Embodied COT

Technical Analysis Robot performing a task via COT Large-language Models (LLMs) and Vision-Languag Models (VLMs) have recently been adpoted in robotics too, creating new models named Vision-Action-Language (VLA) models. Despite their capabilities, they still struggle to proper apply commnon reasoning, which is of mahor importance in robotics. Here comes "Robotic Control via Embodied Chain-of-Thought Reasoning", an approach that shows how instilling reasoning in VLAs can benefit robotics. Motivation of the Work Current robotic control policies often struggle to generalize when faced with tasks or environments outside their specific training data. To address this, Vision-Language-Action Models (VLAs) have emerged as the state of the art, integrating large pre-trained vision-language models (VLMs) into robotic policies to enhance robustness. However, typical VLAs function through "reactive" control, directly mapping sensory observations to immediate low-level actions without intermediate reasoning. While reasoning methods like Chain-of-Thought (CoT)—where models "think step-by-step"—have dramatically improved performance in purely linguistic tasks, applying naive CoT to robotics is ineffective. This is because robot policies require their reasoning to be strongly grounded in actual sensory observations and the robot's state, rather than relying only on purely semantic sub-tasks. Contribution and How They Tackle the Problem The core motivation is to move beyond reactive control…