A research report from Guotai Haitong Securities Co., Ltd. indicates that Figure Helix 02 has achieved a critical technological breakthrough, propelling humanoid robots from technical demonstrations towards practical application. A single neural control system has overcome the industry bottleneck of integrating mobility and manipulation. The core barriers for humanoid robots are concentrated in three key capabilities: multi-modal data perception and fusion, closed-loop stability for motion/force control, and deep integration of hardware and software. Recommended areas include: 1) Real-time dynamic perception - 3D vision, LiDAR; 2) High-precision sensing - IMU; 3) Real-time control - encoders; 4) High-precision control - force control loops; 5) Motors & actuators. The catalyst for this analysis was a new demonstration video released by Figure on March 10, showing the Helix 02 humanoid robot performing a full-cycle cleaning task in a living room entirely autonomously, without remote control or human intervention. The main points from Guotai Haitong Securities Co., Ltd. are as follows:
Helix 02 has made continuous breakthroughs in whole-body, end-to-end control. The Helix 02 model represents a fundamental leap for humanoid robots from upper-body, end-to-end control to whole-body, end-to-end control. While Helix 01 focused on upper-body manipulation, achieving breakthroughs in precise actions like grasping and assembly, it employed a sequential execution logic for walking and operating, leading to high overall response latency and limited scenario adaptability. Helix 02 integrates whole-body movement and upper-limb operation into a single control system, enabling parallel processing of tasks such as walking, balancing, moving, and manipulating. This allows the robot to perform precise operations simultaneously while moving dynamically, significantly enhancing task coherence and real-time responsiveness. Helix 02 utilizes a single neural network system that directly controls the robot's full-body actions from pixel input, enabling dexterous operation throughout an entire room and long-duration autonomous behavior. In real-world scenarios like industrial work, home services, and complex environment inspection, this parallel control mode can substantially improve operational efficiency and environmental robustness.
Multi-modal data integration, dual closed-loop stability, and hardware-software unification are the core barriers. Whole-body, end-to-end control is a central technical bottleneck in the industrialization of humanoid robots. The industry generally faces three high-barrier challenges that collectively form the core competitive moat: 1) Whole-body multi-modal data fusion: Robots are equipped with various sensors like vision, touch, force, and inertial measurement units, generating high-dimensional, highly asynchronous, and heterogeneous massive data streams. This data requires alignment, feature extraction, noise filtering, and semantic fusion within milliseconds. The costs of data acquisition, the difficulty of annotation, and the complexity of model training are far greater than for upper-body control scenarios, placing extremely high demands on data systems and algorithm frameworks. 2) Long-duration task stability and closed-loop control: The system must simultaneously couple a motion control loop, which ensures stability for walking, balancing, and moving, with a force control loop, which supports the precision and safety of manipulation. These two loops need to cooperate in real-time and adapt dynamically to resist external disturbances and eliminate cumulative errors in long-duration, unstructured environments, maintaining continuous and reliable system operation. 3) Hardware-model co-design: Whole-body, end-to-end control requires deep coupling between the hardware platform and the algorithm model, necessitating native matching from actuator performance and sensor layout to computing architecture and neural network structure.
Risks include the potential for technological progress and industrialization to fall short of expectations, as well as intensifying industry competition.
Comments