Sunday Robotics scaling data drops generalization gap to 3% · Digg

/TECH

Sunday Robotics scaling data drops generalization gap to 3%

The Memo robot's lab-developed skills successfully transferred to thirty unseen homes.

Combined views

26.4K

2 posts, first seen 13h ago

133 likes

Sunday Robotics scaling data drops generalization gap to 3%

The Memo robot's lab-developed skills successfully transferred to thirty unseen homes.

Commentary on X

IV IvyLeagueRedNeck@Ivy_League_Red

10:31 AM · Jul 16, 2026

@tonyzzhao Good luck to http://Sunday.AI. One of many steps on the road to a world of physical AI created abundance.

MC Michael Cho - Rbt/Acc@micoolcho

3:45 PM · Jul 16, 2026

@tonyzzhao This one single Solve will get the data flywheel going, Tesla autopilot style

SA Sean Austin@stlgotmynikeson

11:40 AM · Jul 16, 2026

@tonyzzhao Soo how can I get one to beta test? I’m all in.

TZ Tony Zhao@tonyzzhao

10:22 AM · Jul 16, 2026

784TECH

This property allows us to hill-climb performance in our office, and trust those gains to hold in unseen homes Our fleet of Memos runs in parallel to rapidly advance reliability, quality, and speed. Left: fleet-scale improvement in-house Right: Memo working across unseen homes

Combined views

26.4K

2 posts, first seen 13h ago

133 likes

2 comments

19 saves

7 reposts

Sentiment

Positive100%0%Negative

Positive Read

Users are excited about scaling pretraining for Memo Robots because it promises physical AI abundance and a Tesla-style data flywheel for emergent capabilities.

Based on 3 visible X reactions from 6 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.

Commentary on X

IV IvyLeagueRedNeck@Ivy_League_Red

10:31 AM · Jul 16, 2026

@tonyzzhao Good luck to http://Sunday.AI. One of many steps on the road to a world of physical AI created abundance.

MC Michael Cho - Rbt/Acc@micoolcho

3:45 PM · Jul 16, 2026

@tonyzzhao This one single Solve will get the data flywheel going, Tesla autopilot style

SA Sean Austin@stlgotmynikeson

11:40 AM · Jul 16, 2026

@tonyzzhao Soo how can I get one to beta test? I’m all in.

TZ Tony Zhao@tonyzzhao

10:22 AM · Jul 16, 2026

784TECH

This property allows us to hill-climb performance in our office, and trust those gains to hold in unseen homes Our fleet of Memos runs in parallel to rapidly advance reliability, quality, and speed. Left: fleet-scale improvement in-house Right: Memo working across unseen homes

View all

CP Chris Paxton@chris_j_paxton

9:06 PM · Jul 16, 2026

856TECH

I think this will be the biggest lesson of the year: that most of the work we have done in the past needs to be thrown out, as it relied on a number of priors

View all

CP Chris Paxton@chris_j_paxton

9:06 PM · Jul 16, 2026

856TECH

I think this will be the biggest lesson of the year: that most of the work we have done in the past needs to be thrown out, as it relied on a number of priors

The Conversation on X

2 posts

TZ Tony Zhao@tonyzzhao

10:22 AM · Jul 16, 2026

784TECH

This property allows us to hill-climb performance in our office, and trust those gains to hold in unseen homes Our fleet of Memos runs in parallel to rapidly advance reliability, quality, and speed. Left: fleet-scale improvement in-house Right: Memo working across unseen homes

CP Chris Paxton@chris_j_paxton

9:06 PM · Jul 16, 2026

856TECH

I think this will be the biggest lesson of the year: that most of the work we have done in the past needs to be thrown out, as it relied on a number of priors

Nishant Desai@nishantsdesai

Turns out learnings from low data scale and overfit scenarios don’t translate to the high data regime. Last year we relied on inductive biases and extremely strict data collection procedures. The result was a capable but brittle policy. This year we embraced chaos and allowed our dataset to span the real world state distribution. The result was a set of capabilities we would never have predicted would emerge.

2 comments

19 saves

7 reposts

Sentiment

Positive100%0%Negative

Positive Read

Users are excited about scaling pretraining for Memo Robots because it promises physical AI abundance and a Tesla-style data flywheel for emergent capabilities.

Based on 3 visible X reactions from 6 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.