For a given instruction and image, the VLA model first generates a textual reasoning chain: high-level reasoning (task understanding → plan → sub-task) and low-level grounding (motion description → gripper position → object bounding boxes). Only after generating these steps does the model predict the next robot action. Training data is produced by a pipeline that uses foundation models (object detectors, VLMs, LLMs) to automatically annotate the Bridge V2 demonstrations with reasoning chains. The base model (OpenVLA, built on the Prismatic VLM) is then fine-tuned to generate reasoning and actions.
VLA models often generalize poorly and act as a black box, mapping observation to action without explicit reasoning. ECoT introduces an explicit, perception-grounded reasoning chain, improving generalization (by 28 percentage points absolute over OpenVLA on challenging tasks) and making the policy interpretable and correctable in natural language — without additional robot data.
Task understanding, plan development and decomposition into sub-tasks.
Movement and gripper-position descriptions linking reasoning to a concrete action.
Object bounding boxes and end-effector position providing perceptual awareness.
The VLA model that predicts the robot action after generating the reasoning chain.
Automatic annotation of Bridge V2 demonstrations with reasoning chains using foundation models.
Generating a long reasoning chain before acting slows real-time control.
Synthetic reasoning chains generated by foundation models may contain errors propagated into the policy.
Incorrect bounding boxes or gripper positions can lead to wrong actions.
Introduction of chain-of-thought for language models — the conceptual precursor of ECoT.
An open VLA model serving as the base model for ECoT.
The paper “Robotic Control via Embodied Chain-of-Thought Reasoning”; +28 pp success over OpenVLA.
Generating a long chain of reasoning tokens before each action increases inference latency compared with a direct VLA policy.
Which steps make up the chain: task, plan, sub-task, motion, gripper position, bounding boxes.
The VLA model the reasoning is built on (OpenVLA in the original work).
The base VLA model (OpenVLA / Prismatic) is a dense transformer; ECoT does not change the execution mode, it adds an explicit reasoning stage.
The reasoning chain is generated autoregressively token by token, which is inherently sequential and lengthens time-to-action.
The VLA model (billions of parameters) needs GPU acceleration for training and near-real-time inference.
