Skip to main content

Hardware Architecture

4-Tier Architecture Overview​

To effectively manage the complexity of humanoid robotics and leverage distributed computing resources, this course employs a robust 4-tier architectural model. This tiered approach ensures clear separation of concerns, optimizes computational load, and balances real-time performance with scalable AI training.

graph TD
    A[Tier 1: Simulation Rig] -->|Training Data| B[Tier 2: Edge Brain]
    B -->|Sensor Data| C[Tier 3: Sensors]
    B -->|Control Commands| D[Tier 4: Robot Actuators]
    C -->|Raw Data| B
    D -->|State Feedback| B
    
    A -->|Trained Models| B
    B -->|Telemetry| A

Tier 1: Simulation Rig (Development and Training)​

Description​

High-performance workstations and cloud instances for development, simulation, and AI training.

Components​

On-Premise RTX Lab​

Workstation Specifications:

  • GPU: NVIDIA RTX 4070, 4080, or 4090
  • CPU: Intel Core i7/i9 or AMD Ryzen 7/9 (8+ cores)
  • RAM: 32GB minimum, 64GB recommended
  • Storage: 1TB+ NVMe SSD
  • OS: Ubuntu 22.04 LTS

Purpose:

  • ROS 2 application development
  • Digital twin creation (Gazebo/Unity)
  • Isaac Sim photorealistic simulation
  • AI model training (vision, locomotion, manipulation)

Cloud "Ether" Lab​

Cloud Infrastructure:

  • Platform: AWS g5/g6e instances
  • GPU: NVIDIA A100 or H100
  • Scalability: On-demand compute resources

Purpose:

  • Large-scale AI training
  • Distributed simulation
  • Batch data processing

The Latency Trap ⚠️​

Critical Consideration: While cloud resources offer immense flexibility, round-trip communication latency between physical robots and remote servers can severely hinder real-time control loops.

Impact: Delays of 50-200ms are common in cloud communication, which is unacceptable for:

  • Balance control (requires <10ms response)
  • Collision avoidance (requires <50ms response)
  • Dynamic manipulation (requires <20ms response)

Solution: Use cloud for offline training and edge devices for real-time control.

Tier 2: Edge Brain (Real-time Processing)​

Description​

NVIDIA Jetson Edge Kits deployed on or near the physical robot, serving as the localized "brain" for real-time processing.

Jetson Edge Kits​

NVIDIA Jetson Orin Nano​

Specifications:

  • GPU: 1024-core NVIDIA Ampere
  • CPU: 6-core Arm Cortex-A78AE
  • RAM: 8GB
  • AI Performance: 40 TOPS
  • Power: 7-15W

Use Cases:

  • Entry-level edge AI
  • Basic perception tasks
  • Sensor data preprocessing

NVIDIA Jetson Orin NX​

Specifications:

  • GPU: 1024-core NVIDIA Ampere
  • CPU: 8-core Arm Cortex-A78AE
  • RAM: 16GB
  • AI Performance: 100 TOPS
  • Power: 10-25W

Use Cases:

  • Advanced perception
  • Real-time SLAM
  • Local AI inference
  • Multi-sensor fusion

Functions​

✅ Execute real-time control loops
✅ Run critical ROS 2 nodes (motor controllers, VSLAM)
✅ Perform on-device AI inference
✅ Handle sensor data preprocessing
✅ Maintain low-latency operation

Tier 3: Sensors (Perception Input)​

Description​

Sensory hardware attached to the robot for environmental perception and proprioception.

Common Sensors​

Intel RealSense D435i​

Type: RGB-D Camera with IMU

Specifications:

  • RGB Resolution: 1920×1080 @ 30fps
  • Depth Resolution: 1280×720 @ 90fps
  • Depth Range: 0.3m to 3m
  • FOV: 87° × 58° (depth)
  • IMU: 6-axis (accelerometer + gyroscope)

Applications:

  • Visual SLAM
  • Object detection and recognition
  • Depth estimation for manipulation
  • Obstacle avoidance

IMU (Inertial Measurement Unit)​

Type: 6-axis or 9-axis sensor

Measurements:

  • Accelerometer: Linear acceleration (3-axis)
  • Gyroscope: Angular velocity (3-axis)
  • Magnetometer: Magnetic field (3-axis, if 9-axis)

Applications:

  • Robot orientation estimation
  • Balance control
  • Odometry
  • State estimation

ReSpeaker Mic Array​

Type: Multi-microphone array

Specifications:

  • Microphones: 4-6 channels
  • Sampling Rate: 16kHz
  • Features: Beamforming, noise suppression, echo cancellation

Applications:

  • Far-field speech recognition
  • Voice command input
  • Sound source localization
  • Human-robot interaction

Sensor Integration​

# Example: Multi-sensor data fusion
class SensorFusion:
    def __init__(self):
        self.camera_sub = rospy.Subscriber('/camera/image', Image, self.camera_callback)
        self.imu_sub = rospy.Subscriber('/imu/data', Imu, self.imu_callback)
        self.depth_sub = rospy.Subscriber('/camera/depth', Image, self.depth_callback)
    
    def fuse_data(self):
        # Combine visual and inertial data for robust state estimation
        visual_pose = self.vslam.get_pose()
        imu_orientation = self.imu_filter.get_orientation()
        
        # Kalman filter fusion
        fused_state = self.kalman_filter.update(visual_pose, imu_orientation)
        return fused_state

Tier 4: Robot Actuators (Physical Interaction)​

Description​

The physical body of the humanoid robot, comprising joints, motors, and end-effectors.

Robot Platforms​

Unitree Humanoid Robots​

Unitree H1

Specifications:

  • Height: 180cm
  • Weight: 47kg
  • DOF: 25+ joints
  • Actuators: High-torque servo motors
  • Battery: 15Ah (90 minutes runtime)
  • Payload: 30kg

Capabilities:

  • Bipedal locomotion
  • Whole-body manipulation
  • Dynamic balance
  • Outdoor operation

Unitree G1 (Alternative)

Specifications:

  • Height: 130cm
  • Weight: 35kg
  • DOF: 23 joints
  • Focus: Research and education

Actuator Control​

# Example: Joint control interface
class JointController:
    def __init__(self, robot):
        self.robot = robot
        self.joint_states_sub = rospy.Subscriber('/joint_states', JointState, self.state_callback)
        self.joint_cmd_pub = rospy.Publisher('/joint_commands', JointCommand, queue_size=10)
    
    def send_position_command(self, joint_positions):
        """Send position commands to robot joints"""
        cmd = JointCommand()
        cmd.names = self.robot.joint_names
        cmd.positions = joint_positions
        cmd.kp = [100] * len(joint_positions)  # Position gain
        cmd.kd = [10] * len(joint_positions)   # Damping gain
        self.joint_cmd_pub.publish(cmd)

Lab Configurations​

Proxy Lab (Simulation-Focused)​

Equipment:

  • On-Premise RTX Lab workstations (5-10 units)
  • Optional Cloud "Ether" Lab access
  • No physical robots

Approach:

  • All development in simulation
  • Digital twins for all experiments
  • Focus on software and algorithms

Cost: Low (hardware only)

Best For: Budget-constrained institutions, remote learning

Mini Lab (Hybrid)​

Equipment:

  • On-Premise RTX Lab workstations (5-10 units)
  • Jetson Edge Kits (2-3 units)
  • Limited Unitree robots (1-2 units)

Approach:

  • Primary development in simulation
  • Limited hardware deployment
  • Demonstrations on physical robots

Cost: Medium

Best For: Universities with moderate budgets

Premium Lab (Full Hardware)​

Equipment:

  • On-Premise RTX Lab workstations (10+ units)
  • Dedicated Jetson Edge Kits per team (5+ units)
  • Multiple Unitree robots (3-5 units)
  • Full sensor suite for each robot

Approach:

  • Simulation for rapid prototyping
  • Extensive hardware testing
  • Real-world deployment focus

Cost: High

Best For: Research institutions, industry partnerships

Data Flow Architecture​

sequenceDiagram
    participant Sim as Tier 1: Simulation
    participant Edge as Tier 2: Edge Brain
    participant Sensors as Tier 3: Sensors
    participant Robot as Tier 4: Actuators
    
    Sim->>Edge: Trained AI Models
    Sensors->>Edge: Raw Sensor Data
    Edge->>Edge: Process & Infer
    Edge->>Robot: Control Commands
    Robot->>Edge: State Feedback
    Edge->>Sensors: Sensor Configuration
    Edge->>Sim: Performance Telemetry

Key Takeaways​

🔑 4-tier architecture separates development, edge processing, sensing, and actuation
🔑 Simulation rigs enable fast development and AI training
🔑 Edge devices handle real-time control to avoid latency issues
🔑 Sensor fusion combines multiple modalities for robust perception
🔑 Lab configurations can be adapted to different budgets and goals

Navigation:
← Assessments | Glossary →