Figure unveils Index, a vast dataset for general-purpose robot training

Figure has officially emerged from stealth mode to introduce Index, a groundbreaking robot training dataset. This initiative aims to gather extensive real-world physical data, which is crucial for developing truly versatile robots. The company previously found that existing data vendors could not meet the specific requirements for throughput, diversity, or quality needed by its AI system, Helix, prompting Figure to build its own data collection pipeline.

Index is the rebranded public launch of an app that Figure had been testing in stealth for four months. This app allows individuals, referred to as 'Creators,' to record and upload videos of various human tasks from their homes or workplaces. The platform is now available on both Google Play and the App Store, marking a significant step in Figure's strategy to scale its data acquisition efforts.

Since its stealth launch, the Index app has achieved over 264,000 downloads across more than 100 countries, accumulating over 44,000 weekly active users. Creators have uploaded more than 16 million videos to the platform, with the app processing 30 minutes of video uploads every second, which translates to 4.9 years of human work uploaded daily. Figure has compensated Creators with $15 million to date for their contributions.

The company has committed to investing over $1 billion in data and compute resources over the next 12 months, aiming for a 100x increase in its data collection capacity. The collected data demonstrates significant diversity, with every 1,000 hours of video containing 373 unique tasks, 1,146 distinct manipulated objects, and 116 different environments.

Index is an initiative by Figure, a company focused on robotics and artificial intelligence. The dataset is specifically designed to train Figure's proprietary AI stack, known as Helix. The company developed an exclusive pipeline to ensure high throughput, broad diversity, and strict quality standards for the data collected.

Figure asserts that generalization in AI, particularly for robotics, is fundamentally a data problem. The company believes that a truly general-purpose robot requires a vast and diverse dataset of real-world physics, which is not readily available online. Index addresses this challenge by enabling large-scale collection of physical data directly from human activities, providing the varied input necessary for Helix to learn and improve.

The diversity of data, sourced from individuals performing everyday tasks in unique environments with various objects, is crucial for developing robust and adaptable robotic systems. This approach allows for capturing long-tail variations in tasks that are difficult to pre-define, ultimately validating Figure's thesis on data-driven generalization in robotics.

Figure's long-term vision for Index is to establish the foundation for a future where robots can be ordered as a service. The company envisions a shift from human assistance for daily chores to fully automated robotic solutions. The continuous expansion and refinement of the Index dataset will be key to achieving this ambitious goal, enabling robots to perform a wide array of tasks in diverse settings.