GPT-6 Astra's Robotic Capabilities May Stem from Specialized Training, Study Suggests

Recent research indicates that GPT-6 Astra's robotic capabilities may stem from specialized training data rather than emergent general intelligence. By testing 50 models on forward and inverse kinematics tasks with varying robot structures, researchers found that GPT-6 Astra, along with GPT-6.1 Sol and Gemini Argon, consistently produced results matching a specific robot (Franka Panda) even when provided with parameters for different robots. This suggests these models have overfit to particular robotic embodiments rather than mastering generalizable kinematic laws. The study highlights the distinction between memorizing specific robot structures and understanding universal physical principles, urging caution in interpreting VLA model performance as evidence of true cross-embodiment capability.