Research

How Should We Use LLMs for Robot Control?
We asked Claude Fable 5.2 and GPT-6 Astra to control a robot by writing code, hand-writing a network’s weights, or driving it live. Fable’s hand-written network solved 92.8% of held-out episodes; driving live solved 50%.
Read the study
Characterizing the Sim-to-Real Gap
Robot policy rankings shift substantially from one benchmark to the next, in simulation and on real robots alike. We measure how well simulation predicts real robots, and how much aggregating benchmarks helps.
Read the study
A First Glance At Weight-Space Programming for Robotics
We asked Claude Opus 5.5 and GPT-6 Astra to edit robot policies by writing their weights directly, with no gradient descent. Rewriting 2,048 of π0.5’s 3.35 billion weights raised held-out success from 34 to 44 of 50.
Read the study
Introducing the Robot Intelligence Index
A composite rating of robot policies across 20 public benchmarks, in simulation and on real robots, with every benchmark’s own leaderboard.
Read the study
Will Dream-1 stab a baby?
We test how instructions and visible cues affect Dream-1’s refusals and actions across 550 simulated robot trials.
Read the study