Notes from the field

Blogs

Background evals

Measuring Background Impact on Performance

How task-irrelevant visual context affects robot performance, and whether offline action comparisons can identify harmful backgrounds.

Sunday Evals

Laundry Folding Eval Intelligence

An interactive analysis of robot evaluation failures, error classes, and performance.

Research note

How to Get Your Robot to Almost Play With Its Toys

What our experiments taught us about robot data, interventions, and policy performance.

Auto-evals

How to Make Toys Your Robots Will Play With for Hours

Building self-resetting, self-scoring environments that run without a human babysitter.