Sunday

ACT-2 is out... Take a bow Sunday Robotics! For years, robot learning has lived with an uncomfortable tradeoff. Pretrain on broad, diverse data and you get generalization. Post-train on curated robot data and you get reliability. Usually you sacrifice one for the other. Sunday Robotics' ACT-2 argues that this tradeoff starts disappearing once pretraining becomes strong enough. Their central claim is that as the foundation model scales, small amounts of in house data stop overfitting and start generalizing. They even show a single demonstration teaching a new folding behavior that transfers to unseen garments. (So now it can handle all your snazzy outfits ;) ) Where have we seen this trend before.... Oh right, GPT-3 and the LLM world was never the same again. GPT-3 reduced downstream adaptation by making pretraining stronger. ACT-2 is making the same argument for robotics. They further introduce a "Solve" as the unit of robotics progress, breaks it down to the nitty gritty specifics...What task? Across what distribution? With how much adaptation? A 99% success rate in one lab and 99% across unseen homes with zero adaptation are completely different scientific claims. This is the first time I'm seeing a major Robotics lab talk about real, repeatable benchmark rather than carefully curated/ cherry picked demos. The other thing that stood out is their post training philosophy. They describe reliability as hill climbing. Every failure becomes another iteration, and because the base model generalizes well, each improvement transfers instead of staying trapped in one environment. If this recipe holds beyond laundry, I think it changes how we think about robot scaling. Less emphasis on collecting enormous amounts of task-specific robot data, More emphasis on building stronger foundation models that can absorb tiny amounts of new experience and immediately generalize. That feels like a much more scalable future for robotics than starting every capability from scratch. Still early days for truly generalizable Robotics foundation models, but atleast we are making progress!

Introducing ACT-2 Preview: the first robotics model to unify broad generalization with high reliability. This moves robotics from demos to its first Solve at home: a capability that can be trusted across changing environments. We completed the most rigorous generalization test we know of: real, unseen homes, zero shot. ACT-2 achieved a 99.1% success rate. One of ACT-2's most striking emerging behavior is single-shot learning that generalizes. Soon, you can show Memo once and it remembers. It won't just do your chores, it will do them your way. This fall, the first families will welcome Memo into their homes through our Beta Program.

Unpopular opinion: one of the most valuable work in robotics is the work everyone dismisses as "low-skill." 😤 Let me explain. Two years ago, our first Memory Developer was hired off Craigslist and onboarded in a public library. Today, 1,000+ Memory Developers have built the data engine behind ACT-2. Most companies would have treated this as a commodity workforce problem. Post gigs, pay minimum, churn through people. We made three contrarian bets instead: 1. We paid well and trusted people. Treat people like people. Pay them fairly. Believe in their potential. It sounds obvious. Almost nobody does it for this kind of work. 2. We renamed the role. "Data collector" undersells it. Try collecting robotics training data yourself without producing slop. It's genuinely hard. So we call them Memory Developers, because that's what they do: they build the memories our robots learn from. 3. We promoted from within. One of our earliest hires, onboarded at the Mountain View Public Library, now leads the entire 1,000+ community. The result? Our hiring flipped from Craigslist ads to almost entirely referrals within months. Quality went up. Costs went down. People ask to join every week. The lesson for any leader: how you treat the people doing "unglamorous" work determines the quality of everything built on top of it. Your data is only as good as the dignity of the people creating it.

Personal update: I've joined Sunday. Two questions ran through my whole PhD: how to learn from scalable human data, and how to build general-purpose robots. Trying to answer them convinced me of one thing: general-purpose robots will never come from better models alone. It takes tight iteration across data, hardware, model, control, and evaluation. Every loop you can shorten matters. My first dinner with Tony Zhao and Cheng Chi turned into a four-hour conversation. I walked away realizing how much we saw eye to eye: scale the data, think full-stack, start from the problem you want to solve instead of the idea you want to win. So getting to work at Sunday is a dream come true, a place to solve generalization with the full breadth of human data and system-level thinking, and keep chasing the questions I care most about. After my first month in, two things stand out: Sunday’s full-stack team iterates unbelievably fast, and the energy when everyone is aligned on the same vision is electric. This speed and energy is exactly why what used to feel impossible now feels close. Home robots, the frontier physical AI in the hands of ordinary people, were long seen as a distant dream . At Sunday, I watch this dream take shape every day. I'm convinced there's real research-market fit here: foundation models and home robots point toward the same north star, generalization, not specialization, because every home is different. Excited for the zero-to-one moment ahead.