Are frontier VLMs trained on robotics specific data? (github.com)

Recent research has explored whether frontier Vision Language Models (VLMs) can be effectively trained using robotics-specific data, focusing on the performance comparison of these models on Franka Panda kinematics versus synthetically generated kinematics. By analyzing 50 both open and closed-source models, the study aims to uncover any discrepancies in model efficacy, revealing insights into how well these advanced AI systems can adapt to specialized tasks in robotics. This investigation is significant for the AI/ML community as it sheds light on the importance of contextual training data for enhancing model performance in real-world applications. By using joint angles sampled from DROID and employing OpenRouter for model queries in November 2026, the researchers have laid out a framework for further exploring the capabilities of VLMs in understanding and executing complex robotic movements. The implications of this study could influence future VLM designs, encouraging a shift towards more tailored training datasets that improve operational accuracy and relevance in robotics.