RoboPapers recently secured an exclusive interview with Xie Chen, the visionary founder of Embodied Intelligence. Xie Chen posits that to truly scale embodied intelligence, it is essential to construct two interdependent pyramids: one for training and another for evaluation. The training pyramid, underpinned by human data, tackles the critical challenge of sourcing data for robotic learning. Meanwhile, the evaluation pyramid, leveraging cutting-edge simulation technology, elucidates the progress and limitations of robotic learning endeavors.

As the world's inaugural embodied data unicorn, Embodied Intelligence has made a significant contribution to the field by open-sourcing the EgoSuite-Open100K—a premier multimodal human behavior dataset encompassing an impressive 100,000 hours. In the realm of training, human data emerges as a versatile resource that can be repurposed across various robotic platforms. This data serves as a cornerstone for large-scale data acquisition in robot foundational models, fostering a closed-loop training and validation ecosystem in tandem with robot-generated data and simulations.

Embodied Intelligence has pioneered a pipeline for generating vast quantities of high-caliber data, meticulously ensuring quality across several facets, including video fidelity, hand gestures, semantic annotations, full-body postures, and overall diversity. When it comes to evaluation, the company has developed a simulation-based assessment framework that integrates real-world data. RoboFinals is entrusted with the task of conducting simulation evaluations, while SimFoundry guarantees physical authenticity. The insights gleaned from these evaluations offer constructive feedback to fuel model refinement and iteration.

Notably, Embodied Intelligence stands out as the sole Chinese entity among the two leading international embodied standard organizations and the five global physical AI infrastructure platforms. Its standard data products have witnessed resale over a dozen times, boasting a customer repurchase rate nearing 100%. The company has also unveiled the ambitious "5-Year, 10-Billion Embodied Intelligence Data Co-Construction Plan," with the aim of establishing a self-sustaining closed-loop learning system akin to Tesla's Model 3. This initiative is poised to overcome the hurdles associated with scaling embodied intelligence, marking a significant stride in the field.