MicrocosmAI is working on developing embodied artificial agents that, through their interaction with other agents and their physical environment, exhibit emergent behaviors, like communication and cooperation. These agents are designed to tackle challenging aspects of their tasks, which include learning to control their bodies, memorizing past experiences, as well as predicting future events. However, their primary objective is to develop a novel communication protocol grounded in their interaction with each other and their physical world. This emergent communication facilitates collaboration on cooperative tasks — a complex challenge in the field of AI. We believe this to be a necessary step towards genuine machine cognition and human-level comprehension for agents that are able to navigate through the real world and pragmatically communicate in a grounded language with other agents and people.
MicrocosmAI’s embodied language agents leverage neural models, trained through Reinforcement Learning, to perceive and navigate their environments effectively. This emergent communication facilitates collaboration on cooperative tasks — a complex challenge in the field of AI. To address this challenge, we utilize several state-of-the-art techniques, including World Models, Hindsight Experience Replay, and our distributed version of the Proximal Policy Optimization algorithm for multi-agent scenarios.
Design embodied AI agents that
learn and adapt by interacting with
simulated physical environments
to develop decision-making skills.
Create diverse environments and
custom AI models, where agents
can cooperate or compete in
complex problem-solving tasks.
Analyze how fundamental
principles of intelligence
and adaptability arise from
simple multi-agent dynamics.
Commit to open-source solutions
for science, openly share research,
and accelerate innovation and
collaboration in the AI community.
Inspire and educate a new
generation of AI practitioners
with hands-on learning and
innovative student projects.
Effectively communicate the
impact and potential of our
research to both technical and
non-technical audiences.