All intelligence is collective intelligence.

Keisuke Okumura, Guang Yang, Zhan Gao, Heedo Woo, Amanda Prorok

We present concrete planning, a hybrid approach that captures real-world continuous dynamics while maintaining scalable guaranteed planning via discrete search. The framework integrates advances in robot dynamics learning, optimal control, and anytime complete planning into a modular system deployed with 40 robots — 20 aerial, 8 ground, and 12 obstacle robots — operating in a compact laboratory space.

Michael Amir, Matteo Bettini, Amanda Prorok

An investigation of when and why behavioral diversity benefits cooperative multi-agent learning, establishing conditions under which heterogeneous policies outperform homogeneous ones.

Rishabh Jain, Keisuke Okumura, Michael Amir, Pietro Liò, Amanda Prorok

Pairwise interactions captured by standard GNNs miss higher-order dependencies in multi-agent pathfinding. This work introduces hypergraph neural networks that model group-level constraints, improving solution quality for dense navigation scenarios.

Michael Amir, Manon Flageat, Amanda Prorok

A method for embedding remotely detectable watermarks into robot policies, enabling verification of policy provenance without requiring access to the model parameters.

Rishabh Jain, Keisuke Okumura, Michael Amir, Amanda Prorok

Historically, learning-based approaches to multi-agent pathfinding have struggled to outperform classical search-based methods. In this work, we introduce a hybrid approach that combines GNNs with a classical search algorithm, to achieve superior performance to both purely learning-based and classical methods.

Jasmine Bayrooti, Sattar Vakili, Amanda Prorok, Carl Henrik Ek

A no-regret Thompson sampling algorithm for finite-horizon Markov decision processes using Gaussian process models, providing efficient exploration in model-based reinforcement learning settings.

Amanda Prorok

The current trend toward generalist robot behaviors powered by massive, monolithic AI models is unsustainable. This viewpoint article argues for a paradigm shift toward collective robotic intelligence — a "mixture-of-robots" approach where diverse, specialized agents learn and work together, inspired by natural systems where cooperation among specialized components leads to greater intelligence and resilience.

Michael Amir, Guang Yang, Zhan Gao, Keisuke Okumura, Heedo Woo, Amanda Prorok

A hybrid, decentralized framework combining optimization-based control with adaptation provided by MARL. Rather than discarding expert controllers, ReCoDe improves them by learning additional dynamic constraints that capture subtler behaviors — for example, constraining agent movements to prevent congestion in cluttered scenarios.

Yixiao Zhang, Keisuke Okumura, Heedo Woo, Ajay Shankar, Amanda Prorok

An optimization method for generating kinodynamically feasible and collision-free multi-robot trajectories that exploits an incremental denoising scheme from diffusion models. Evaluated for differential-drive and holonomic teams with up to 16 robots in 2D and 3D worlds.

Joshua Bird, Jan Blumenkamp, Amanda Prorok

A decentralized visual monocular SLAM system for multi-agent systems, enabling collaborative mapping and localization without centralized infrastructure.

Steven Morad, Ajay Shankar, Jan Blumenkamp, Amanda Prorok

Offline reinforcement learning approach conditioned on natural language commands for decentralized multi-robot navigation, bridging high-level human instructions and low-level robot coordination policies.

Jasmine Bayrooti, Carl Henrik Ek, Amanda Prorok

An efficient model-based reinforcement learning method using optimistic Thompson sampling to balance exploration and exploitation in complex sequential decision-making problems.

Zhan Gao, Guang Yang, Amanda Prorok

Treats the environment as a co-decision variable alongside agent policies, jointly optimizing both the physical layout and the decentralized navigation policies for multi-agent systems.

Matteo Bettini, Amanda Prorok, Vincent Moens

A standardized library for benchmarking multi-agent reinforcement learning, enabling seamless mixing and matching of MARL algorithms, tasks, and models while maintaining rigorous reproducibility and standardization.

Zhan Gao, Guang Yang, Jasmine Bayrooti, Amanda Prorok

A provably safe online method for multi-agent navigation in unknown environments, combining control barrier functions with decentralized planning to guarantee collision avoidance while maintaining liveness.

Jan Blumenkamp, Steven Morad, Jennifer Gielis, Amanda Prorok

A decentralized visual spatial foundation model enabling real-time, platform-agnostic pose estimation and spatial comprehension for autonomous robots. Provides accurate pose estimates and local bird's-eye-view without needing camera overlap, deployed on wheeled platforms and a quadruped.

Matteo Bettini, Ryan Kortvelesy, Amanda Prorok

DiCo (Diversity Control) controls behavioral diversity to an exact value of a given metric by representing policies as the sum of a parameter-shared component and dynamically scaled per-agent components. Applied directly to the policy architecture, leaving the learning objective unchanged.

Jan Blumenkamp, Steven Morad, Jennifer Gielis, Amanda Prorok

An agile multi-robot research platform built on DJI RoboMaster hardware, designed for real-world experiments in decentralized multi-agent coordination, supporting GNN-based policy deployment and sim-to-real transfer.

Ryan Kortvelesy, Steven Morad, Amanda Prorok

A generalized f-mean aggregation framework for graph neural networks that subsumes common aggregation functions (sum, mean, max) as special cases while enabling learnable, task-adaptive aggregation strategies.

Steven Morad, Ryan Kortvelesy, Stephan Liwicki, Amanda Prorok

A new memory model for RL that serves as a super-efficient plug-in replacement for RNNs and transformers. Runs up to two orders of magnitude faster with linear space complexity, and sets new records on the POPGym benchmark for partially observable RL.

Matteo Bettini, Ajay Shankar, Amanda Prorok

We study cooperative multi-robot tasks where the team is composed of agents with structurally different observations, action spaces, and reward functions. We introduce HetGPPO, a parameter-sharing paradigm enabling heterogeneous behaviors in multi-agent reinforcement learning with graph neural networks, achieving superior performance over role-blind baselines.

Steven Morad, Ryan Kortvelesy, Matteo Bettini, Stephan Liwicki, Amanda Prorok

A comprehensive benchmark suite for partially observable reinforcement learning, providing a unified evaluation framework across diverse POMDP environments to compare memory-based RL architectures.