Goldman Sachs Report Says AI Robotics Need 10 Million Hours of Data

General-purpose artificial intelligence robotics could need about 10 million hours of high-quality physical interaction data to reach production-level capabilities. A new Goldman Sachs report points out that while AI models are developing fast, physical interaction data limits, precision issues, and processing speed problems are major barriers to commercial sales.

Technical Hurdles

The report states that embodied AI needs multimodal interaction data rather than passive or noisy recordings. Unlike autonomous vehicles that mostly use steering, acceleration, and braking, advanced robots must manage dozens of mechanical degrees of freedom, continuous physical contact, and tactile feedback. The technology is currently comparable to self-driving cars before 2015 or around 2017 to 2018. However, progress in physical AI could move two to four times faster than the history of autonomous driving.

Engineering Bottlenecks

The study points to three main engineering problems: three-dimensional spatial perception, inference latency, and inference economics. Current spatial fidelity is not enough for millimeter-level physical interactions. Processing speeds also remain inadequate for real-time dynamic stabilization and microsecond-level reflexive recovery. First-person video recordings help models understand tasks but lack physical force data, while physics simulations struggle to mimic real-world friction and material deformation.

Context

Goldman Sachs published these findings in a report titled China AI Robotics: Expert Call on Frontier Foundation Models in Physical AI & Embodied Robotics, drawing on insights from a robotics algorithm engineer. The analysis suggests that collaborative robots and industrial automation systems may offer practical near-term opportunities compared to humanoid robots alone.