Current progress in humanoid robotics is centered on optimizing vision-language-action (VLA) models, integrating multimodal data, and enhancing instruction comprehension as well as the ability to interpret human intent. Training relies heavily on world models, human video data, and VR-based remote training, with increasing emphasis on first-person perspectives to strengthen perception. While the ultimate goal is to achieve general-purpose humanoids, development remains constrained by significant challenges, leading Western and Chinese companies to pursue divergent technological pathways.

Key Highlights

• Humanoid robotics focuses on optimizing vision-language-action (VLA) models and enhancing multimodal data integration.
• Improving instruction comprehension and human intent interpretation is a core development area.
• Training relies heavily on world models, human video data, and VR-based remote training, with growing emphasis on first-person perspectives.
• The ultimate goal is to achieve general-purpose humanoids, but major technical challenges persist.
• Western and Chinese companies are pursuing different technological pathways in response.

Table of Contents

1. Vision Models as the Core of Robotic Perception
   • Figure 1: Humanoid Robot Model Operation Framework
   • Figure 2: Training Data for Humanoid Robots
   • Table 1: Comparison of First-Person and Third-Person View Algorithms
   • Figure 3: Apple HAT Model Overview
   • Table 2: Summary of First-Person Datasets
2. Strategic Moves by Humanoid Robot Model Developers
   • Figure 4: ViLLA Architecture
3. TRI’s View

<Total Pages: 12>