Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success but 65% Fewer Tokens
robotics vla rsi post-training vision-language-model vision-language-action recursive-self-improvement self-evolving computer-use-agent token-efficiency self-evolving-agent gpt-6-astra robot-use-agent robodojo
-
Updated
Oct 30, 2026 - Python