> For the complete documentation index, see [llms.txt](https://strikerobot.gitbook.io/strikerobot/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://strikerobot.gitbook.io/strikerobot/sr-agentics-and-sr-platform.md).

# SR-AGENTICS & SR-PLATFORM

Two systems, one Physical-AI flywheel

Strike Robotics builds two complementary systems. **SR-Agentics** is the embodied AI that runs on the
\
robot - task-conditioned, open-vocabulary navigation. **SR-Platform** is the SaaS that produces the
\
simulation environments and (in V2) the training datasets robots learn from. This report gives a onepage picture of each, then shows how they reinforce each other

<table><thead><tr><th width="170">System</th><th width="150">Status</th><th>What it is</th><th>Role in the stack</th></tr></thead><tbody><tr><td><strong>SR-Agentics</strong></td><td>R&#x26;D →
 product</td><td>On-robot agentic navigation
 (Task-Conditioned Scene
 Graphs)</td><td>Consumer of simulation + data; the
 deployed robot brain.</td></tr><tr><td><strong>SR-Platform V1</strong></td><td>Production</td><td>Natural-language → MuJoCo
 simulation scene generator</td><td>Producer of simulation environments.</td></tr><tr><td><strong>SR-Platform V2</strong></td><td>Planned</td><td>Physical-AI Data Factory (NL →
 training-ready dataset)</td><td>Producer of photorealistic, annotated
<br>training data</td></tr></tbody></table>

## 1. SR-Agentics - the on-robot brain

SR-Agentics makes the spatial representation a function of the task: a Task-Conditioned Scene Graph (TC-SG) whose ontology and build budget are chosen from the instruction and deployment context (indoor / outdoor / SAR). It runs as a dual-system agentic stack - System 1 (edge, fast, always-on) is
&#x20;a classic safety loop (SLAM/VIO + obstacle avoidance + Nav2); System 2 (cluster, slow, event-triggered) runs the embodied-reasoning model, the TC-SG representation policy, and the planner.

<figure><img src="https://3817982748-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDF4m7JHke7C5cjSeIrBU%2Fuploads%2FPNE80wwIcO1OOmjukCkg%2Fimage.png?alt=media&amp;token=a4213f81-733f-42fa-a2ed-8a7ce7cb6bee" alt=""><figcaption></figcaption></figure>

***Figure 1**. SR-Agentics dual-system architecture. System 1 (edge) is a classic safety loop; System 2 (cluster, eventtriggered) runs the embodied-reasoning model, TC-SG policy and planner. Security & Reliability planes (top) gate every action. Full SLOs / security / anti-hallucination live in the production proposal.*

{% hint style="info" icon="message-exclamation" %}
Full production SLOs, agent-security model and anti-hallucination detail live in trikeRobot\_Agentics\_Production\_Proposal. Targets: safety loop ≤100 ms, query ≤8 s p50 / ≤15 s p95, SR ≥0.85, bounded memory.
{% endhint %}

### Inside SR-Agentics

#### Task-Conditioned Scene Graph (the core IP)

A representation policy picks or synthesizes the graph ontology from the task via an informationbottleneck objective - applied at construction, not just retrieval. Ships with a reusable schema library (indoor-service, outdoor-urban, SAR) plus open-vocab synthesis of new relation types on demand. Result: leaner graphs, lower onboard latency, bounded memory, and transfer-by-template instead of a new mapping system per client.

#### Dual-system runtime

System 1 (edge, always-on): a classic reactive safety loop - SLAM/VIO, obstacle avoidance, Nav2,
\
Safety Governor - that keeps the robot safe even with no cloud link. System 2 (cluster, enttriggered): an agent harness running a perceive → reason → act → reflect loop over open-vocab perception, the TC-SG policy (M3/M4) and the IB planner (M5).

#### Provider-agnostic model stack&#xD;

Every model role is a swappable interface, pinned per deployment at the M6 eval gate: agent harness (LangGraph-class), embodied-reasoning VLM (System 2), open-vocab perception (SAM 3 → ounding DINO / YOLO-World), visual features (DINOv3 + compact captioner). The System-1 / perception defaults ship with open weights - no vendor lock; sensitive sites run fully self-hosted.

#### Reliability & security built in&#xD;

Confidence gating + abstain/fallback, IB-grounded retrieval, lazy VLM verification, ontology constraints, citation-grounded answers, replan-on-update. A security plane wraps every action:
\
mTLS/gRPC, RBAC, sandboxed tools, action allow-list, immutable audit log, red-teaming.

### Edge deployment - Jetson and Qualcomm

Because System 1 and the on-device runtime are a swappable interface, SR-Agentics is not tied to one silicon vendor. The reactive safety loop (plus a quantized slice of System-2 perception) runs on-edge on either target:

<table><thead><tr><th width="176">Edge target</th><th width="221">Silicon</th><th>Fit</th></tr></thead><tbody><tr><td><strong>NVIDIA Jetson</strong></td><td>Jetson Thor / Orin - Isaac stack,
 CUDA</td><td>Highest on-device compute; native to the same
 Isaac Sim / training toolchain used for 
development</td></tr><tr><td><strong>Qualcomm</strong></td><td>Dragonwing / QCS-series ·
 Snapdragon + Hexagon NPU
 (Qualcomm AI Engine)</td><td>Low-power, fanless, embedded form factors for
 fleet-scale and battery-constrained robots.
 INT8-quantized models run the System-1 loop
 and on-device perception on the Hexagon
<br>NPU - broadening BOM options and letting
<br>Strike ship where Jetson is over-spec, costly, or
 supply-constrained.</td></tr></tbody></table>

*Both targets honor the same dual-system contract - identical System-1 safety loop + System-2 client - so a robot moves
&#x20;between Jetson and Qualcomm hardware as a build/config change, not a redesign.*

***

## 2. SR-Platform V1 - scene generator

SR-Platform V1 turns a natural-language description into a ready-to-run MuJoCo (MJCF) simulation
\
scene in under a minute - no XML, no CAD. A robotics engineer types "warehouse with a forklift AGV,
&#x20;3 pallet stacks, overhead camera" and gets a complete, standards-checked scene. It is a full SaaS (auth, credits, job queue, browser viewer), live in production.

<figure><img src="https://3817982748-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDF4m7JHke7C5cjSeIrBU%2Fuploads%2Fog1ERa4brjmtnfT6GAyw%2FChatGPT%20Image%20Jun%2026%2C%202026%2C%2011_59_58%20AM.png?alt=media&amp;token=b7c582a5-b992-443b-9edc-12c45a646b25" alt=""><figcaption></figcaption></figure>

<figure><img src="https://3817982748-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDF4m7JHke7C5cjSeIrBU%2Fuploads%2Fb9ESsRJVkX80ldREpsOR%2Fimage.png?alt=media&amp;token=7ac49012-319b-4e29-ae17-11a931f99e39" alt=""><figcaption></figcaption></figure>

### V1 feature set (all live in production)

<table><thead><tr><th width="227">Feature</th><th>What it does</th></tr></thead><tbody><tr><td>1 · Scene generation (NL)</td><td>English description → complete MJCF scene (rooms, objects, robot at correct spawn). Async job returns a <code>job_id</code> in &#x3C;100 ms; poll for the result.</td></tr><tr><td>2 · Asset Forge</td><td>No CAD needed: Qdrant semantic search over the asset library; if no match, an LLM writes <strong>CadQuery</strong> (OCCT) code to generate a new STL (OpenSCAD fallback). View / edit / re-render asset source.</td></tr><tr><td>3 · Smart layout + safety checks</td><td>Auto-places objects and verifies <strong>ISO 10218 / NEC 79 / NFPA 70E</strong> compliance; robot-scale calculator sizes the robot to the room.</td></tr><tr><td>4 · Real-time NL scene editing</td><td>Insert / remove / move / delete objects and swap robots by English command on a live scene.</td></tr><tr><td>5 · Robot library (20+)</td><td>MuJoCo Menagerie: UR/KUKA/Franka arms, Anymal/Go2, humanoid H1/G1, AGVs - one-command load with auto-scale.</td></tr><tr><td>6 · Export / import</td><td>ZIP (scene + STL; robot or scene-only); import external scenes; saved-scene library with one-call activation.</td></tr><tr><td>7 · Brainstorm assistant</td><td>Chat freely about an idea → a composed, detailed generation prompt; refine iteratively.</td></tr><tr><td>8 · Platform infrastructure</td><td>JWT auth + per-user state, credit/quota + rate limiting, ARQ job queue, MinIO storage, browser MuJoCo-WASM viewer, pre-made environment templates.</td></tr></tbody></table>

***Stack**: FastAPI · LangGraph · ARQ+Redis · PostgreSQL · MinIO · Qdrant · InfluxDB · CadQuery (OCCT) · MuJoCo-WASM, shipped
&#x20;as a 9-service Docker Compose. Customers: research labs, robotics startups, industrial integrators, education - pricing $0
&#x20;(research) → $499 (starter) → $2,499 (growth) → $15K+ (enterprise).*

### Interactive & articulated objects

Generated scenes aren't just static props - V1 emits MJCF with real physics and articulation, so the
\
environment is interactive on MuJoCo today:

#### Articulated objects (openable / movable)

Doors, drawers, cabinets, levers and valves are emitted with joints (hinge / slide) - a door swings
\
open, a drawer slides out, a handle turns. The robot can plan and execute "open the door, then
\
enter."

#### Dynamic & breakable objects

Rigid and fragile objects obey contact dynamics: a vase knocked off a shelf falls, collides and
\
breaks apart, a box stack topples. Authored directly in MuJoCo's rigid-body + contact physics.

*This makes V1 scenes valuable to SR-Agentics beyond static maps: dynamic / moving / disappearing objects are exactly what
&#x20;the agent's change detection and SAR hazard reasoning need to train against. V2 (Newton) later raises the physics fidelity
&#x20;(deformables, fracture) and speed.*

***

## 3. SR-Platform V2 - Physical-AI Data Factory

V2 turns the scene generator into a data factory: natural language in, a training-ready dataset out. It
\
keeps V1 as the foundation and adds three open-source engines, integrated end-to-end so each stage's
&#x20;output is the next stage's exact input. The aim: what took 20 people six months becomes one engineer
&#x20;one day, attacking the sim-to-real gap at the infrastructure level.

<figure><img src="https://3817982748-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDF4m7JHke7C5cjSeIrBU%2Fuploads%2Fn0YiRD2Yp7O9F92ftiu9%2Fimage.png?alt=media&amp;token=380ef894-9b9e-49c0-8008-4db649ce41bf" alt=""><figcaption></figcaption></figure>

*Object interaction (articulated doors/drawers, falling & breaking objects) is already available in V1 on MuJoCo - see §2
&#x20;above. V2's Newton engine only upgrades the fidelity (contact-rich, deformable and fracture physics at 70–100× speed), and
&#x20;Cosmos annotates the resulting cause → effect traces for VLA training.
&#x20;V2 is the subject of a $2.5M seed proposal (DSC Labs). Market: synthetic robot training data $1.3B (2025) → $9.7B (2030).*

***

## 4. How they connect - the Physical-AI flywheel

The two systems are producer and consumer of the same physical-AI loop. SR-Platform manufactures
&#x20;the worlds and the data; SR-Agentics learns in them and is validated on them; once deployed, real
&#x20;robots surface new scenarios and asset needs that flow back into the Platform. Each turn makes both
&#x20;stronger.

<figure><img src="https://3817982748-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDF4m7JHke7C5cjSeIrBU%2Fuploads%2FXacuJGxoskYpV4owVwcm%2Fimage.png?alt=media&amp;token=0970f152-93a5-4337-9d85-0986341c0fb7" alt=""><figcaption></figcaption></figure>

### Concrete hand-off points

<table><thead><tr><th width="205">From → To</th><th>What flows</th><th>Why it matters</th></tr></thead><tbody><tr><td>Platform V1 → Agentics</td><td>MJCF scenes for the indoor/outdoor/SAR benchmark; robot library (Go2/G1); domain-randomized variants; <strong>interactive / articulated + dynamic-event scenes</strong> (openable doors, falling/breaking objects) -MuJoCo-native.</td><td>The Agentics proposal needs exactly this: a MuJoCo/Isaac scenario suite for the cross-domain benchmark, sim-to-real training, and dynamic-event scenes that train change detection &#x26; SAR hazard reasoning.</td></tr><tr><td>Platform V2 → Agentics</td><td>Photorealistic, reasoning-annotated datasets (Cosmos); higher-fidelity contact-rich / deformable / fracture physics (Newton).</td><td>Closes the sim-to-real gap and supplies supervision for the planner/perception models, raising the fidelity of the interactions V1 already simulates.</td></tr><tr><td>Agentics → Platform</td><td>Real deployment scenarios, failure cases, schema-library templates, new object/asset requests.</td><td>Field data defines which scenes and assets the Platform should generate next - and feeds the compounding asset library.</td></tr></tbody></table>

#### Shared substrate

Both standardize on MuJoCo / MJCF and the Unitree Go2 / G1 platforms, so scenes and robots
\
move between systems without conversion friction.

#### One strategic story

Strike owns both the data layer (Platform) and the embodied product (Agentics) for physical AI -
\
a rare, defensible end-to-end position as Newton/Cosmos/TRELLIS mature.

***

*Strike Robotics - Systems Overview, June 3, 2026. Sources: SR-Agentics production proposal; SR-Platform V1 product brief &
&#x20;technical spec; SR-Platform V2 investment proposal (DSC Labs); STaR (arXiv:2602.09255). V2 and SR-Agentics product status are
&#x20;forward-looking.*