Cost reduced to 1/10! Say goodbye to the crowd collecting data, how can the embodied data token factory run smoothly?

Intelligent Things Author | Wang Han Editor | Mo Ying In the past few years, the two large-scale developments of AI are inseparable from sufficient data supply. Large language models can directly use the web page and code data accumulated on the Internet for many years to improve model capabilities through continuous training; autonomous driving companies can rely on the shadow mode of mass-produced vehicles to continuously collect driving data and manual takeover records from real roads. Taking Tesla as an example, it integrates data collection into the daily driving process through millions of mass-produced vehicles around the world, without the need to organize large-scale manual collection. But in the field of embodied intelligence, such data supply conditions do not exist. In the past two years, the industry has mainly relied on real machine remote operation and first-person perspective (EGO) wearable devices to collect data, and the cumulative number of effective real machine data and EGO data collected is only a few million hours. For robotics companies and large model companies that have invested a lot of money, it is still not easy to obtain high-quality training data of sufficient scale. On October 30, the spatial intelligence company InSpatio launched the country’s first embodied simulation data mining center in Hangzhou. Yingsu estimates that the center will produce 1 million hours of simulation training data in a single month, and the comprehensive cost of a single piece of data is about 1/10 of traditional real machine remote operation collection. The entire industry has spent two years accumulating millions of hours of data. Why does Yingsu plan to increase production capacity to one million hours in a single month? To answer this question, we need to first see clearly the difficulties faced by current embodied intelligence data collection, and how shadow tracing can transform real-life scenes into robot training data. 1. Why is it that the current embodied industry cannot provide large quantities of data? A qualified embodied token needs to contain three elements at the same time, namely three-dimensional space observation, real physical contact feedback and executable action sequence. In other words, the data must not only record what the robot saw, but also how the robot interacted with objects and what changes occurred after the action occurred. Around these requirements, there are currently two main data acquisition paths in the industry: one is in the physical world, through real machines and wearable devices.