As of Oct 8, MirroS released AgentGarten: executable code environments supply physics and state while a learned renderer streams first-person frames above 30 fps so agents can act, observe and revise written playbooks between rounds.
MirroS released AgentGarten: executable code environments supply physics and state while a learned renderer streams first-person frames above 30 fps so agents can act, observe and revise written playbooks between rounds. In a hide-and-seek revisit, hiders built panel shelters by round 4 and seekers used ramps by round 10—behaviors that took tens of millions of self-play RL episodes in OpenAI's 2019 study—then the same loop improved companion-dog, one-lane bridge, herding and quarry-loader tasks over four rounds. Blog, report and code are public.
Items older than 72 hours never go in Top stories and show their original date.