What is it about?

This work tackles a big challenge in robotics: how to get teams of smart robots to work together efficiently in real time. These systems, called cooperative embodied agents, are used in tasks like cleaning, cooking, or exploring - but they often face delays in planning, communication, and decision-making. We introduce ReCA, a new system that helps these robot teams think and act faster by: 1. Running AI models directly on the robots (instead of relying on servers), 2. Organizing memory and teamwork better, and 3. Using custom hardware to speed up low-level actions. As a result, ReCA makes robots over 10x faster and more successful at completing complex tasks, even in large teams. It brings us closer to deploying intelligent robots in the real world.

Featured Image

Why is it important?

ReCA tackles a key challenge in robotics: how to make teams of intelligent agents collaborate efficiently and in real time. While today’s systems are powerful, they’re often too slow and don’t scale well in real-world tasks. ReCA is the first to co-optimize AI models, system design, and hardware together for cooperative embodied agents. It runs language models locally, introduces a smart memory and planning system, and uses custom hardware to speed up low-level actions. ReCA delivers 10× faster performance and better task success - enabling multi-agent systems that are ready for real-world use in homes, factories, and beyond.

Perspectives

What I find most exciting about this work is how it brings together multiple layers - algorithms, systems, and hardware - into a single, unified framework for cooperative embodied AI. Instead of treating planning, memory, and execution as separate problems, we explored how they interact and can be co-optimized for real-time, scalable performance. ReCA is meaningful to me because it shows that with thoughtful design, it’s possible to run advanced multi-agent AI systems efficiently and locally, without relying on expensive cloud calls or heavy infrastructure. From designing a hierarchical planning scheme to building a lightweight hardware accelerator, this project was about making intelligent agent collaboration practical and deployable. To me, ReCA represents not just a system, but a step toward realizing embodied agents that can operate seamlessly in the real world - working together, responding quickly, and adapting on the fly.

Zishen Wan
Georgia Institute of Technology

Read the Original

This page is a summary of: ReCA: Integrated Acceleration for Real-Time and Efficient Cooperative Embodied Autonomous Agents, March 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3676641.3716016.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page