High-end robots Data is often collected using the most basic methods.
According to Business Insider, Tesla A data collection team is being used to train the Optimus robot to act like a human. During eight-hour shifts, data collectors repeat the same actions hundreds of times, including picking up cups, wiping tables, and drawing curtains. Before each shift, they receive task assignment documents and operating manuals to ensure the data collection is performed correctly. Each employee is required to collect at least four hours of usable video footage per shift.
It's worth noting that Tesla is now collecting data via cameras, rather than using motion capture suits or manual controls to train robots as it has done in the past . In June of this year, after the departure of Milan Kovac, the head of the Optimus project, the company informed the data collection team of this decision: "The scale of data collection would be much larger without motion capture suits."
Data collection work places high demands on the physical strength of employees. In addition to the cameras around the environment, each data collection employee needs to carry about five cameras on their helmet and heavy backpack to record the entire process. According to sources, at Tesla 's Fremont, California factory, data collectors wear headphones and backpacks, organize vehicle parts, and work on conveyor belts.
The long hours of physical labor have taken a toll on the employees' health. One employee revealed that their back and neck had been injured during work due to uneven weight distribution from their backpacks. Another employee pointed out that a data collector had suffered from motion sickness due to prolonged use of head-mounted displays.
Regarding the size of the data collection team, the report points out that at its peak, Tesla... More than 100 employees are engaged in data collection . Behind this is Musk's humanoid robot. Ambitious Vision: During Tesla's Q3 earnings call this year, Musk declared that the company would achieve its goal of producing 1 million Optimus units annually. He also stated that in the future, the humanoid robot business would account for approximately 80% of Tesla's value.
Data has long been considered key to improving the generalization capabilities of humanoid robots . (Huachuang Securities) It is pointed out that the fusion of multimodal training datasets will significantly improve the robot's environmental perception and multitasking capabilities. In terms of classification, the data is mainly divided into real data and simulated data. Real data is widely recognized as the "golden data" with the best training effect, but it is more expensive to collect and has inconsistent formats. In contrast, simulated data can be generated on a large scale at a low cost, but the models trained on it are often unsuitable.
Against this backdrop, "virtual-real integration" is currently the mainstream solution for data collection in the robotics industry . Liu Yufei, Deputy General Manager of the National-Local Jointly Built Humanoid Robot Innovation Center, stated that virtual-real integration training grounds have been established in eight provinces and cities across the country. Li Junlan, Research Manager at IDC China, predicts that the industry will build a solid data foundation based on massive amounts of highly realistic physical data combined with high-quality real-world data collection, thereby rapidly improving the generalization capabilities of intelligent systems.
According to Research Nester's "Data Acquisition Systems Market Size and Forecast" report released in September of this year, the data acquisition systems market size exceeded $2.4 billion in 2025 and is projected to reach $3.98 billion by 2035, with a CAGR of approximately 5.2% during the forecast period (2026-2035). The industry size for data acquisition systems is estimated at $2.51 billion in 2026.
However, at the same time, there are signs that future robot training may become "AI-driven." For example, Tesla recently announced that it is training Optimus on its self-developed world model. (Dongwu Securities) It is noted that currently, enterprises mostly use methods such as world models, remote operation, simulation transfer, and simulation training, but these all have limitations and cannot achieve generalization capabilities. Learning and training methods for embodied intelligence still need to be explored.
(Article source: CLS)
