Training Data for Stanford SAIL

Stanford SAIL's robotics groups have collectively shaped the modern landscape of accessible robot learning. The ALOHA system — a $20K bimanual teleoperation rig built by Tony Zhao and Zipeng Fu under Chelsea Finn's supervision — demonstrated that graduate students can collect high-quality manipulation data without million-dollar hardware. Mobile ALOHA, costing $32K including onboard power and compute, extended this to whole-body household tasks: the robot learned to saut shrimp, clean wine spills, push in chairs, and call elevators. ALOHA Unleashed, a collaboration with Google DeepMind, pushed dexterity further by training robots to repair other robots and perform tasks requiring fine motor control.

Octo created an open-source generalist policy that anyone can fine-tune. Bridge V2 established a cross-institution data sharing standard with 60,000+ demonstrations formatted in the RLDS (Reinforcement Learning Datasets) format. Together, these projects form an ecosystem where hardware is cheap, software is open, and data is shared — democratizing robot learning in ways that previously required corporate-scale resources.

This accessibility-focused research philosophy creates a specific data bottleneck: the quality and diversity of policies trained on ALOHA is directly limited by the diversity of environments where ALOHA data has been collected. Most ALOHA data comes from a handful of university labs in the San Francisco Bay Area — the same kitchens, the same tables, the same lighting. A policy trained to fold towels in a Stanford lab kitchen overfits to those specific visual features and fails in a kitchen with different countertops, cabinet styles, or lighting conditions.

The Bridge dataset faces the same challenge at larger scale. While Bridge V2 aggregated data from multiple institutions, the contributing labs are primarily in the western United States and have similar laboratory setups. Expanding Bridge to include data from diverse global environments — different architectural styles, object categories, cultural objects, and regional home designs — would significantly improve the generalization of models like Octo that are trained on it.

Claru's distributed collection network addresses this gap directly. By coordinating ALOHA-compatible and Bridge-compatible data collection across 100+ cities with diverse home environments, Claru can provide the environmental variety that the Stanford ecosystem needs to train policies that work outside the Bay Area. Each location contributes unique visual characteristics — different countertop materials, cabinet styles, lighting conditions, household objects — that make trained policies more robust. The data can be formatted in RLDS for direct compatibility with Bridge V2 and Octo training pipelines.