Embodied Reasoning
Introducing Embodied Reasoning, a data engine that adds reasoning to robotic datasets. Training on reasoning data allows the model to iteratively reason about intermediate representations over complex tasks before predicting actions. Our reasoning engine, applicable to both open-source and proprietary datasets, improved the base OpenVLA model by 27% on the Bridge-v2 dataset.
Robot Reasoning Improves Policies
VLAs trained on reasoning datasets achieve 27% higher accuracy
STAY UPDATED
Get research updates, news, and events.