Abstract
We explore Astra’s capabilities, agent design, and future possibilities in robot control. Using Ditto-Bench, a suite of six manipulation tasks, we evaluate physical reasoning and interaction, examine how control interfaces and observation context affect performance, and investigate collaboration with specialized vision-language-action models. The study highlights promising capabilities alongside limitations in precision, reliability, and latency.
Publication
Research blog
