CT-VAM: A Cerebello-Thalamic-Inspired Vision-Action Model for Efficient Visuomotor Control

A research team has proposed a compact vision-action model for robot control that separates high-level task understanding from low-level motor execution, drawing inspiration from the brain’s cerebello-thalamic pathways [1][2]. The model, called CT-VAM, is designed to function as a local execution policy that predicts action sequences from dual-view visual data, proprioceptive feedback, and a lightweight task condition [1][2]. The architecture introduces a component named TARS, or Thalamic Action Routing Stream, a stream-separated conditional attention decoder that independently routes action, visual, and task information to prevent dense sensory tokens from overwhelming compact task-relevant signals [2]. The researchers argue that raw language is primarily needed to specify task intent, not to be repeatedly processed during high-frequency low-level execution, a separation that could enable a cloud-edge paradigm where large models handle semantic reasoning while fast closed-loop control runs on local hardware [1][2]. With 68 million parameters, CT-VAM achieves success rates on the LIBERO benchmark that are competitive with substantially larger vision-language-action models, while reducing inference latency [1][2]. The system also incorporates flow-consistent inpainting for asynchronous chunk execution, supporting high-frequency control and demonstrating robust real-world deployment on resource-constrained robotic platforms [2]. The paper was submitted to the arXiv preprint repository on June 8, 2026 [1]. arXiv, which began operating in August 1991, is an open-access repository of electronic preprints that are moderated but not peer-reviewed, and it passed the two-million-article milestone by the end of 2021 [6]. The repository now receives approximately 24,000 submissions per month as of November 2024 [6]. Large language models, which underpin many modern vision-language-action systems, are neural networks trained on vast text corpora for tasks including generation, summarization, and translation, and are typically built on transformer architectures [8]. The CT-VAM work departs from that trend by minimizing reliance on language processing during real-time control, instead using a task condition that is lightweight enough for edge deployment [2].