Narrative Audit — Phase 1

Date: 2026-09-06
Scope: research claims only; no visual or layout changes made in this phase.

Approved definition

Building Spatial Intelligence for the Physical World: enabling AI to move from incomplete multimodal observations to persistent representations of dynamic physical spaces, reason about how entities relate and evolve, predict what happens next, and support grounded action.

The public framework is:

Capability	Visitor question	Evidence boundary
PERCEIVE	What is there?	Published work establishes robust multimodal / wireless physical perception in agricultural sensing settings.
REPRESENT	Where is it and how is it related?	Published work establishes cross-modal knowledge transfer and physical-state reconstruction; a unified spatial relation representation remains an agenda.
MODEL	How is the world changing?	Published work estimates physical state and leaf wetness; persistent dynamic world modeling is not yet a published result on this homepage.
REASON & PREDICT	Why is it changing and what happens next?	Future research agenda; no published evidence is claimed.
ACT	What should the agent do?	Current human-centered driving safety work is an application bridge toward grounded action; it is not presented as a completed embodied-intelligence system.

Research trajectory

Published Foundations — what the previous work establishes

Yimeng’s published work forms a coherent foundation in physical observation and representation:

Reliable multimodal perception: Hydra combines mmWave radar and RGB camera information for leaf-wetness detection. Its paper reports up to 96% accuracy in evaluated plant scenarios and around 90% in rainy, dawn, or poorly lit farm conditions. Source: Hydra PDF.
Cross-modal knowledge transfer: Proteus uses an RGB teacher to guide a mmWave SAR student, with noise reduction and phase-angle features. Its paper reports up to 96.3% accuracy across varied environmental scenarios. Source: Proteus PDF.
Physical-state estimation: Adonis introduces Leaf Wetness Level and estimates fine-grained wetness from mmWave signals. It reports MAE 4.43 in controlled conditions and 6.49 in real farm conditions, compared with 11.84 and 14.32 for traditional sensors. Source: Adonis PDF.
Benchmarking and reproducibility: Hydra-Bench contributes synchronized mmWave raw data, SAR images, and RGB images collected over six months across five plant species and controlled/outdoor environments. Source: Hydra-Bench arXiv record.

These papers demonstrate sensing, cross-modal learning, and physical-state estimation. They do not demonstrate a general-purpose persistent world model, spatial-relation reasoning, future-state prediction, or a closed-loop embodied agent.

Current Research — what is being built now

The current public project page documents human-centered driving safety research: driving-simulator assessment, continuous in-car monitoring, realistic simulation tests, and community education. Source: Senior Driving project.

This is a credible bridge toward grounded action and safety-aware physical AI, but the current repository contains no public artifact supporting the stronger homepage claim of an LLM-grounded CARLA system. That claim must be removed or replaced until a verifiable project page, paper, codebase, or result is added.

Future Research Agenda — where the work is going

The long-horizon agenda is to turn multimodal observations into a persistent spatial state, maintain entity identities and relations over time, model change, reason about interaction, predict future states, and support grounded action. This is a research direction, not retroactive evidence for the published papers.

Project → capability claim matrix

Project	Capability it can honestly evidence	Maturity	Approved public claim	Do not claim
Hydra	Multimodal perception under adverse conditions	Published Foundation	mmWave + RGB complement one another for physical sensing when visual conditions degrade.	A general human/vehicle perception model; “first” unless independently documented.
Proteus	Cross-modal learning / transfer	Published Foundation	An RGB teacher guides a mmWave SAR student, transferring useful features across modalities.	A complete unified 3D spatial representation or a general world model.
Adonis	Physical representation / state estimation	Published Foundation	mmWave signal processing and contrastive feature extraction estimate fine-grained Leaf Wetness Level.	A dynamic world model, broad sparse-observation reconstruction, or reasoning system.
Hydra-Bench	Multimodal benchmark / research infrastructure	Published Research Artifact	Synchronized raw mmWave, SAR, and RGB data support evaluation across modalities and environments.	A flagship Spatial Reasoning result.
Senior Driving Safety / Cognitive Driving direction	Human-centered sensing and safety-aware intervention bridge	Current Research	Simulator assessment and in-car monitoring explore earlier, safer support for older drivers.	LLM-grounded CARLA reasoning, completed action policy, or published embodied intelligence.
Dynamic spatial world modeling	Persistent state, relation tracking, prediction	Future Agenda	A long-term research direction built on the sensing foundations.	A capability already demonstrated by Hydra, Proteus, or Adonis.

Copy audit

Keep

“Building Spatial Intelligence for the Physical World” as the identity-level H1.
The incomplete-observation → persistent-state → reasoning → action framing.
The five visitor questions, because they make the research problem operational.
Hydra, Proteus, Adonis, and the driving direction as distinct capability evidence.
Explicit Published Foundation / Current Research / Future Agenda labels.

Modify

Replace the current Cognitive Driving: LLM-grounded ... in CARLA sentence with language grounded in the public senior-driving project page.
Replace Proteus problem copy about slow/costly scanning with the paper-supported issue of mmWave SAR imaging quality and cross-modal guidance.
Replace Adonis “wider scan distances” / “sparse RF observations” wording with the verified Leaf Wetness Level and mmWave physical-state estimation result.
Add the quantitative results above to the evidence layer rather than relying on broad adjectives such as “faster, finer, and more robust.”
Keep leaf wetness in paper titles and evidence captions, but remove it from the top-level research definition and trajectory headline.

Remove or demote

Any “first” claim that is not directly supported by a citable source.
Any implication that the published sensing series already solved world modeling, spatial reasoning, prediction, or embodied action.
The unsupported LLM / CARLA claim until a matching public artifact exists.
Low-value homepage content that does not help the visitor understand identity, problem, demonstrated capability, current build, or future agenda.

Required 60-second interpretation

Yimeng Liu studies Spatial Intelligence: how AI can move from incomplete multimodal observations to persistent representations of dynamic physical spaces, understand relationships and change, predict what happens next, and ultimately support grounded action. His previous multimodal sensing and physical-state estimation work establishes the foundations; his current human-centered driving research provides a bridge toward safety-aware action; and his future research extends those foundations toward dynamic world models, spatial reasoning, prediction, and embodied intelligence.

Phase 1 acceptance

The definition is fixed and reusable across Hero, About, Vision, Featured Research, and future project pages.
Every flagship has a capability, maturity label, evidence boundary, and prohibited overclaim.
Published evidence is limited to claims supported by the local paper artifacts or verified public project / benchmark records.
Current and future work are explicitly separated from published results.
The unsupported LLM/CARLA driving claim is identified as a hard bug and scheduled for correction before Phase 4 evidence design.