Hydra
Fuse mmWave SAR depth slices with camera images to sense wetness on real leaves, day and night.
≈96% accuracy indoors and ≈90% in the field, including rain, dawn and poorly lit nights.
Yimeng Liu
I am a Ph.D. candidate in Computer Science and Engineering at Michigan State University, advised by Zhichao Cao.
I study how intelligent systems understand the physical world from incomplete observations. My published work builds on multimodal sensing and learning in real environments. Before MSU, I studied mathematics at Virginia Tech and worked on biomimetic sonar with Rolf Müller.
Current focus
Spatial intelligence: systems that know what is where, in what state, how it is changing and what can be done about it—first for one observer, then for many observers sharing one space.
8 first-author papers · 3 published · 1 under review · 4 preprints
2023 —Ph.D. candidate, Computer Science and EngineeringMichigan State University
2019 — 2023B.S., MathematicsVirginia Tech
From a single sensor to a space that understands itself
Take one room: a plant, a table, a window. Machines have entered it in five steps. Select a step to see what each one added, and what it still could not do.
-
Machines start to measure the world.
Signals everywhere, but no understanding.
-
Learning turns signals into labels.
Recognition for one task, from one view, in one place.
-
AI learns to reason and explain.
Knowledge without a body: it has never been in this room.
-
Sensors, a model and a body close one loop.
One agent senses, decides and acts—alone.
sensors + model + robot = embodied intelligence
-
The space and its agents share one model.
The room as a whole can sense, understand and act.
environment + many agents + shared model = spatial intelligence
Spatial intelligence shapes our future.
Intelligence is moving from text into rooms, fields and roads. To be useful there, a system must know what is where, in what state, how it changes and what can be done—and, increasingly, know it together with other systems that share the same space.
“The truth would be literally nothing but the shadows of the images.”
- An inverse problem
- Every sensor sees a projection of the world: O = h(W) + ε. Understanding means asking which world explains all the projections—and what stays unobservable even when they are combined.
- From shadows to community
- Plato: we see projections. Gibson: we perceive in order to act. Aristotle and Xunzi: we became intelligent by living and knowing together.
- One loop, many scales
- Sense → understand → act, built across sensors, models, bodies and networks, and evaluated where its assumptions break.
Sense, understand, act: closing the loop for one observer
Each stage of the loop has its own failure mode. My first-author papers address one stage each; one lead paper per stage is highlighted.
-
1Sense
The state that matters is often hidden.
Cameras miss water and material; any single sensor fails somewhere.
-
2Understand
Sparse evidence admits many worlds.
Fast sensing means few measurements, and few measurements fit many shapes and meanings.
SPIRIT
Treat sparse-aperture radar as inverse rendering: reconstruct an explicit 3D surface through a physical model of visibility, phase and propagation.
Over 90% less data collection while keeping low geometric and signal error.
-
3Act
In the physical world there is no reset.
Actions change future states, errors compound, and models go stale.
Mímir
A physics-grounded LLM agent: language proposes, a simulator checks and repairs, and physical authority stays with fixed numerical mechanisms.
About 51% less irrigation than the historical schedule; no monotonic gain from larger language models.
Next: from one observer to shared spatial intelligence
Today’s intelligence lives on islands: a field sensor sees one leaf, a car one viewpoint, a robot one room, a language model words rather than places. One observer is a sparse aperture on the world. My next step is to let observers combine what they see, believe and do—without giving up local authority.
- Field sensorone leaf, one number
- Driver assistanceone car, one viewpoint
- Home robotone room, one task
- Language modelknows words, not places
1Observe together
When do several partial views determine what one cannot?
Share evidence across sensors and agents, and measure what becomes identifiable only when views are combined.
Challenge: alignment, timing and bandwidth across heterogeneous sensors.
2Believe together
How should shared beliefs be revised—and when should they not be?
Maintain a common model of the space with explicit uncertainty, admitting revisions only when evidence supports them.
Challenge: avoiding double-counted evidence and conflicting sources.
3Act together
Who may act on shared understanding?
Coordinate decisions across agents while each keeps physical authority bounded and local.
Challenge: safety, accountability and privacy when consequences are physical.
The bridge: AURA, ongoing
AURA reads one driving situation through several streams—driver behavior, vehicle state, road and weather context—over trips and years. Its next test is an occluded intersection: a roadside unit sees a pedestrian that one car cannot, another vehicle adds a view, and the car keeps the final decision.
AURAWhen is sharing better?
Not always. Sharing helps when observers complement each other, adds little when they are redundant, and can hurt when the same evidence is counted twice. Measuring which regime a space is in is part of the science.
Core papers & research
Selected work that shaped my current research direction.
-
2026arXiv'26
-
2026arXiv'26
-
2026Under review
Snotra: Continually Adapting Irrigation World Models under Physical Change
Manuscript under review (2026)
-
2026arXiv'26
-
2025SenSys'25
-
2025INFOCOM'25
-
2024MobiCom'24
Academic service & community
Service
Program committee
- IEEE DSAA 2025
Invited reviewer
- UbiComp / ISWC · UbiSense 2025
- IEEE Internet of Things Journal
- IEEE Transactions on Mobile Computing
- IEEE Internet Computing
- HAI 2025 · Poster track
Conferences & talks
- ACM SenSysPresented Proteus
- ACM MobiComPresented Hydra
- NSF Workshop on Sustainable ComputingParticipant
- Invited talk, University of Hawai'i at Mānoa — October 2024
Recent updates
The SPIRIT preprint is on arXiv: sparse-aperture SAR as physics-informed inverse rendering.
The Mimir preprint is on arXiv: a physics-grounded LLM agent for long-horizon irrigation control.
Our driving assessment preprint discusses continuous, contextualized and explainable assessment for older adults.
How I approach research
- Start with the question, not the model.
- Define what must be understood or decided before choosing how to model it.
- Keep the observation in view.
- Ask what the evidence supports, what it cannot distinguish, and which assumptions connect observation to physical state.
- Test where the assumption breaks.
- Evaluate where a system works, where its assumptions fail, and whether it can recognize when the evidence is no longer sufficient.
Get in touch
I welcome conversations about multimodal sensing, physical representations and reliable intelligent systems, especially where these questions meet a concrete measurement problem.
liuyime2@msu.edu