Training Data for Agility Robotics

Agility Robotics occupies a unique position in the humanoid landscape: rather than pursuing general-purpose intelligence, they have narrowed Digit's initial deployment to warehouse logistics — a domain with predictable task structures but highly variable environmental conditions. This strategic focus means their data needs are simultaneously more specific and more demanding than general humanoid research. They do not need data covering every possible household task, but they need exhaustive coverage of the environmental variation within warehouses.

The core technical challenge for Digit in warehouse settings is robust bipedal locomotion on real warehouse floors. Commercial warehouses have polished concrete with varying friction coefficients, occasional liquid spills, uneven seams between floor sections, and dock plates with slope transitions. Digit's spring-loaded legs — a design inherited from Jonathan Hurst's PhD research at CMU on series-elastic actuators — give it compliance advantages over rigid-legged humanoids, but this compliance also means the locomotion controller is more sensitive to ground-contact dynamics. Simulated locomotion policies trained on flat terrain in Isaac Gym transfer poorly to these conditions. Agility needs hours of real locomotion data collected in actual warehouses with ground-truth foot contact measurements, IMU recordings, and environmental surface characterization.

RoboFab represents a critical inflection point. When Agility opened the world's first humanoid robot factory in Salem, Oregon in 2023, they announced capacity to produce 10,000 Digits per year. This manufacturing scale creates an urgent data scaling challenge: each deployed Digit needs policies robust enough for the specific warehouse it operates in, while benefiting from shared representations learned across the full fleet's experience. The ratio of training data to deployed units must increase dramatically as production scales.

Manipulation presents a different but equally demanding challenge. Digit's end-effectors are designed for grasping standardized warehouse totes — a seemingly simple task that becomes complex when totes vary in weight (1-25 kg), fill level, surface friction (dry versus condensation-covered plastic), and stacking configuration. A tote at the bottom of a stack behaves differently than one on top. A tote packed with heavy items has different inertial properties than one with lightweight goods. Real-world manipulation data must capture these variations across thousands of picks to build policies that generalize beyond the narrow distributions seen in laboratory demonstrations.

The human-robot interaction dimension adds a safety-critical data requirement. Digit operates in shared spaces with human workers at Amazon's BOS27 facility and GXO warehouses, requiring accurate perception of human proximity, intent prediction, and safe motion planning. This demands training data collected in authentic mixed human-robot work environments — something that cannot be ethically or practically simulated at scale.