ACT-2 Preview: Generalizing Reliability

Summary

Sunday Robotics blog (cite sunday2026act2preview) previewing ACT-2 on the Memo home robot. Claims the first end-to-end recipe where a single SFT demo teaches a new folding behavior that generalizes, because scaling diverse sensorized-human pretraining drives the in-domain vs OOD success gap from 82 pp (no pretrain) to 0 at full scale. Laundry folding framed as a Solve (performance, declared scope, adaptation cost): 99.1% ±0.3% zero-shot success on 785 autonomous attempts across 9 garment types (XXS–8XL) in unseen homes, zero per-home data/finetune; 778 successful folds mean quality 4.72/5, median 2 min 13 s. Discord/X is Cheng Chi’s follow-on: during ACT-2 home eval, cloud inference failed on real WiFi/Starlink, so Arpit Kalla shipped edge inference in two days; argument that on-robot inference is required once the policy is reliable enough that network jitter dominates. Closed model/robot; no public weights or paper.

Keywords

act-2 · sunday-robotics · memo · home-robot · laundry · edge-inference · solve · x · blog

Topics

home robotics, robot foundation models, generalization

Research notes

• Primary: Sunday blog. Discord/X https://x.com/chichengcc/status/2078159388288938298 via fxtwitter (7-tweet edge-inference thread; careers https://sunday.ai). Closed robot/model; access left blank rather than guess paid vs beta. Not a hosted corpus, so no datasets_local row.