GPT-6 Astra on robot arms (openai.robocurve.org)

OpenAI's GPT-6 Astra has demonstrated significant advancements in robotic manipulation by completing tasks with a new YAM robotic arm setup. In tests involving picking up and placing objects, Astra successfully placed a red block into a bowl in 19 out of 20 trials, greatly outperforming previous models, Fable 5.1 and Fable 5. The task was completed in an average of 2.5 minutes at an estimated cost of just $0.94 per run. However, Astra's performance dipped on a more complex task involving placing a puzzle piece in a groove, achieving only 2 out of 20 completions, mirroring the limitations seen in older models. The outcomes from this comparison are significant for the AI/ML community, as they highlight the progress in reinforcement learning and fine-tuning control in AI-driven robotics. The advanced capabilities of GPT-6 Astra to manipulate objects efficiently and economically open new avenues for applications in automation, manufacturing, and beyond, although the challenges with more intricate tasks illustrate that further developments are still needed. The study utilized a bimanual setup with six degrees of freedom per arm, incorporating absolute end-effector poses and multiple camera views for improved manipulation accuracy.