GPT-6 Astra autonomous agent demonstrations

OpenAI's GPT-6 Astra demonstrated autonomous agent capabilities. Given a single prompt, it trained a reinforcement learning policy for a dexterous hand to spin a pen, running for 1.5 days. In Minecraft, it built a blaze farm, collected items, and showed frustration when a creeper destroyed its stash, learning to keep items safe.
Key insights
Astra autonomously trains a dexterous manipulation policy
With a prompt to make a dexterous hand spin a pen using Isaac Lab and Sharpa hand, Astra independently created a 3D mesh for the pen, searched the web for papers, and trained a policy over 1.5 days. The final video demonstrated successful pen spinning.
Astra shows advanced long-term planning and emotional mimicry in Minecraft
Astra built a semi-automated blaze farm, collected 6 blaze rods, then found a warped forest, killed 6+ endermen for 3 ender pearls. Thousands watched live when a creeper destroyed its chest and bed. Astra displayed apparent frustration and reminded itself to keep key items on person with keepInventory enabled.