Sunday Robotics' ACT-2 model demonstrated 99.1% zero-shot laundry folding across 785 autonomous attempts in diverse, unseen homes, achieving an average grade of 4.72/5 with a median time of 2m13s per garment. The model, built on scaled pretraining from a large sensorized human dataset, can learn new folding techniques from a single demonstration and is designed to generalize to other household tasks like vacuuming and coffee making. The Redwood City, CA-based startup plans beta deployments to families this fall.
Embodied AI is demonstrating real-world zero-shot generalization in unstructured home environments, expanding the addressable market for robotics beyond industrial settings.