-aligned tactile sensing were released on arXiv.
-
Placed second in the visual–tactile grasping track at the Shenzhen Intelligent Robot Dexterous Hand Competition.
-
Led the AI stack for a real-robot reinforcement-learning demonstration at WRC 2023, covering perception, policy learning, and sim-to-real transfer for robotic-arm juggling.
Earlier work
Human perception and robot learning
Visual Geo-Localization with Knowledge Graphs from the Human Perspective
Investigated how people solve visual geo-localization tasks and how those cognitive strategies can inform machine representations and retrieval.
Reinforcement-Learning Sim2Real Robot Juggling
Led the end-to-end AI stack for robotic-arm ping-pong juggling: reinforcement learning, computer vision, and sim-to-real transfer.
Me · North Star
Me
I enjoy reading books, watching movies and dramas, and gaming — mostly sci-fi, scene comedy, philosophy, and historical dramas. I'm drawn to new experiences, so I'm always learning something different (recently kendo, sign language, and French). I love to travel and collect different perspectives to draw my own conclusions. 01 When I was 7 or 8, I was drawn to reading sci-fi novels and short stories. Later, watching The Big Bang Theory sparked something deeper — Leonard's character influenced my professional path in ways I didn't expect. Combining my passion for sci-fi with the academic template from the show, I entered college studying laser physics, dreaming of becoming an experimental scientist. The opportunity to study physics in Germany felt like living Sheldon's dream. But somewhere between the equations and lab work, I discovered something unexpected: I was more fascinated by the human side of physics than the pure science itself. That realization led me to an interdisciplinary major — a decision that would reshape how I think about everything. 02 Physics taught me to see the universe as equations dancing in perfect harmony. But equations, I discovered, don't capture the tremor in a voice or the weight of a decision. They don't explain why humans reach for the stars or why we fear the very technologies we create. Leaving pure physics felt like stepping out of a cathedral of certainty into a wilderness of questions. In this interdisciplinary space, I learned to think like a weaver — pulling threads from physics, psychology, philosophy, and beyond, watching them intertwine into patterns I never could have imagined alone. The most beautiful discoveries, I realized, don't live within the neat boundaries of disciplines. They emerge in the liminal spaces, where different ways of seeing the world collide and spark something entirely new. It's in these intersections that I found my calling — not just to understand how things work, but to understand how they can shape us. 03 As a sci-fi fan turned researcher, I've moved from consuming visions of the future to actively shaping them. We're not riders of the tech wave — we're the ones who shape its direction. That's both our privilege and our burden. Standing at this unique position of defining cutting-edge technology, I believe we have a responsibility to choose our path carefully. While most people can only passively accept the results of technological development, we as definers have both the opportunity and obligation to pursue a more human-centered technological trajectory. I want to build AI and robots that embody the future we choose, not the one we fear. Systems that evolve smartly through interaction, measured not by efficiency alone, but by how they elevate the human experience. Because technology should serve what humans truly need: not just survival, but dignity. This vision guides every research decision I make and every line of code I write.The Big Bang Theory
Between Worlds
A Human-Friendly Future
Build, research, collaborate
Interested in real-world robot intelligence?
I am open to conversations about embodied AI, tactile dexterous manipulation, research collaboration, technical leadership, and industry-aligned doctoral work.
