Reliability in unseen homes, with no special setup, is the real test of embodied general intelligence. Sunday Robotics has previewed ACT-2, its home-robot foundation model, with a headline result: 99.1% zero-shot laundry folding across diverse, unseen homes, with zero per-home adaptation (no home- or garment-specific data, no demos, no fine-tuning at deployment). 785 autonomous attempts across 9 garment types, folds graded 4.72/5 on average, median 2m13s each. The recipe: - Scaling pretraining (on a large, sensorized human dataset) closes the gap between in-house and real-world performance, so lab gains actually hold in the wild - With that base, a single demonstration can teach a new folding technique that generalizes to unseen garments (they claim a first for end-to-end robot models) - That unlocks a fast loop: fix failures with minimal in-house data, and the fix generalizes Laundry is just the starting point: the AI is built to generalize to any household task, and the model is already learning vacuuming, tidying, and coffee making. The robot is called Memo. The Redwood City, CA based startup will begin beta deployments to families this fall.