Goldman Sachs report on data requirements for production-level AI robotics
Overview
A Goldman Sachs report highlights that general-purpose AI robotics may require approximately 10 million hours of high-quality, multimodal interaction data to achieve production-level capabilities. The report identifies key challenges in physical interaction data, spatial perception, and real-time execution, suggesting that advances in AI models alone may not be sufficient for commercial viability without improvements in physical precision and reliability.
From Robotics — Web news discovery
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · interaction data hours: 10000000 hours · Basis not reported · Differing source assertions
“AI robotics may need 10 million hours of interaction data to reach production-level capabilities: Goldman Sachs”
Exact source · revision 2Source owner not reported
Reported quantity · interaction data hours: 10000000 hours · Basis not reported · Differing source assertions
“New Delhi [India], October 30 (ANI): General-purpose artificial intelligence (AI) robotics could require an estimated 10 million hours of high-quality physical interaction data to reach production-level capabilities, highlighting the challenges in achieving commercial viability, even as technological progress in the sector is projected to be two to four times faster than the historical pace of autonomous driving tech”
Exact source · revision 2Source owner not reported
Reported quantity · interaction data hours: 10000000 hours · Basis not reported · Differing source assertions
“ber 28 interaction with a robotics algorithm engineer, the report estimated that embodied AI would require approximately 10 million hours of high-quality, multimodal interaction data to achieve generalised production capabilities. "Embodied AI will likely require an estimated 10 million hours of high-quality, multi-modal interaction data," the report noted, emphasising the importance of reliable physical interaction”
Exact source · revision 2Source owner not reported
Reported quantity · training_data_hours: 500000 hours · Basis not reported · Differing source assertions
“全球能用的机器人训练数据，截至2026年初，只有约50万小时。”
Exact source · revision 2Source owner not reported
Reported quantity · training_data_hours: 10000000 hours · Basis not reported · Differing source assertions
“要训练一个实用的具身智能机器人，至少需要1000万小时的数据，缺口超过99%。”
Exact source · revision 2Source owner not reported
Reported quantity · data collection hours: 500000 hours · Basis not reported · Differing source assertions
“行业里流传的一组数字：国内真实物理交互场景的合规数据只有约50万小时，而机器人真正商业化落地需要数千万小时。”
Exact source · revision 2Source owner not reported
Reported quantity · data collection hours: 1000000 hours · Basis not reported · Differing source assertions
“Meta 那个用视频学习物理规律的模型，一口气喂了超过100万小时，业界认为离终点还远。”
Exact source · revision 2Source owner not reported
Reported quantity · data collection cost: 200000000 other · Basis not reported · Differing source assertions
“一条30秒的真机操作数据，采集成本在10到15元；业内测算，要把数据量凑到20万小时，得花2亿元以上。”
Exact source · revision 2Source owner not reported
Reported quantity · data collection cost: 10 other · Basis not reported · Differing source assertions
“一条30秒的真机操作数据，采集成本在10到15元；业内测算，要把数据量凑到20万小时，得花2亿元以上。”
Exact source · revision 2Source owner not reported
Reported quantity · data collection hours: 22 other · Basis not reported · Differing source assertions
“可这20万小时连续放起来，不过22年——对于预训练来说，仍然是杯水车薪。”
Exact source · revision 2Source owner not reported
Evidence dependency changes (6)
- 2026-10-30T19:30:06.251Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:30:01.716Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:25:33.732Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T19:25:20.458Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T17:28:53.506Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T17:27:23.378Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
Report timeline
Follow attributed reports and material updates.
- Robotics — Web news discoverySignalAI Robotics May Need 10 Million Hours of Interaction Data to Reach Production-Level Capabilities
A Goldman Sachs report highlights that general-purpose AI robotics may require approximately 10 million hours of high-quality, multimodal interaction data to achieve production-level capabilities. The report identifies key challenges in physical interaction data, spatial perception, and real-time execution, suggesting that advances in AI models alone may not be sufficient for commercial viability without improvements in physical precision and reliability.
- Chinese robotics — Dataset and collection discoveryPhysical AI: Robots Lack Touch, Not Intelligence
This article discusses the development of physical AI through the creation of high-quality datasets for real-world interaction. It outlines the challenges of data scarcity, cost, and the three main approaches to data collection: real-world teleoperation, simulation, and human video recording. The article also highlights the strategic initiative by Jiangsu province to build these datasets and the importance of standardization, data sharing, and ethical considerations in the development of physical AI systems.
- Chinese robotics — Dataset and collection discoverySignalRobotics Training Data Gap and Emerging Data Collection Infrastructure
As of early 2026, global robot training data amounts to about 500,000 hours, far below the 10 million hours needed for practical embodied intelligence. Data collection is expensive and labor-intensive, with new dedicated hardware and infrastructure emerging to address these challenges. The article details recent data sets, training facilities, and policy support for the field.
Event coverage history
There is not enough continuous observation data to show a trend.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).