# Closing the 100,000 year "data gap" in robotics Source: https://www.youtube.com/watch?v=Fkqlow_lQ5Y Recap page: https://rapidrecap.app/video/Fkqlow_lQ5Y Generated: 2026-01-23T14:03:20.341+00:00 --- ## Quick Overview The key to closing the 100,000x "Robot Data Gap" lies in developing Data Flywheels, which involve Innovating, Commercializing, and Collecting Data, as simulation is insufficient for manipulation tasks and YouTube data lacks 3D structure, offering a path toward general-purpose robots capable of moving, making, and maintaining things. **Key Points:** - The "Robot Data Gap" is a 100,000x disparity between the data available for solving vision/language problems (billions of hours/tokens) and the data available for robotics (e.g., 10K hours for Pi0). - Simulation fails to capture the nuances required for manipulation tasks, as demonstrated by the failure to transfer sim-trained manipulation skills to the real world (08:51). - YouTube videos, while abundant (e.g., Qwen-2.5 trained on 1.2B hours), lack the crucial 3D structure needed for robotic learning in the physical world. - Human teleoperation (teleop) is slow and tedious (1x speed for simple tasks, 4x speed for improved gripper responsiveness), equating to many person-years of effort (10:27). - The proposed solution is the Data Flywheel: Innovate -> Commercialize -> Collect Data, which generates massive amounts of high-quality robot data (e.g., AmbiSort systems collecting data 24/7). - Ambi Robotics' PRIME-1 model, trained on less than 1% of their collected data, already outperforms previous AI models, illustrating the power of this data-centric approach (15:27). - AmbiStack, an example of this system, is projected for February 2025 to handle package stacking with high accuracy by combining data from PRIME-1 and simulation RL. ![Screenshot at 15:15: The slide comparing LLM data \(up to 790M hours for Llama-3\) to robot data \(10K hours for Pi0\) graphically illustrates the massive data gap that needs closing for advanced robotics.](https://ss.rapidrecap.app/screens/Fkqlow_lQ5Y/00-15-15.jpg) **Context:** Ken Goldberg from UC Berkeley discusses the immense data deficit—the "100,000 Year Robot Data Gap"—that separates the training data available for large language models (LLMs) and computer vision from what is available for robotics, especially for complex manipulation tasks. He frames the problem by contrasting the billions of hours of text/image data available online with the mere 10,000 hours collected for robotics projects like Pi0. Goldberg outlines the limitations of current approaches (simulation, YouTube data, human teleoperation) and proposes the 'Data Flywheel' model as the necessary path forward for creating robots capable of general-purpose manipulation. ## Detailed Analysis Ken Goldberg opens by highlighting the 100,000x 'Robot Data Gap,' noting that while vision and language AI models use billions of hours of data (e.g., GPT-4 used 685M hours), robotics struggles with much less (Pi0 used 10K hours). He dismisses three current approaches to closing this gap: 1) Simulation, which fails for manipulation tasks due to inaccuracies in modeling physical contact (08:51); 2) YouTube Videos, which lack the necessary 3D structure for robotic learning (09:16); and 3) Human Teleoperation, which is slow and tedious, requiring many person-years (10:27). The proposed solution is the "Data Flywheel," which loops through Innovate, Commercialize, and Collect Data (13:06). This method, exemplified by Ambi Robotics' AmbiSort systems processing millions of items daily, generates massive amounts of real-world data. Goldberg credits this data collection strategy for the success of their PRIME-1 model, trained on less than 1% of their total data, which outperforms previous models (15:27). He concludes by showing AmbiStack (February 2025), a system using PRIME-1 and simulation RL for package stacking, demonstrating that this flywheel approach provides the path for robots to learn the necessary skills to Move, Make, and Maintain things. ### The Robot Data Gap - Population aging creates a need for robots to perform tasks like moving, making, and maintaining things (00:15); The data required for robotics (6+ D state space) vastly exceeds that for vision (2D) or language (1D), where massive training data is readily available (05:16). ### Limitations of Current Data Sources - Simulation fails for manipulation due to physical inaccuracies (08:51); YouTube videos lack 3D structure (09:16); Human teleoperation is slow and tedious, requiring many person-years (10:27). ### Data Volume Comparison - LLM data scales to hundreds of millions of hours (e.g., GPT-4 at 685M hours, Llama-3 at 790M hours), while early robot data sets like Pi0 are only 10K hours (06:10). ### The Data Flywheel Solution - Ambi Robotics proposes a cycle of Innovate -> Commercialize -> Collect Data (13:06) to continuously generate massive amounts of real-world data (14:45). ### Ambi Robotics Milestones - Founded in 2018 by five UC Berkeley co-founders (12:48); Their AmbiSort systems are deployed across the US in major logistics companies (14:21); PRIME-1, trained on <1% of their data, outperforms previous models (15:27); They forecast AmbiStack for package stacking by February 2025 (15:58). ![Screenshot at 00:02: The initial graphic establishing the theme of AI for Good Global Summit.](https://ss.rapidrecap.app/screens/Fkqlow_lQ5Y/00-00-02.jpg) ![Screenshot at 00:09: Ken Goldberg begins his presentation on "How to Close the 100,000 Year Robot 'Data Gap'".](https://ss.rapidrecap.app/screens/Fkqlow_lQ5Y/00-00-09.jpg) ![Screenshot at 00:22: Slide illustrating the demographic shift showing populations are aging, with fewer workers supporting more retirees by 2025 \(00:22\).](https://ss.rapidrecap.app/screens/Fkqlow_lQ5Y/00-00-22.jpg) ![Screenshot at 03:46: Graphic highlighting that the ability to grasp arbitrary objects is a 'Grand Challenge for Robotics' compared to the vast data available for vision and language.](https://ss.rapidrecap.app/screens/Fkqlow_lQ5Y/00-03-46.jpg) ![Screenshot at 15:14: Slide comparing the scale of LLM training data \(up to 790M hours\) against Ambi Robotics' 200K hours of robot data, equating 1PB of robot data to 22 years of collection.](https://ss.rapidrecap.app/screens/Fkqlow_lQ5Y/00-15-14.jpg)