Skip to content

Hardware Selection Guide

Robot hardware selection is the first step of system design, directly determining project cost, capability range, and development timeline. This article provides a systematic hardware selection decision framework.

Selection Decision Tree

graph TD
    START[Start Selection] --> USE_CASE{Use Case?}

    USE_CASE -->|Research/Academic| RESEARCH[Research Project]
    USE_CASE -->|Teaching/Entry| EDU[Teaching Project]
    USE_CASE -->|Product/Industrial| PROD[Product Development]

    RESEARCH --> BUDGET_R{Budget?}
    BUDGET_R -->|< $500| R_LOW["SO-100 / Koch v1.1<br/>+ USB Cameras<br/>+ Laptop"]
    BUDGET_R -->|$500 - $5K| R_MID["Koch Dual-arm / GELLO<br/>+ Jetson Orin NX<br/>+ RealSense D435i"]
    BUDGET_R -->|$5K - $50K| R_HIGH["ALOHA / Mobile ALOHA<br/>+ Jetson AGX Orin<br/>+ Multi-camera + LiDAR"]
    BUDGET_R -->|> $50K| R_TOP["Franka + LEAP Hand<br/>+ GPU Workstation<br/>+ MoCap + Force Sensors"]

    EDU --> EDU_CHOICE["SO-100<br/>+ USB Cameras<br/>+ LeRobot"]

    PROD --> PROD_TYPE{Product Type?}
    PROD_TYPE -->|Manipulation| PROD_MANIP["UR/Franka + OnRobot<br/>+ AGX Orin + Industrial Camera"]
    PROD_TYPE -->|Mobile| PROD_MOBILE["Custom Base<br/>+ Orin NX + LiDAR<br/>+ Multi-RealSense"]
    PROD_TYPE -->|Humanoid| PROD_HUMANOID["Custom QDD Actuators<br/>+ Jetson Thor<br/>+ Full-body Sensors"]

    style START fill:#e3f2fd
    style R_LOW fill:#e8f5e9
    style R_MID fill:#e8f5e9
    style R_HIGH fill:#e8f5e9
    style R_TOP fill:#e8f5e9
    style EDU_CHOICE fill:#fff3e0
    style PROD_MANIP fill:#fce4ec
    style PROD_MOBILE fill:#fce4ec
    style PROD_HUMANOID fill:#fce4ec

Budget-Tiered Solutions

Educational Entry Level: < $500

Goal: Low-cost algorithm validation, learning the LeRobot pipeline.

Component Choice Cost
Robot arm SO-100 (Leader + Follower dual-arm teleop) $220
Cameras 2x USB cameras (Logitech C920) $100
Compute Personal laptop + cloud GPU (Colab/Lambda) $0 (existing)
Other USB Hub, mounts, 3D printed parts $50
Total ~$370

What you can do:

  • LeRobot data collection and training (ACT, Diffusion Policy)
  • Simple pick-and-place tasks
  • Algorithm comparison experiments
  • Understand the complete teleoperation and data collection workflow

Limitations: Low torque, low precision, 5 DOF, average servo quality.

Research Prototype Level: $500 - $5K

Goal: Complete manipulation learning research platform.

Component Choice Cost
Robot arm Koch v1.1 dual-arm (Leader + Follower) $680
Teleop (optional) GELLO (adapts to multiple target arms) $200
Compute (inference) Jetson Orin NX 16GB $600
Compute (training) RTX 4090 workstation or cloud $1,600 or pay-per-use
Depth camera Intel RealSense D435i x 2 $600
Force sensor (optional) Bota SensONE $2,000
Other Mounts, cables, power $200
Total $2,000 - $5,000

What you can do:

  • Complete dual-arm manipulation research
  • Vision + force multimodal policies
  • 6DOF manipulation tasks (folding clothes, pouring water, etc.)
  • MoveIt 2 motion planning integration
  • Full LeRobot research pipeline

Advanced Research Level: $5K - $50K

Goal: Publication-ready research platform.

Component Choice Cost
Robot arm ALOHA dual-arm system $20,000
Or mobile version Mobile ALOHA $32,000
Dexterous hand (optional) LEAP Hand $2,000
Cameras 4x RealSense D435i + ZED 2 $2,000
LiDAR (mobile version) Livox Mid-360 $500
Tactile sensor (optional) GelSight Mini x 2 $1,000
Compute (inference) Jetson AGX Orin 64GB $1,600
Compute (training) 2x A100 80GB or rental $30,000 or pay-per-use
Total $20,000 - $50,000

What you can do:

  • Complete ACT / Diffusion Policy reproduction
  • Mobile manipulation tasks
  • VLA model fine-tuning and deployment
  • Dexterous manipulation (with LEAP Hand)
  • High-quality data collection, publish at top conferences

Production Pilot Level: > $50K

Component Choice Cost
Robot arm Industrial arms (Franka Emika Panda x 2) $60,000
Dexterous hand LEAP Hand / Allegro Hand \(2,000-\)15,000
Sensor suite ATI F/T + GelSight + multi-camera $20,000
Motion capture OptiTrack (6 camera) $30,000
Compute (inference) Jetson AGX Orin 64GB $1,600
Compute (training) 8x H100 DGX or cloud $200,000+
Safety system E-stop, force limits, fencing $5,000
Total $100,000+

What you can do:

  • Near-production-ready robot manipulation system
  • Large-scale data collection and VLA training
  • Human-robot collaboration safety validation
  • Extended unattended experiments

Sensor Selection Matrix

Vision Sensors

Need Primary Alternative Price Rationale
Tabletop manipulation (indoor) RealSense D435i Orbbec Gemini 2 ~$300 Depth+RGB, low cost
Close-range manipulation RealSense D405 — ~$300 Short-range high-precision depth
Mobile navigation (indoor) RealSense D455 ZED Mini ~$350 Wide baseline
Outdoor ZED 2i — ~$500 Passive stereo unaffected by sunlight
High-speed scenarios FLIR Blackfly (global shutter) Event camera ~$500 No motion blur
Multi-view coverage 3-4x Logitech C922 — ~$70/ea Low-cost multi-view

LiDAR

Need Primary Alternative Price Rationale
2D navigation RPLIDAR A1/S2 — ~$100 Low cost
3D SLAM (low cost) Livox Mid-360 — ~$500 Excellent value
3D SLAM (high precision) Ouster OS1-64 Velodyne VLP-16 $5K+ High channel count
Outdoor long range Livox HAP — ~$1,000 150m range

Force/Tactile Sensors

Need Primary Alternative Price Rationale
Precision force-controlled assembly ATI Mini45/Nano17 OnRobot HEX-E ~$5K Accuracy + bandwidth
Collision detection Robotiq FT 300 Bota SensONE ~$2K Sufficient + low cost
Dexterous manipulation research GelSight Mini DIGIT ~$300 High-resolution vision-tactile
Simple contact detection FSR film sensor — ~$5 Lowest cost
Large-area coverage ReSkin XELA uSkin $200+ Film array

Position/Pose Sensors

Sensor Use Accuracy Price
Joint encoder (built-in) Joint angle 0.088 deg Included in motor
OptiTrack External MoCap Sub-millimeter $10K+
iPhone ARKit 6-DoF tracking (UMI) cm-level (Use existing iPhone)
Vicon Precision MoCap 0.1mm $50K+

Computing Platform Selection

By Workload

Workload Minimum Config Recommended Config Power
2D navigation + avoidance Jetson Orin Nano 8GB Jetson Orin NX 8GB <15W
3D SLAM + navigation Jetson Orin NX 8GB Jetson Orin NX 16GB 10-25W
Object detection (YOLO) Jetson Orin Nano 8GB Jetson Orin NX 16GB 10-25W
Visual policy inference (small model) Jetson Orin Nano 8GB Jetson Orin NX 8GB <15W
VLA inference (3B) Jetson AGX Orin 32GB Jetson AGX Orin 64GB 30-60W
VLA inference (7B+) Jetson AGX Orin 64GB Jetson Thor 40-60W
Multi-sensor fusion Jetson Orin NX 16GB Jetson AGX Orin 32GB 15-50W

Training vs Inference Separation

Typical training-deployment separation approach:

Training: Cloud/workstation GPU (A100/H100/4090)
  | Model export (PyTorch -> ONNX)
  | Inference optimization (ONNX -> TensorRT FP16/INT8)
Inference: Onboard Jetson (Orin NX/AGX Orin)

For more detailed computing platform information, see Computing Platforms.

Imitation Learning Entry

Robot arm: Koch v1.1 or SO-100 ($110-$250)
Camera: 2x USB cameras ($100)
Compute: Laptop + Jetson Orin Nano ($300)
Software: LeRobot
Total cost: ~$600-$800

Dual-arm Manipulation

Robot arm: ALOHA dual-arm system ($20K)
Camera: 4x C922 (ALOHA standard) ($280)
Compute: Jetson Orin NX + RTX 4090 workstation ($2,200)
Software: ACT / Diffusion Policy
Total cost: ~$23K

Dexterous Manipulation

Robot arm: Franka Panda or Koch v1.1
Dexterous hand: LEAP Hand ($2K)
Sensors: Tactile (GelSight/DIGIT) + force sensor
Compute: AGX Orin inference + A100 training
Software: Isaac Gym (simulation) + custom real-robot deployment
Total cost (low-cost): ~$5K | (high-end): ~$40K

Mobile Manipulation

Platform: Mobile ALOHA ($32K)
Additional sensors: LiDAR (Livox Mid-360, $500)
Compute: AGX Orin 64GB ($1,600)
Software: ROS2 Nav2 + manipulation policy
Total cost: ~$35K

Practical Procurement Advice

China Procurement Channels

Component Channel
Dynamixel motors ROBOTIS Taobao flagship / Taobao distributors
Feetech servos Feetech official Taobao store
Intel RealSense JD.com self-operated / Intel authorized distributors
Jetson NVIDIA authorized distributors / Yahboom / Waveshare
LiDAR Manufacturer official Taobao stores (Livox, RPLIDAR)
3D printed parts Self-print / Taobao JLCPCB 3D printing
Aluminum extrusion Taobao (MISUMI, Euro 2020/2040 standard)

Dynamixel Motor Notes

  • Distinguish Leader/Follower use: XL330 (Leader, low torque) vs XL430/XM430 (Follower, high torque)
  • Volume discounts: Contact distributors for bulk pricing (10+)
  • Spare parts: Buy 1-2 extra motors as backups
  • Power supply: 12V series must use stable power supplies; poor-quality ones cause unstable communication

Common Pitfalls

Selection Pitfall Guide

  1. Over-pursuing DOF: Many tasks only need 5-6 DOF, 7 DOF is not necessary
  2. Ignoring software ecosystem: Even great hardware is hard to use without companion software (LeRobot/ROS2)
  3. Underestimating camera importance: Visual policy performance largely depends on camera quality and layout
  4. Neglecting force sensors: Contact-rich tasks need force feedback; vision alone is insufficient
  5. Over-specifying compute platform: Many policy models run smoothly on Orin NX; AGX is not always needed immediately
  6. Ignoring safety: Even research phases need emergency stop buttons and force limits
  7. USB bandwidth contention: Multiple cameras on the same USB controller compete for bandwidth; distribute across controllers
  8. 3D printing tolerance: Joint mating surfaces should have 0.2-0.3mm clearance
  9. Thermal issues: Jetson throttles under high load; ensure cooling (fans/heatsinks)
  10. Cable management: Use flexible flat cables at moving joints, maintain adequate bend radius

评论 #