Skip to content

Typical Robot Teardowns

Overview

Teardown analysis is one of the best ways to learn robot engineering. By tearing down successful commercial products, we can learn:

  • How engineers make tradeoffs between cost, performance, and reliability
  • System architecture design philosophy
  • Engineering details of cable management, thermal design, and waterproofing
  • Real-world component selection choices

This section provides detailed teardowns of four representative robots.

1. Unitree Go2 (Quadruped Robot)

Product Positioning

  • Type: Consumer/Education-grade quadruped robot
  • Price: $1,600 (Air) ~ $2,800 (Pro) ~ $8,000+ (Edu)
  • Weight: ~15 kg
  • Dimensions: 70cm L × 31cm W × 40cm H
  • Battery Life: 1-2 hours (standard use)

System Architecture

graph TB
    subgraph Computing Platform
        J[Jetson Orin NX / Nano<br/>Main Controller] 
        ESP[ESP32<br/>Auxiliary/WiFi/BT]
        MCU[STM32<br/>Motor Control]
    end

    subgraph Perception
        LIDAR[Head LiDAR<br/>Hesai XT16 / Livox]
        CAM[Front Stereo Camera]
        IMU_S[IMU<br/>In Body]
        UW[Ultra-Wideband UWB]
    end

    subgraph Actuation
        M1[LF Hip/Knee/Ankle<br/>BLDC × 3]
        M2[RF Hip/Knee/Ankle<br/>BLDC × 3]
        M3[LR Hip/Knee/Ankle<br/>BLDC × 3]
        M4[RR Hip/Knee/Ankle<br/>BLDC × 3]
    end

    subgraph Power
        BAT[32 × 18650 Battery Pack<br/>~800Wh]
        BMS[Battery Management System]
        DC[DC-DC Converter]
    end

    subgraph Thermal
        FAN1[Fan × 2]
        HS[Aluminum Heatsink]
    end

    J --> MCU
    ESP --> J
    MCU --> M1
    MCU --> M2
    MCU --> M3
    MCU --> M4
    LIDAR --> J
    CAM --> J
    IMU_S --> MCU
    BAT --> BMS --> DC
    DC --> J
    DC --> MCU
    FAN1 --> HS

Motor System

The Go2 uses 12 high-performance BLDC outrunner motors (3 degrees of freedom per leg: hip abduction, hip pitch, knee pitch).

Motor Specifications (estimated from teardown):

Parameter Specification
Type BLDC Outrunner
Slot/Pole Configuration 36 slots / 42 poles
Peak Torque ~40 N·m (after reduction)
Gear Ratio ~9:1 (planetary)
Encoder 14-bit magnetic encoder
Drive Method FOC (Field Oriented Control)
Communication CAN Bus

Motor + reducer + encoder + driver integrated in a single joint module -- this is Unitree's core engineering capability.

Joint module structure:

┌─────────────────────┐
│  Output Flange       │ ← Connects to next link
├─────────────────────┤
│  Planetary Reducer   │
│       (9:1)          │
├─────────────────────┤
│  BLDC Motor          │ ← 36S42P Outrunner
│  (Stator+Rotor+      │
│   Magnets)           │
├─────────────────────┤
│  Magnetic Encoder    │ ← AS5047P or similar
│       (14-bit)       │
├─────────────────────┤
│  FOC Driver Board    │ ← DRV8353 + STM32
│  (MCU + Power Stage) │
├─────────────────────┤
│  CAN Interface       │ ← 4-pin connector
└─────────────────────┘

Battery System

  • Cells: 32 × 18650 lithium cells (8S4P configuration)
  • Nominal Voltage: \(8 \times 3.7V = 29.6V\)
  • Capacity: Approximately \(4 \times 2600mAh = 10.4Ah\)
  • Energy: \(29.6V \times 10.4Ah \approx 308Wh\)
  • BMS: 8S protection board (overcharge/over-discharge/overcurrent/short circuit/balancing)

Sensing System

Head LiDAR:

  • Pro/Edu versions: Hesai XT16 or Livox MID-360
  • Used for SLAM mapping and navigation
  • Mounted on the head, yaws with the head

Front Camera:

  • Stereo or depth camera
  • Used for visual perception/obstacle avoidance

IMU:

  • High-precision 6-axis IMU (accelerometer + gyroscope)
  • Mounted at the center of the body
  • Used for attitude estimation, state estimation

Computing Platform

  • Main Controller: NVIDIA Jetson Orin NX or Jetson Nano (varies by version)
  • Auxiliary Controller: ESP32 (WiFi/Bluetooth communication, remote controller interface)
  • Motion Control: STM32 series MCU (1 kHz control loop)

Thermal Design

  • Dual fans: Active cooling
  • Aluminum heatsinks: Cover computing platform and motor drivers
  • Thermal pads: Connect chips to heatsinks
  • Body shell: Aluminum alloy itself acts as a heatsink

Design Insights

  1. Highly integrated joint modules: Motor + reducer + encoder + driver all-in-one, simplifying assembly and wiring
  2. CAN Bus communication: 12 joint modules daisy-chained via CAN Bus, minimal cabling
  3. Cost control: 18650 battery pack (rather than custom battery packs) reduces cost
  4. Modular legs: All four legs share the same structure, requiring only one set of molds

2. iRobot Roomba j7 (Robot Vacuum)

Product Positioning

  • Type: Home robot vacuum
  • Price: ~$600
  • Weight: ~3.4 kg
  • Dimensions: 33.9cm diameter × 8.7cm height

System Architecture

┌──────────────────────────────────┐
│           Main PCB                │
│  ┌─────────┐  ┌──────────┐      │
│  │ ARM SoC │  │ Motor     │      │
│  │ (Main   │  │ Driver    │      │
│  │ Control)│  │ Circuits  │      │
│  └────┬────┘  └─────┬────┘      │
│       │              │           │
│  ┌────┴────┐  ┌─────┴────┐     │
│  │ Camera  │  │ Sensor    │     │
│  │ (Front) │  │ Interface │     │
│  └─────────┘  └──────────┘     │
└──────────────────────────────────┘

Drive System

Suction Motor:

  • High-speed BLDC motor
  • Speed ~10,000 RPM
  • Generates vacuum suction

Drive Motors:

  • One DC geared motor per side (left and right)
  • Differential drive (differential steering)
  • Encoder feedback for odometry

Brush Motors:

  • Main brush: DC motor driving dual rubber rollers
  • Side brush: Small DC motor, 3-arm rotating brush

Sensing System

Cliff Sensors (6 IR sensors):

  • 6 infrared emitter/receiver pairs on the bottom
  • Emit infrared light, receive reflected light
  • No reflection -> cliff/stair edge -> stop
\[d_{cliff} = f(V_{reflected})\]

Detection distance approximately 1-3 cm, response time < 10 ms.

Front-View Camera:

  • Core upgrade in the j7 series: front-facing RGB camera
  • AI object recognition (charging cables, socks, pet waste, etc.)
  • Uses iRobot Genius platform (on-device + cloud inference)

Bump Sensor:

  • Front arc-shaped bumper
  • Internal microswitches or optical sensors
  • Triggers avoidance behavior on contact

LDS LiDAR (Low-cost laser rangefinder):

  • Rotating 2D LiDAR (protruding top section)
  • Used for room mapping and navigation
  • Resolution ~1°, range ~6m
  • Rotation speed ~5 Hz

Charging Dock

Dock Alignment:

  • Dock emits infrared signal (fan-shaped beam)
  • Roomba has infrared receivers on the bottom
  • Guided alignment via signal strength
  • Charging contacts on the front underside
  • Early Roomba: Random bump-and-go (cheap but inefficient)
  • j7: vSLAM (visual SLAM) + LiDAR mapping
  • Memorizes room layouts, supports zone cleaning

Design Insights

  1. Extreme cost optimization: Every sensor uses the cheapest option (IR cliff sensors instead of ToF)
  2. LiDAR cost reduction: Uses low-cost rotating LDS, sufficient for 2D mapping needs
  3. AI differentiation: The j7's front camera + AI is the core selling point (obstacle avoidance)
  4. Maintainability: Brushes, filters, and other consumables are easy to replace
  5. Circular design: Simplifies post-collision kinematics (can rotate in place without getting stuck on walls)

3. Franka Emika Panda (Collaborative Robot Arm)

Product Positioning

  • Type: 7-DOF collaborative robot arm (academic research standard)
  • Price: ~EUR 20,000 (discontinued, replaced by Franka Research 3)
  • Payload: 3 kg
  • Reach: 855 mm
  • Weight: ~18 kg
  • Repeatability: ±0.1 mm

Joint Structure

Panda's core innovation is that every joint integrates a torque sensor.

Single Joint Module:

┌─────────────────────┐
│  Output Flange       │
├─────────────────────┤
│  Harmonic Drive      │ ← 100:1 ratio
├─────────────────────┤
│  Torque Sensor       │ ← Strain gauge, joint side
├─────────────────────┤
│  BLDC Motor          │
├─────────────────────┤
│  Encoder × 2         │ ← Motor side + Joint side
├─────────────────────┤
│  Drive Electronics   │
└─────────────────────┘

7-Joint Parameters:

Joint Type Range Max Torque Gear Ratio
J1 Revolute ±170° 87 N·m 100:1
J2 Revolute ±105° 87 N·m 100:1
J3 Revolute ±170° 87 N·m 100:1
J4 Revolute ±175° 87 N·m 100:1
J5 Revolute ±170° 12 N·m 100:1
J6 Revolute ±215° 12 N·m 100:1
J7 Revolute ±170° 12 N·m 100:1

Harmonic Drive

Every Panda joint uses a harmonic drive reducer:

Working Principle:

  • Wave Generator: Elliptical cam
  • Flex Spline: Thin-walled elastic gear
  • Circular Spline: Rigid internal ring gear
\[\text{Gear Ratio} = \frac{N_{flex}}{N_{circular} - N_{flex}}\]

Typical \(N_{circular} = 202\), \(N_{flex} = 200\) -> Gear ratio 100:1.

Advantages: Zero backlash, high gear ratio, compact. Disadvantages: Flex spline fatigue life limited, efficiency ~80%.

Torque Sensing

Every joint has a built-in torque sensor -- this is Panda's most distinctive feature:

  • Strain gauge-based, mounted at the reducer output
  • Resolution: 0.05 N·m (large joints) / 0.01 N·m (small joints)
  • Sampling rate: 1 kHz

This enables Panda to directly implement:

  • Joint torque control
  • Collision detection (no external sensors needed)
  • Impedance control
  • Dynamics estimation

Force-Sensitive Gripper

Panda's parallel gripper also integrates force sensing:

Parameter Specification
Stroke 80 mm (40 mm per side)
Max Grip Force 70 N
Force Resolution 0.05 N
Closing Speed 50 mm/s

Control System

1 kHz Real-Time Control Loop:

\[\tau_{cmd} = M(q)\ddot{q}_d + C(q,\dot{q})\dot{q}_d + g(q) + K_d(q_d - q) + D_d(\dot{q}_d - \dot{q})\]
  • Main controller: x86 real-time Linux (PREEMPT-RT kernel)
  • Communication: EtherCAT (to all joint drives)
  • User interface: FCI (Franka Control Interface) or libfranka

Control Modes:

Mode Description Use
Joint position control Send joint angle targets Basic motion
Joint velocity control Send joint angular velocities Smooth motion
Joint torque control Send torques directly Force control research
Cartesian position control End-effector pose targets Task-space control
Cartesian velocity control End-effector velocity Force guidance

Design Insights

  1. 7-DOF redundancy: One more DOF than 6-DOF allows obstacle avoidance while completing end-effector positioning
  2. Joint torque sensing: Enables force control, the core selling point for academic research
  3. Dual encoders: Motor-side + joint-side encoders eliminate reducer flexibility effects
  4. 1 kHz control: High control frequency ensures force control stability
  5. Clean white appearance: Industrial design oriented toward laboratory environments

4. Unitree G1 (Humanoid Robot)

Product Positioning

  • Type: General-purpose humanoid robot
  • Price: ~$16,000 starting
  • Height: Approximately 127 cm
  • Weight: Approximately 35 kg
  • Degrees of Freedom: 23 DOF (standard) / 43 DOF (dexterous hand version)

Joint Distribution

Body Part DOF Joint Types
Head 2 Yaw + Pitch
Each Arm 5-7 Shoulder(3) + Elbow(1) + Wrist(1-3)
Waist 3 Roll + Pitch + Yaw
Each Leg 5 Hip(3) + Knee(1) + Ankle(1)

QDD Joint Modules

The G1 uses QDD (Quasi-Direct Drive) joints -- Unitree's core technology:

Design Philosophy: Low gear ratio + high torque density motor

\[\tau_{output} = N \cdot K_t \cdot i - \tau_{friction}\]

where \(N\) is the gear ratio (in QDD, \(N \approx 6 \sim 9\), much lower than harmonic drives' 100:1).

QDD vs. High Gear Ratio Approach:

Aspect QDD (Quasi-Direct Drive) Harmonic Drive
Gear Ratio 6-9:1 50-160:1
Backdrivability Good (torque transparent) Poor (nearly non-backdrivable)
Force Control Naturally suited Requires additional force sensors
Torque Density Moderate High
Bandwidth High Low (limited by compliance)
Efficiency ~95% ~80%
Cost High motor cost High reducer cost

QDD's "torque transparency" means external torques can be directly sensed through motor current -- no dedicated torque sensors needed.

Joint Module Integration:

┌─────────────────────┐
│  Output Flange       │
├─────────────────────┤
│  Planetary Reducer   │ ← Low gear ratio
│       (6-9:1)        │
├─────────────────────┤
│  BLDC Motor          │ ← High torque density
│  (Large diameter,    │
│   thin profile)      │
├─────────────────────┤
│  Magnetic Encoder    │ ← High precision
├─────────────────────┤
│  FOC Driver Board    │ ← Integrated in module
└─────────────────────┘

Vision System

Head Stereo Camera:

  • Intel RealSense D435i or similar depth camera × 2
  • Provides RGB + Depth + IMU data
  • Forward-facing + downward-facing (estimated)

Computing Platform

  • Main Controller: NVIDIA Jetson Orin NX (8GB / 16GB)
  • Motion Control: Dedicated MCU (STM32H7 class)
  • Communication: CAN FD (joint bus) + Ethernet (perception)

Battery System

  • Installed in the torso (maintains low center of gravity)
  • Lithium battery pack, ~\(48V\) system
  • Battery life approximately 1-2 hours (walking)

Battery Placement and Center of Gravity

\[z_{CoM} = \frac{\sum m_i z_i}{\sum m_i}\]

The battery is the heaviest single component (~20-30% of total weight), and placing it in the torso:

  • Lowers the overall center of gravity
  • Maintains symmetry (left-right balance)
  • Facilitates replacement

Design Insights

  1. QDD joints are the core: Low gear ratio enables force control and safe human-robot interaction
  2. Modular joints: Different torque levels use different sizes of standardized modules
  3. 23 DOF choice: Sufficient without being excessive -- every DOF adds cost and complexity
  4. Jetson Orin NX: Balances computing power and energy consumption (GPU inference + low power)
  5. CAN FD bus: Higher bandwidth than traditional CAN (up to 8 Mbps), suitable for multi-joint high-frequency communication

Cross-Product Comparison

Dimension Go2 Roomba j7 Panda G1
Motors BLDC+Planetary DC Brushed+BLDC BLDC+Harmonic BLDC+QDD
Gear Ratio ~9:1 N/A 100:1 6-9:1
Force Sensing Current estimation None Joint torque sensors Current estimation (QDD)
Main Controller Jetson ARM SoC x86 RT Linux Jetson Orin
Communication CAN Onboard bus EtherCAT CAN FD
Control Frequency 500 Hz ~50 Hz 1 kHz 500-1000 Hz
Navigation LiDAR SLAM vSLAM+LDS N/A Vision
Price $1.6K-8K ~$600 ~EUR 20K ~$16K+

Core Design Philosophies

Product Design Philosophy
Go2 High integration + economies of scale
Roomba j7 Extreme cost optimization + AI differentiation
Panda Performance first + academia-friendly
G1 Quasi-direct drive force control + general platform

Do-It-Yourself Teardowns

Beginner Intermediate Advanced
Servo (SG90) RC car Quadruped robot
Stepper motor Robot vacuum Drone
Mouse (encoder) Gimbal (IMU+motor) Collaborative robot arm
Phone vibration motor 3D printer Surgical robot

Teardown Record Template

=== Teardown Record ===

Device Name: _______________
Date: _______________

1. Appearance and Dimensions
   - L × W × H: _______________
   - Weight: _______________
   - Material: _______________

2. Electrical System
   - Battery: Type___ Voltage___ Capacity___
   - Main Controller: _______________
   - Motors: Type___ Quantity___ Driver___

3. Sensors
   - List: _______________

4. Mechanical Structure
   - Joint count: _______________
   - Reducer: _______________
   - Material: _______________

5. Design Highlights
   - _______________

6. Improvement Suggestions
   - _______________

References

  • Unitree Robotics: unitree.com
  • iFixit Roomba Teardown: ifixit.com
  • Franka Emika Documentation: franka.de
  • YouTube: "Unitree Go2 Teardown", "Roomba j7 Teardown"
  • Katz, B., "Mini Cheetah: A Platform for Pushing the Limits of Dynamic Quadruped Control," ICRA, 2019

评论 #