# Sunday's Resource Concentration Strategy

Sunday's Resource Concentration Strategy

Full article: https://haiguangboy.com/posts/sunday-act2

## Core Approach

Most teams choose to 'cast a wide demo net': ten tasks each done to sixty percent, relying on visual impact to fuel fundraising narratives. Sunday made the opposite bet: self-developed embodiment + data collection hardware + model + fleet all held in-house, betting on only one thing—pushing the single task of folding laundry to deployment-grade reliability before moving to the next.

## Key Results

↳ What is the organizational logic?
① Full-stack in-house: no reliance on third-party hardware or data vendors, allowing the improvement loop to run fast enough
② Cheap proxies for initial screening: validation loss strongly correlates with real-robot success rate (R²>0.9), using this proxy metric to iterate data ratios in the lab, saving the most expensive real-robot evaluations for the final step
③ Rewriting the metric: proposing 'Solve' (declared scope + adaptation cost), shifting the game rules from 'does the demo look good' to 'can reliability be measured on a self-chosen track'—a standard that happens to be the strength of full-stack companies and the weakness of pure-demo companies

Effect: 785 fully autonomous evaluations, 99.1% success rate across 9 garment categories; pretraining scale ramped from 0 to 100%, closing the in-domain vs. out-of-domain gap from 82 percentage points to 0; with the same data volume, curated 12.5% of data hit 75.6%, while uniform sampling only reached 43.8%—resource investment truly translated into a verifiable curve.

## Related Route Comparisons

- Same route · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: Route bet: strong pretraining rewrites the scaling equation—minimal internal data driving long-tail improvements validates that joint co-training on heterogeneous human-robot data is a structurally suboptimal approach
- Same route · [π0.5 shakes in place after failing to grab a spoon, while Orca's physics intuition from watching videos goes further](https://haiguangboy.com/posts/orca) `orca_2026_07`: The key to a world model is readable states

## Boundaries

· All figures self-reported, no third-party evaluation
· 99.1% covers only the folding task family—it's intra-task generalization, not cross-task emergence
· Capabilities like vacuuming or coffee-making 'not yet validated under the Solve standard'
· Excludes deformable items like socks and underwear—these edge cases are precisely the part the industry agrees is hardest to transfer

## Author's Assessment (not from the paper, synthesized across papers)

Sunday's bet directly conflicts with another known path: the Fe0 team's conclusion is that 'piling on data won't solve physical generalization, and action transfer will hit a hard ceiling'; the edge cases Sunday excluded precisely confirm that wall still stands—scaling addresses generalization at the perception and decision layers, but the physical execution layer's ledger isn't fully settled here.

It's also a three-way fork in organizational strategy: Qianxun piles up data mountains first, scaling out only when the model nears a critical point; Sunday punches through the deployment line at a single point, letting deployment itself kickstart the flywheel; Zibianliang uses human scaffolding, treating deployment as data collection. Which of the three paths solidifies the 'reliability curve' first will be clear in six months.