Core Insights - The "Embodied Intelligence" forum during the 2025 World Artificial Intelligence Conference (WAIC) highlighted key challenges and advancements in the field of embodied intelligence, featuring prominent scholars from the Berkeley alumni group [1][2][4]. Group 1: Key Challenges in Embodied Intelligence - A significant bottleneck in embodied intelligence is the acquisition of high-quality data, which is essential for training robots to perform tasks effectively [4][9]. - The "data pyramid" model for embodied intelligence training emphasizes the need for diverse data sources, with the top tier consisting of remote operation data, the middle tier comprising human behavior data, and the base layer consisting of vast internet data [11][16]. - Current data collection methods are insufficient, as the largest publicly available dataset contains fewer than 1 million trajectories, which is significantly lower than the data available for language models [16]. Group 2: Future Vision and Development Stages - The envisioned future for robots includes three stages: integration into production systems, self-manufacturing capabilities, and assisting humans in expanding their capabilities, such as space exploration [5][6][8]. - The ultimate goal is to develop multi-agent systems where multiple embodied agents can interact and collaborate, enhancing their functionality and complexity [20][26]. Group 3: Data and Model Structure Enhancements - The introduction of "TactileVLA" aims to incorporate tactile feedback into the existing VLA model, improving robots' ability to perform tasks that require touch sensitivity [17]. - A new model called "OneTwoVLA" combines fast and slow thinking processes, allowing robots to analyze tasks and execute actions more effectively [18]. - The concept of "data scaling" is proposed, focusing on improving either world sampling or path sampling to enhance the training of embodied intelligence systems [34].