具身智能前沿展望：数据、模型与系统演进

Key Points

• The aim is to explore the evolution of embodied intelligence, addressing challenges in data, models, and systems.
• Reviewed the current state of embodied intelligence across four dimensions: data, models, systems, and evaluation.
• Proposed a 'data pyramid' to structure data collection and utilization for physical understanding.
• Examined mainstream visual-language-action models for their expansion limitations and introduced the concept of 'world models'.
• Analyzed the shift from end-to-end models to layered architectures for better semantic planning and motor control separation.
• Evaluated current assessment frameworks for authenticity and reproducibility challenges.
• Introduced a structured approach to data that enhances physical common sense and behavioral mapping.
• Identified 'world models' as a promising new direction to improve physical intuition.
• Noted the transition towards layered architectures, allowing for greater flexibility in planning and control.
• Highlighted ongoing challenges in current evaluation systems, particularly in terms of realism and reproducibility.

Abstract