SheetMN-00
TitleArchitecture & Thesis
Issue2026.08
Scopereal → sim → real

Get the Physics Right, and the Machine Stops Fighting It

microNature builds the real-world groundwork physical intelligence runs on. Instead of bending hardware to fit a simulator, we work with the environment and the robot's own embodiment exactly as they physically are — never simplifying one to fit the other. Every bit of friction removed at the physical level is brute force the robot no longer has to supply.

Domain-Native Physical Intelligence Infra Scene-native · 100+ deliveries in specialized industries In-house physics engine + spatial intelligence
Thesis A better model alone won't close the gap between what works in the lab and what works on site. Closing that gap takes infrastructure grounded in the real physical world. That's what we're here to build.
00Contents
01Architecture

microNature · Real-Sim-Real

microNature runs the full real → sim → real loop. Take the whole thing, or just the real-sim half. At the centre sits a domain-native physical scene that any kind of intelligent hardware can plug into, keeping the path from training through validation to rollout lightweight.

TASK & BODY INPUT LLM TASK task encoding · 0.1–1 Hz ROBOT FILES platform digital twin · URDF REAL real site · domain-native vision lidar thermal mmWave tactile force multi-modal capture of physical quantities PHYSICS-NATIVE SIM microNature PHYSICS ENGINE scene-native · geometry + physics + dynamics FULL-DOMAIN PERCEPTION full-domain fusion · runs on physical quantities EA · ENVIRONMENT AFFORDANCE affordance annotation · adversarial validation · scoring lightweight training · validation DEPLOY heterogeneous intelligent hardware legged wheeled UAV arm AV machines that pass the exam go on site R2S S2R DATA FLYWHEEL co-built with the industry domain · scene data ↔ validation standards · both ends accumulate DEPLOYMENT FEEDBACK PLUG-IN STACKS · swappable SENSING STACK any sensing stack plugs in POLICY STACK any policy stack plugs in microNature — real-sim-real · a lightweight deployment across training, validation and deployment
FIG.01microNature architecture · real-sim-real

The chain works as one piece, and it breaks down cleanly into three products.

M1 · REAL→SIM

We digitise a real scene and rebuild it physics-first. You get a reusable digital proving ground: the 3D base map is built once, and every platform after that reuses it.

M2 · EA VALIDATION

Environment Affordance™ delivered on its own. We annotate affordances against your platform specs, run adversarial validation, and hand back scores and samples.

M3 · R2S2R FULL CHAIN

Lightweight training → validation → deployment, closed end to end. Built together with the industry, so what happens on site flows back and the data flywheel keeps turning.

Core component Environment Affordance™
Where robots and intelligent equipment of every kind get aligned, rehearsed and cleared for entry

EA, Environment Affordance, is what sets the microNature architecture apart. We validate from the robot's point of view and from the task, inside a reusable, physically real scene. The question is not what a general-purpose robot can do. It is what this robot can and cannot do on this site, and the answer comes back as a score and a report.

REUSABLE REAL SCENE reusable real scene ROBOT BODY robot platform specs ALIGNED TASK the aligned task EA ENGINE robot's point of view · task-aligned affordance computation · perception adversary closed-loop scoring SCORE REPORT ADMISSION exam before deployment
FIG.02Environment Affordance™ validation loop

Cost ↓

Not every job needs a general-purpose robot. Validate the job on hardware you already own, and your robot budget comes down.

Interoperability ↑

Mixed fleets are validated in one proving ground against one standard, so they work better alongside whatever is already on site.

Threshold ↓

You do not need to understand models. You need to read a score. That is about as low as the barrier gets.

Admission →

One proving ground across domains. Whatever the field, the robot sits the exam before it goes in.

Product

Environment Affordance™ — independent validation before deployment

We take your platform specs, re-annotate the real site as an affordance map, prioritize the scenes most likely to mislead perception systems, and issue an auditable, reproducible verdict before deployment starts. Available on its own, and also the core component of the full microNature chain.

Open the product sheet →
02Difference

A Domain-Native Path, Seen from the Robot's Point of View

Scaling laws for language models (scaling law) sit on thirty years of internet data infrastructure. Scaling physical intelligence needs infrastructure of its own. Without a data foundation drawn from real environments, sim2real alignment has nothing to stand on. That leads to two answers: a different technical path, and different trade-offs.

Point one · technical pathWe fit the simulation to the domain, not the hardware to the simulator

Robots do not end up working in synthetic scenes. They end up in plants, mine sites, campuses, substations and warehouses. General-purpose approaches tend to miss one fact here. Seen from the robot and from the job, different working environments need different physics engines. Low light, dust and explosion-proofing down a mine drift. Cleanliness and geometric precision in a cleanroom. Unstructured haul roads in an open pit. The physical properties, the sensing modalities and the motion constraints are all different. Force every scene into one template and the physics comes out wrong.

Two of our core capabilities exist for exactly this reason (for the task spectrum see FIG.03).

Physics engine · scene-native

We model each domain's working environment physics-first, with geometry, physical parameters and dynamic behaviour built as one thing. Physics is not bolted on after the geometry is finished. It grows with the scene from the start. Scene-native means the simulation behaves the way the real site behaves.

Full-domain sensing · fusion of physical quantities

Different environments call for different ways of capturing physical quantities. Vision in one place, temperature and humidity in another, millimetre wave, touch, force control. Our full-domain sensing fuses all of them and runs them inside one physics engine. Physical quantities in, physical quantities out, nothing collapsed along the way.

Domain-Native We lead with domain-native physical scenes that any intelligent hardware can engage with, robots of every kind included. The simulation takes the hardware on the domain's terms, rather than asking the hardware to accommodate the simulator. That is the main thing separating us from a general-purpose simulator.

Point two · how we handle the trade-offsFrom the robot's point of view, at the pace of industry

The robot's view: know your limits

Not every job needs a top-of-the-line general-purpose robot. From the robot's side, finishing the job and staying operational matter more than being able to do everything. Line up the job, the site and what the platform can actually do, and a stripped-down, low-cost, purpose-built robot with the right infrastructure behind it will often beat an expensive general-purpose one.

Take a mine. What it actually needs is autonomous vehicles, quadrupeds, drones and wheeled robots working together, not one embodied robot handling everything. Environment Affordance™ validates from the robot's point of view, so every piece of hardware already on site gets used for what it is good at.

The pace of industry: development runs in order

The usual position is that you solve highly structured tasks first on today's infrastructure, semi-structured next, and general capability last. That order does not reverse.

Most of what manufacturing and industrial sites urgently need today sits between structured and semi-structured. Layouts change slowly, and physical accuracy, hardware coordination and cost control carry the most weight. Infrastructure should take this stretch on first. It is the largest, and it is the most certain.

STRUCTURED repeatable SEMI-STRUCTURED bounded variation UNSTRUCTURED open-ended / general industry's main demand sits here already solved at scale density 470 per 10,000 workers · IFR 2023 scenes change infrequently · physical accuracy hardware coordination · cost control frontier research · data gap >99% not yet urgent structured first → semi-structured next → general after that
FIG.03the task spectrum, and where industry's demand currently sits
03Proof

Proven Technology, New Application

microNature is a new idea built on mature technology. We hold a rare capability in physics engines and spatial intelligence, backed by a full in-house team, and all of it has been through real industrial delivery.

INPUT · 10+ years of industrial delivery OUTPUT · a proving ground at the application layer ··· MATURE FOUNDATION 01 Physics engine and spatial base GIS · 10+ years of 3D simulation TB-scale loading · domestic CPU 02 EA algorithms and adversarial strategy red vs blue · Deception Library geospatial algorithms carried over 03 Perception, visualisation and manipulation heterogeneous quantity fusion physical-quantity interfaces · one stack throughout 04 Edge compute R&D FeFET compute-in-memory edge inference · low power PROVEN → APPLIED field-proven · translated into applications APPLICATION INNOVATION pre-align rehearse admit → Environment Affordance™ application-layer innovation · digital proving ground new idea · affordance annotation · adversarial validation · scoring CAP-1 · SPATIAL BASE large-scene databases fused from many sources (oblique imagery · point cloud · BIM) fast TB-scale 3D loading · every domestic CPU architecture 10+ years of delivery in specialized industrial settings CAP-2 · ADVERSARIAL STRATEGY affordance annotation · Sensor Deception Library · red-blue adjudication geospatial algorithms turned into tools for testing perception limits red vs blue · field-proven CAP-3 · THREE LAYERS, ONE STACK full-domain sensor fusion · large-scale 3D visualisation physical-quantity control interfaces · top-tier across all three layers perception · visualization · manipulation CAP-4 · EDGE COMPUTE research on FeFET ferroelectric compute-in-memory chips latency and power · addressed at the hardware level compute-in-memory · edge inference
FIG.04mature foundation → application innovation
Two hard numbers Two numbers show how real this is. The gap between how impressive a product looks and how well it works on a real site is still widening. Meanwhile the shortfall in high-quality embodied data runs past 99 per cent, and physical intelligence has none of the ready-made data infrastructure language models had when they took off. The full argument, covering token and compute efficiency, the four data paths taken apart, and the cost of teleoperation, is in Insight · the capability–deployment paradox.
Sheet MN-02 · Team

See the full team

Urban regeneration, agricultural technology, cognitive neuroscience, compute-in-memory, optical switching. Where the team comes from, and what it works on.

About us →
04Now & Next

Closing the Gap, Building Intuition

NOW

Right now we are building the layer of infrastructure that sits between a demo and a deployment. The gap comes from the absence of real physical infrastructure, so we are not chasing larger models. The moat builds along the way, as scene data accumulates and validation standards take hold.

NEXT

Later, we give robots the physical intuition their domain and task demand — built into the infrastructure itself, not something the robot has to acquire on its own. A robot entering an unfamiliar site no longer has to learn the world from nothing. The infrastructure has already handed it the intuition.

Platform Vision We intend to become the layer every robotic agent connects to — in effect, a task-environment ontology. It assigns the task, the robot executes adaptively, and any sensing or policy stack can plug in. Every agent gets intuition, a task and a score. A robot that connects knows what it can and cannot do here. That is what delivering physical intuition looks like when it is complete.
Why this is unavoidable Industry is not going to skip the infrastructure and land on general capability. Build the training, validation and deployment infrastructure for the real physical world properly, and the admission standard for the next stage will follow.
App.Sources

Sources

  1. IFR, World Robotics. Robot density in Chinese manufacturing reached 470 units per 10,000 employees in 2023, third worldwide, against a global average of 162. New installations in China in 2023 came to 276,000 units, roughly 51 per cent of the world total.
  2. Gartner and industry estimates. The shortfall in high-quality data for humanoid robots runs to four orders of magnitude. Teleoperation data costs roughly CNY 2,000 per hour.
  3. Gasgoo Auto Research Institute. High-quality embodied data in existence stands at about 500,000 hours against roughly 10 million hours needed for general capability, a shortfall above 99 per cent. The volume of physical AI data is about 1/20,000 of the corpus available to language models. Simulation can bring the cost of a single data item down to between 1/20 and 1/200.
  4. Frost & Sullivan. The market for embodied intelligence solutions in China is projected at roughly CNY 142.6 billion by 2030.
  5. Insight, the capability–deployment paradox. Token and compute consumption is high while capability does not grow in proportion. Total data volume is enormous but data efficiency is very low, and an hour of video may carry less usable physical knowledge than a thousand words of text. Teleoperation, ego-centric capture, synthetic data and video learning, the four mainstream data paths, all fall short on physical fidelity. Taken together, no scaling law for robots has appeared to date.
  6. Insight, comparison of data costs. Teleoperation data costs roughly CNY 2,000 per hour (Gartner), while simulation can bring the cost of a single data item down to between 1/20 and 1/200. The physical corpus is about 1/20,000 of the corpus available to language models.