logo Mr.Robota 具身智能学习手册
/
中文
Chapter 3: 03 Robot Hardware Lerobot And Rdk X5
Core Features Franka Research 3 Kinova Gen 3 (7-DOF) KUKA LBR iiwa UR5e (Uro) UFACTORY xArm 7 RealMan RX / RM65 AgileX Piper (Sungling) ViperX 300S Ark Infinite π0 Hello Robot Stretch D-Robotics MMK2 Rokae xMate CR7 Flexiv Rizon 4
Positioning and Tags Force-controlled ceiling, RL algorithm benchmark All-round lightweight, king of manipulation Force-controlled Rausmayer, medical-grade standard Industrially enduring research tool, vision benchmark Cost-effective embodied system, scale-first choice Domestic lightweight innovator, human-like dual arms Domestic low-cost embodied entry point, ALOHA alternative International imitation learning champion, original ALOHA Ultimate cost-performance desktop level, fixed base North American home interaction dark horse, minimal lift arm Domestic mobile embodied platform, built-in lift chassis No control cabinet design, flexible collaboration Adaptive robot, hand-eye coordination
Reference Price (single arm/unit) Approximately ¥250,000 - 300,000+ Approximately ¥200,000 - 300,000+ Approximately ¥350,000 - 500,000+ Approximately ¥200,000 - 250,000 Approximately ¥60,000 - 80,000 Approximately ¥35,000 - 50,000 Approximately ¥14,900 Approximately ¥35,000 - 40,000 Approximately ¥28,000 - 35,000 (set) Approximately ¥180,000 ($25,000) 约 ¥8万 - 12万 (整机) 约 ¥8万 - 12万 约 ¥45万+ ($65,000+)
Payload 3.0 kg 2.0 kg (continuous)/4.0 kg (peak) 7.0 kg / 14.0 kg 5.0 kg 3.5 kg 5.0 kg 1.5 kg 0.75 kg Approximately 2.0 kg 1.5 kg Approximately 2.0 kg (single arm) 7.0 kg 4.0 kg
Maximum Reach 855 mm 902 mm 800 mm / 820 mm 850 mm 700 mm 740 mm 626 mm 750 mm Approximately 600 mm 520 mm (horizontal telescoping) Approximately 600 mm (with lift mechanism) 988 mm 876 mm
Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation Manipulation
Weight Approximately 18.0 kg 8.2 kg Approximately 22.0 kg / 30.0 kg 20.6 kg 13.7 kg 7.2 kg 4.2 kg 3.5 kg Approximately 10 kg 23.0 kg (whole unit) Approximately 40 kg (whole unit) Approximately 27.0 kg (including built-in controller) 20.0 kg
Joint torque sensor ✅ Present (all 7 joints) ✅ Present (all 7 joints) ✅ Present (all 7 joints) ❌ None (only end can be externally connected) ❌ None (current estimation) ❌ None (current estimation) ❌ None (current estimation) ❌ None ❌ None ❌ None (current threshold for collision prevention) ❌ None ✅ Present (all 7 joints) ✅ Present (all 7 joints)
Infinite rotation capability ❌ None ✅ Present (all 7 joints at 360°) ❌ None ❌ None (most are ±360∘\pm 360^\circ) ❌ None ❌ None ❌ None ❌ None ❌ None ❌ None ❌ None ❌ None (each joint can reach ±360°) ❌ None
Repositioning accuracy ±0.1\pm 0.1 mm ±0.05\pm 0.05 mm ±0.1\pm 0.1 mm ±0.03\pm 0.03 mm ±0.1\pm 0.1 mm ±0.05\pm 0.05 mm ±0.1\pm 0.1 mm Uncalibrated (actuator accuracy is average) Approximately ±0.1\pm 0.1 mm Uncalibrated (designed for soft contact) Approximately ±0.1\pm 0.1 mm ±0.02 mm ±0.05 mm
End-integrated vision ❌ No built-in ✅ Built-in Intel RealSense ❌ No built-in ❌ No built-in ❌ No built-in ❌ No built-in ❌ No built-in (but paired with PIKA gripper with vision) ❌ No built-in ❌ No built-in ✅ Built-in Intel RealSense ✅ Head-integrated multimodal vision ❌ No built-in ❌ No built-in
Low-level Control Frequency 1 kHz (1000 Hz) 1 kHz (1000 Hz) 1 kHz (Fast Research Interface) 500 Hz 100 - 250 Hz 200 Hz 100 - 200 Hz < 100 Hz 100 - 200 Hz 100 Hz Based on RDK X5 scheduling 1 kHz 1 kHz
Main Disadvantages Limited position, prone to jamming; large control box, unable to attach to a mobile chassis. Low stiffness, prone to tremor during heavy contact tasks. Extremely expensive; high barrier to secondary development. Heavy weight, no proprioception; primarily for pure position control. Poor smoothness in high-speed motion, lacks true force control. The software ecosystem is still in its early stages compared to international open-source communities. Very low load, mainly used for lightweight demonstrations and data collection. Short lifespan of servos, prone to overheating, frequent maintenance. Only fixed base, unable to expand to mobile and lifting tasks. Only cylindrical coordinate motion possible, not human-like arm shape. Overall computing power relies on edge computing boards; overall cost exceeds 50,000 budget. High self-weight, high load requirements for small mobile chassis. Extremely expensive, not suitable for large-scale deployment.
Best Application Scenarios Complex contact tasks, reinforcement learning, high-precision physical interaction. Mobile manipulation combined with AGV, flexible tasks in large spaces. Surgical medical applications, ultra-precise compliant assembly. Visual grasping, 24-hour continuous industrial validation. Dual-arm collaboration competitions, large-scale data collection. Replacement of expensive foreign arms, service robot integration. Low-cost reproduction of open-source tutorials such as every-embodied. Pure software algorithm verification, low-cost imitation learning. Desktop-level grasping, enterprise-level fixed workstation data accumulation. Elderly care, auxiliary operations in small home spaces. Complex mobile operations across floors and desktop heights. 3C precision assembly, cabinetless industrial deployment. Complex surface polishing, top-tier force control research.
Single Arm Degrees of Freedom (DOF) 7 7 (optional 6) 7 6 7 6 (RX series is 7) 6 6 6 4-5 6 7 7

In-depth Analysis of Table Content

  1. Embodied AI Tier (The Precision Tier):

    • Franka Research 3 and Flexiv Rizon 4 represent the current academic standard for “physical contact” simulation. Their value lies not in handling tasks, but in the high-sensitivity force feedback at joints.
    • Rokae xMate ER7 Pro (note: not the CR series) became mature in 2026. With 7-axis torque sensing and a frequency of 1kHz, it has become the preferred alternative to Franka for domestic research groups.

  2. Lightweight and Mobility Tier (The Mobility Tier):

    • Kinova Gen 3 is the most unique in this list. With a weight of 8.2kg but 7 degrees of freedom, it is the only true solution for AGV/mobile chassis.
    • Hello Robot Stretch 3 has evolved to a NUC-12 computing version in 2026, with a load capacity increased to 2kg. It is specifically designed for grasping objects within “home narrow passages” and is a game-changer for service robot papers.

  3. Embodied AI Data Tier (The Scaling Tier):

    • AgileX Piper and ViperX exist to reproduce imitation learning algorithms such as ALOHA and GELLO. Their advantage is “low cost,” allowing laboratories to purchase 5–10 units simultaneously to build a data collection array (Robot Farm).

  4. Domestic integrated solutions:

    • D-Robotics MMK2 reflects the new trend in 2026: it no longer only sells robotic arms, but offers a complete system consisting of “computing power (RDK X5) + mobile chassis + arm,” aiming to lower the entry barrier for embodied AI.

## Comparison of Global Robot Manufacturers in 2026

The current market has formed a clear “tripartite confrontation” situation:

1. Established in Europe and America: Algorithm Definers

  • Manufacturers: Franka Emika, Kinova, Universal Robots (UR).
  • Features: The software ecosystem (driven by ROS/ROS 2) is highly mature, and reproduction of the code from papers is almost “zero effort”.
  • Current Status: Although the hardware is expensive and updates are slow, it still holds significant academic influence worldwide.

2. China’s leader: The balancer of performance and cost

  • Manufacturers: Rokae, Flexiv, Unitree - Z1 series.
  • Features: Hardware parameters (load, force control) have caught up to or even surpassed those in Europe and America, and the price is highly competitive (usually 50% of foreign brands’ prices).
  • Current Status: By 2026, more and more UCAS research groups like yours will shift to domestic high-end arms (such as xMate or Rizon), as localized technical support and deep adaptation to domestic AI chips (such as Ascend, Horizon) are lacking in foreign brands.

3. New Forces in Embodied AI: Scale Leaders

  • Vendors: AgileX, UFACTORY, Ark Infinite.
  • Features: Focus on extreme cost reduction, with emphasis on Python usability and open-source community integration (such as Datawhale and Every-Embodied).
  • Current Status: Through the “low price + open source” strategy, they have quickly captured the imitation learning and desktop research market.