Embodied chain-of-thought (CoT) aims to let language reasoning guide robot control, but its effective form has been unclear. The authors revisit this question at an unprecedented scale, building a corpus of 978,743 trajectories with 226.3M samples and 2592.5 hours of data.

Their experiments show that high-level semantic reasoning alone yields marginal gains. What works is grounding that reasoning in concrete linguistic action guidance, such as end-effector movement and image-space trajectories. They also find that explicit CoT does not scale reliably when used as an autoregressive action prefix, because errors compound during inference.

To address this, they propose ERVLA, a vision-language-action model trained with a CoT-dropout strategy. The model uses rich reasoning traces during training but predicts actions directly at inference, bypassing autoregressive instability. ERVLA reaches state-of-the-art results on LIBERO-Plus (86.9%) and VLABench (53.2%), and outperforms baselines in real-robot tests, especially on semantic ambiguity and long-horizon tasks.