> 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/core-framework-and-system-architecture.md).

# CORE FRAMEWORK & SYSTEM ARCHITECTURE

STRIKEROBOT.AI is structured around a modular, six-layer core framework designed to develop, evaluate, and scale embodied physical intelligence for security and safety in high-risk environments. Drawing from the SafeGuard ASF architecture, each layer addresses specific aspects of perception, reasoning, planning, and action, while integrating shared data infrastructure, tool orchestration, and evaluation standards.

### Object Metadata Framework

At the foundation is a JSON-based metadata structure that captures real-time robot state (battery, pose, joint angles, stability, mode), available tools, and user preferences (priorities, cost sensitivity). This enables fleet-wide collective intelligence, with key tools including:

<table><thead><tr><th width="250"> Tool Name</th><th>Description</th></tr></thead><tbody><tr><td><strong>perception_dinov3</strong></td><td>Detects/identifies objects, workers, intruders via DINOv3; returns bounding boxes, 768-dim embeddings, semantic captions.</td></tr><tr><td><strong>perception_pose</strong></td><td>Estimates 3D human pose via GVHMR; returns 18 keypoints for state understanding and anomaly detection.</td></tr><tr><td><strong>knowledge_vector_db</strong></td><td>Queries Milvus DB for identifications using embeddings; supports workers, intruders, incidents.</td></tr><tr><td><strong>navigation_viplanner</strong></td><td>Plans local collision-free paths with semantic costmaps.</td></tr><tr><td><strong>motion_kungfubot</strong></td><td>Executes locomotion policies (WALK, RUN, STAIR, CAREFUL, TURN, STOP)</td></tr><tr><td><strong>intervention_safe_contact</strong></td><td>Performs safe interventions via impedance control (soft_guide, firm_guide, block).</td></tr><tr><td><strong>reasoning_safety</strong></td><td>LLM for risk assessment; analyzes behavior, predicts accidents.</td></tr><tr><td><strong>reasoning_security</strong></td><td>LLM for threat assessment; recommends responses.</td></tr></tbody></table>

### Six-Layer Architecture

The system operates through a hierarchical pipeline where each layer communicates seamlessly for real-time, intelligent behavior.

#### Layer 1: Perception Layer

Objective: Transform raw sensor data (RGB-D cameras) into semantically rich representations using DINOv3 for object detection, visual fingerprints, depth maps, and 30-class semantic costmaps. Outputs enable zero-shot recognition and spatial reasoning, running at 30fps with <15ms latency.

#### Layer 2: Understanding Layer

Objective: Build situational awareness from perceptions, including 3D pose estimation (GVHMR), trajectory prediction (3-10s ahead), behavior classification (e.g., loitering, anomalies), and risk scoring (collision probability, threat levels). This interprets actions like falls or suspicious movements for proactive responses.

#### Layer 3: Object Metadata Layer

Objective: Maintain a unified fleet knowledge base for shared intelligence. Uses Milvus vector DB for sub-20ms queries, JSON schemas for state/events, and communication stacks (MQTT, ZeroMQ, WiFi Mesh). Hierarchical structure: Fleet Commander (global planning), Zone Leaders (local allocation via auctions), Squad Members (task execution).

#### Layer 4: Decision Layer - The Agentic System

Objective: Enable autonomous reasoning via ReAct paradigm with dual modes. Realtime Rules (<1s) for emergencies (e.g., collision avoidance); Strategic AI (>3s) using ToolOrchestra/Gemini for chaining tools, multi-robot coordination, and explainable decisions. Supports adaptive patrols and audits.

#### Layer 5: Planning Layer

Objective: Generate feasible trajectories from decisions. Dual planning: Global A\* for high-level waypoints (updates every 5-10s); Local ViPlanner (10Hz) for 5m paths with dynamic replanning, semantic awareness, velocity profiling, and multi-robot deconfliction.

#### Layer 6: Action Layer - MMA-Inspired Action Learning

Objective: Execute plans via learned policies inspired by MMA for versatile locomotion. KungfuBot provides specialized policies (e.g., WALK for efficiency, RUN for urgency) trained in Isaac Gym with adversarial curricula and sim-to-real transfer. Includes impedance-controlled interventions (GUIDE, BLOCK, SUPPORT) with modes like SOFT\_CONTACT (K=100 N/m, 30N max).

### Rule-Based Foundation

Core rules ensure predictability: Emergency Intervention (physical intercept if collision >0.8 prob), Distance Switching (modes based on time-to-hazard), Safe Contact Protocol (verbal warnings first, yield on resistance).

### Robotics RL Training Platform (Future Offering)

As a Stage 2 product, STRIKEROBOT.AI is developing a cloud-based Reinforcement Learning Training Platform that will allow enterprises to generate real-world environment models and train humanoid robots for their specific operational use cases—offered as an Environment-as-a-Service alongside SafeGuard ASF.

### Sensor-Driven Autonomous Combat

Emphasizing closed-loop autonomy, robots use onboard sensors for dynamic responses to hazards (smoke, heat, leaks, sounds) against baselines. Tactical cues feed cognitive decisions for inspection, intervention, or recovery, with human proxy teleoperation for high-risk scenarios without physical exposure.