🤖 Asimov Architecture

Asimov - Humanoid Robot Platform

A comprehensive monorepo for developing, controlling, and managing humanoid robots. Named after Isaac Asimov, this platform implements a layered architecture for robust robot operations.

🏗️ Architecture Overview

Asimov follows a distributed 4-layer architecture with clear separation between robot-side and cloud-side components:

                    ☁️  CLOUD/BACKEND
┌─────────────────────────────────────────────────────────────┐
│                    Command & Control                        │
│          (Teleoperation, Autonomy, Intelligence)           │
├─────────────────────────────────────────────────────────────┤
│                     Data Engine                             │
│        (Audio/Video/State Storage & Analytics)             │
├─────────────────────────────────────────────────────────────┤
│                       Gateway                               │
│                (Cloud-Robot Interface)                      │
└─────────────────────────────────────────────────────────────┘
                              ↕
                    🤖  ROBOT HARDWARE
┌─────────────────────────────────────────────────────────────┐
│                   Transport Layer                           │
│              (Communication & Streaming)                    │
├─────────────────────────────────────────────────────────────┤
│                Hardware Abstraction                         │
│                   (Robot Runtime)                          │
└─────────────────────────────────────────────────────────────┘

📁 Project Structure

🤖 Robot (/robot)

Everything that runs directly on the robot hardware

• cortex/ - Primary firmware for custom robot development

  • Custom robot implementations: Houses all firmware when developing proprietary robots
  • Core control systems: Primary control loops, motor drivers, sensor interfaces
  • Hardware abstraction: Direct hardware access and real-time control
  • Empty for third-party robots: Not used when working with Unitree G1, AgiBot A2
  • Mission-critical: Primary firmware for custom hardware platforms

• link/ - Robot control interface layer (“Always Present”)

  • Universal robot interface: Always exists regardless of robot type
  • Robot-specific implementations: agibot_a2/, unitree_g1/ adapters for third-party robots
  • Control logic and trajectory planning coordination
  • Inverse kinematics calculations and motor control interfaces
  • Never blocked by Nucleus - maintains robot safety
  • Read/write flow management: read_flow/, write_flow/ for sensor data and commands
  • Works with Cortex: Provides interface layer even for custom firmware

• nucleus/ - Robot gateway and service manager

  • Cloud service interface - handles all external communications
  • Connection management - reconnections, retries, fault tolerance
  • Metrics collection - system health and performance data
  • Can crash safely without affecting core robot functionality
  • Acts as buffer between critical systems and external services

• simulation/ - Robot simulation environment

  • Physics simulation and testing
  • Third-party integrations
  • Virtual development and validation

• common/ - Shared robot utilities

  • Configuration management (config.yaml)
  • Shared memory interfaces
  • Cross-component data structures

☁️ Backend (/backend)

Cloud-based services for robot management and data processing

• gateway/ - Universal service entry point

  • Multi-protocol support: WebRTC, gRPC, HTTP endpoints
  • Service discovery and routing for all backend components
  • Authentication and security layer
  • Load balancing and request distribution
  • Primary interface for all external communications

• command_control/ - Command & Control Layer

  • teleoperation/ - Primary mode: Human operator control systems
  • intelligence_provider/ - AI decision engines for autonomous capabilities
  • cloud_models/ - Cloud-based AI/ML models for complex reasoning
  • local_models/ - Edge computing models for low-latency decisions
  • Flexible input routing based on operational mode

• data_engine/ - Comprehensive data management

  • storage/databases/ - ClickHouse for long-term data storage
  • telemetry/ - Real-time data collection and streaming
  • Stores: Audio streams, video feeds, robot state data, sensor logs
  • Analytics: Historical analysis, performance metrics, behavior insights
  • Time-series data for trajectory and performance analysis

• diagnostics/ - System monitoring and observability

  • replay/ - Historical data replay and debugging tools
  • Prometheus/Grafana stack for real-time monitoring
  • System health dashboards and alerting
  • Performance bottleneck identification

🖥️ Frontend (/frontend)

User interfaces for robot interaction

• website/ - Web-based control dashboard

  • Real-time robot monitoring
  • Configuration management
  • Analytics and reporting

• mobile_app/ - Mobile application

  • Remote robot control
  • Status monitoring
  • Emergency controls

🧪 Testing (/test)

Integration and system-wide testing

• Cross-system integration tests
• End-to-end testing scenarios
• Performance benchmarks

🤖 Development Modes

Asimov supports two primary development approaches:

🔧 Custom Robot Development

When building proprietary robot hardware:

• Full control over hardware, sensors, actuators
• Custom control loops and real-time systems
• Direct hardware abstraction layer implementation

🏭 Third-Party Robot Integration

When working with existing platforms (Unitree G1, AgiBot A2):

• Adapter pattern for different robot manufacturers
• Interface with existing robot SDKs and APIs
• Standardized control interface across platforms

This dual-mode architecture allows Asimov to support both custom hardware development and integration with commercial robot platforms using the same software stack.

🛡️ Safety-First Architecture

Our architecture directly embodies Asimov’s Laws through:

Human-Centric Control

• Teleoperation primacy: Humans remain in control of all critical decisions
• Real-time responsiveness: Low-latency command execution for immediate human control
• Rich feedback: Audio, video, and sensor data provide operators with complete situational awareness

Layered Safety Systems

• Link protection layer: Secondary firmware that acts as final safety arbiter
• Never-blocking guarantee: Safety systems remain responsive regardless of other component states
• Fault isolation: Critical safety functions isolated from non-essential systems

Transparent Operations

• Complete data logging: All robot actions, decisions, and states recorded
• Replay capabilities: Historical analysis ensures accountability and learning
• Human oversight: Operators can review and understand all robot behaviors