Excited to share that our paper "Heterogeneous Policy Networks for Composite Robot Team Communication and Coordination" has been published in IEEE Transactions on Robotics, Vol. 40, 2024. 🤖 Want to learn about improving heterogenous robot team coordination, follow this🧵. (1/ )

Heterogeneous robot teams include agents that perform different tasks and/or have different sensory information. This diversity can be very powerful, however, it can lead to significant challenges in coordinating through communication, as agents may "speak" different languages.

In prior work, we introduced HetNet, graph-attention architecture for multi-agent reinforcement learning (MARL) with communication. HetNet helped heterogenous teams learn efficient communication protocols setting new SoTA in composite team performance. (3/)

In this new work, we create novel MARL frameworks to support generating coordination policies for heterogeneous robot teams. We introduced a multiagent PPO variant (MAH-PPO) that enhances policy learning in difficult environments, improving sample efficiency & training speed.(4/)

We use fully convolutional networks (FCNs) to process tensorized representations of agent message encodings, making HetNet scalable. This improves policy transfer to large-scale domains, boosting sample efficiency by 389.51% & reducing training time by 80.63%. (5/)

Taken together, our empirical results show HetNet sets a new SOTA in learning emergent heterogeneous cooperative policies by achieving improvement of 5.84%–707.65% over baselines in small domain configurations & 32.5%– 1161.35% across larger-scale domain configurations. (6/)

We adapted HetNet for noisy communication channels & introduced new communication loss formulations (range-based & type-based) to thoroughly assess noise robustness. Even with a poor signal-to-noise ratio, HetNet's performance only dropped by 4.57%, showing its robustness. (7/)

Finally, we demonstrated our framework on the Robotarium platform with real robots. Our HetNet-based agents coordinated seamlessly, detecting and extinguishing fires through effective communication and teamwork. Check out the demo in our paper. (8/8) Link: ieeexplore.ieee.org/document/10606…

This work was developed in collaboration with Esmaeil (Esi) Seraj, Rohan Paleja, Matthew Gombolay, Zahi Kakish, and Luis Pimentel, Kin Man Lee, Zheyuan Wang, Daniel Martin, Matthew Sklar, and John Zhang.

Check out our new paper on Multi-Agent Reinforcement Learning in IEEE T-RO, entitled "Heterogeneous Policy Networks for Composite Robot Team Communication and Coordination.” Robotics@GT is Back from Sabbatical! Georgia Tech Computing CORE Robotics Lab ieeexplore.ieee.org/document/10606…

Amazon Robotics + Covariant Exciting!! Very, very exciting..!! Super excited for our FOMO team. Covariant

📢 If you're attending Conference on Robot Learning 2024 next week in Munich & interested in working at the cutting edge of Robotics Research, AI/ML/CV, Foundation Models for Robotics & more, ping me, and let's meet up for a chat! We're actively hiring highly motivated & talented candidates for

Nothing like the Conference on Robot Learning day 1 morning vibes… always a pleasure to be here and looking forward to hearing all about your amazing works! P.S.1: make sure to stop by the Amazon Robotics booth for a quick chat and learning about our work! P.S.2: hit me up for a coffee if

The inverted pendulum theory of robotics… . . Conference on Robot Learning 2024 keynote #CoRL2024

#CoRL2024 set a new level! When I watched the presentation at the Oral session, I immediately knew they will win the outstanding paper award. Congratulations to the authors and Toyota Research Institute (TRI) Conference on Robot Learning. One of the all-time best things (not just presentations) I’ve ever seen..