Sunday Robotics has previewed ACT-2, a landmark robotics model that the company describes as the first to achieve reliability by unifying broad generalization with high performance. The results represent a significant step toward fully autonomous home robot deployment.
Key Results
Across 785 autonomous attempts spanning 9 major garment types, ACT-2 achieved an overall success rate of 99.1% (±0.3% standard error), with a mean fold quality score of 4.72/5. 98.3% of completed folds met the four- or five-star quality bar, including 73.8% that received a perfect score. The model completed successful folds at a median time of 2 minutes and 13 seconds.
The Central Breakthrough: Single-Example Learning
ACT-2's central advance is that a single fine-tuning example can teach the pretrained model a new behavior that generalizes to unseen environments. The company demonstrated this by post-training four independent copies of the same base model, each receiving one demonstration of a different folding technique. All four successfully executed their newly learned techniques on held-out garments in unseen environments.
This breakthrough redefines the scaling equation for robotics: minimal in-house data can drive improvements across the real-world long tail.
Closing the Generalization Gap
The key technical insight is that scaling pretraining closes the generalization gap — the difference between in-domain and out-of-domain success after post-training. At full pretraining scale, the gap narrowed to 0%, with in-domain and out-of-domain success both reaching 100%. High-quality data subsampling proved significantly more efficient than uniform subsampling.
Performance Across Conditions
Performance remained high across all tested conditions: 100% success on shorts, long-sleeved tops, polos, and sleeveless tops; 99.0% on T-shirts; 98.8% on pants; 96.3% on leggings; and 94.7% on blouses. The model maintained reliability across different starting configurations, robot positions, and bed sheet colors.
Emergent Capabilities and Deployment
ACT-2 demonstrated emergent behaviors not explicitly programmed, including edge-case recovery, robustness under disturbance, and whole-body manipulation. The company plans to deploy Memo robots to families through a Beta Program this fall, and is already training the same base model on additional household capabilities including vacuuming, toy organization, and coffee preparation.
