In multi-agent systems, getting independent entities to cooperate effectively is notoriously difficult. Traditional approaches often rely on agents sharing their current observations or historical states. However, this creates a fundamental coordination lag: by the time an agent receives a teammate's state data, that data may already be obsolete. Furthermore, when agents make decisions simultaneously, they fall into a circular dependency—Agent A chooses an action based on what it thinks Agent B will do, while Agent B is simultaneously trying to guess Agent A's next move.
To break this deadlock, a prominent research direction focuses on sharing future intentions rather than current states. By communicating or predicting planned trajectories, sub-goals, and intended actions, agents can proactively align their behaviors, satisfy joint constraints, and maintain global coherence even under severe communication constraints.
The Myopia of State Sharing
Sharing raw sensory data or current environmental states is both bandwidth-heavy and myopic. As Edmund H. Durfee et al. noted in their foundational work, Increasing Coherence in a Distributed Problem-Solving Network, exchanging high-level summaries of planned future activities allows nodes to avoid redundant tasks and focus on complementary activities before conflicts occur.
This forward-looking approach is particularly vital when communication networks are imperfect. In real-world deployments, network lag can cause standard state-sharing protocols to fail. To address this, Kai Lv et al. demonstrate in CoDe: Communication Delay-Tolerant Multi-Agent Collaboration via Dual Alignment of Intent and Timeliness that high-level behavioral intent remains valid over longer horizons, providing a stable signal for teammates to anticipate future actions despite network delays. Furthermore, Marwa Chafii et al. show in Emergent Communication in Multi-Agent Reinforcement Learning for Future Wireless Networks that sharing intentions mitigates the non-stationarity inherent in multi-agent reinforcement learning (MARL), stabilizing the learning process by making peer behavior predictable.
Proactive Coordination and Trajectory Alignment
When agents share their future intentions, coordination shifts from a reactive paradigm to a proactive one. Instead of waiting for a conflict to manifest physically—such as traffic congestion or robot collisions—agents resolve conflicts ahead of time in the planning space.
For example, in urban traffic light control, Zhengxu Yu et al. develop MaCAR: Urban Traffic Light Control via Active Multi-agent Communication and Action Rectification, demonstrating that sharing intended actions allows neighboring intersections to account for future traffic flows before they cause physical gridlock. In autonomous driving, Dechen Gao et al. show in Ego-centric Learning of Communicative World Models for Autonomous Driving that broadcasting planned waypoints acts as a safety force multiplier compared to sharing current state data alone, allowing vehicles to maintain higher safety margins.
This proactive adjustment is often formalized by treating the planned trajectories of peers as dynamic constraints:
- Local Motion Planning: D. Itoh et al. in Development of Multi-Vehicles Navigation System based on Local Information Management (1st Report) incorporate neighbors' planned paths directly into local motion planning algorithms, eliminating reactive stop-and-wait behaviors.
- Formation Control: Seung‐Mok Lee et al. utilize predictive trajectory exchange in Cooperative Particle Swarm Optimization-based Model Predictive Control for Multi-Robot Formation to maintain group cohesion and formation stability decentralizedly.
- Joint Constraints: When safety or fairness constraints span multiple agents, Zhaoxing Yang et al. show in DeCOM: Decomposed Policy for Constrained Cooperative Multi-Agent Reinforcement Learning that sharing intended "base actions" allows neighbors to adjust their local actions to satisfy joint constraints before final execution.
- Infrastructure Assistance: Mai Hirata demonstrates in Roadside-assisted Cooperative Planning using Future Path Sharing for Autonomous Driving that sharing high-resolution planned trajectories with a roadside unit enables proactive conflict resolution before vehicles enter the same physical space.
Breaking Circular Dependencies via Commitment
Simultaneous decision-making often causes agents to pick incompatible pieces of different plans. Intention sharing can break this loop by structuring the communication sequentially, effectively converting a simultaneous game into a sequential one.
Zeyuan Liu et al. address this in Multi-Agent Sequential Decision-Making via Communication by introducing a leader-follower framework where higher-level agents broadcast their selected actions before followers move, transforming the coordination problem into a Stackelberg game.
Building on this concept, Gang Ding et al. argue in Multi-Agent Coordination via Multi-Level Communication that communicating finalized, committed actions provides a more reliable foundation for coordination than sharing noisy or uncertain intentions. When follower agents receive the ground-truth action of a leader, they can optimize their own responses without the noise or ambiguity associated with predicting teammate behavior.
Communication-Efficient Latent Intentions
Broadcasting full, high-resolution trajectories or continuous plans can overwhelm communication bandwidth. To make intention sharing practical, researchers compress intentions into compact latent representations or high-level goals.
Tianshu Chu et al. show in Multi-agent Reinforcement Learning for Networked System Control that sharing abstract intentions and beliefs helps agents build accurate mental models of teammates far more efficiently than sharing raw observations. In ToM2C: Target-oriented Multi-agent Communication and Cooperation with Theory of Mind, Yuanfei Wang et al. use a Theory of Mind framework to communicate high-level intentions rather than raw sensory data, focusing communication strictly on resolving goal-level conflicts. Similarly, Murtaza Rangwala et al. demonstrate in Learning Multi-Agent Communication through Structured Attentive Reasoning that communicating internal intentions allows agents to negotiate shared resources like intersections efficiently.
To further save bandwidth, hybrid approaches combine explicit communication with local prediction. Minglong Li et al. propose Integrating Decision Sharing with Prediction in Decentralized Planning for Multi-Agent Coordination under Uncertainty, where agents explicitly share intentions for immediate steps but use heuristic-based predictions for distant horizons, exponentially decreasing the volume of data exchanged.
Internal Teammate Modeling: Coordination Without Communication
The ultimate extension of intention sharing is internal teammate modeling, where agents use internal simulations to predict their peers' intentions without active data exchange. This is crucial when communication is completely severed or prohibited.
Laëtitia Matignon et al. demonstrate this "empathy" mechanism in Coordinated Multi-Robot Exploration Under Communication Constraints Using Decentralized Markov Decision Processes, where robots simulate their teammates' decision-making processes based on their last known states to maintain coordination during signal dropouts. Similarly, Mustafa O. Karabag et al. in Planning Not to Talk: Multiagent Systems that are Robust to Communication Loss utilize "imaginary play" to maintain internal simulations of teammate behavior, allowing agents to execute complex, long-horizon plans during extended periods of isolation.
Modern generative and sequence models have significantly enhanced this capability:
- Diffusion Models: In MADiff: Offline Multi-agent Learning with Diffusion Models, Zhengbang Zhu et al. show that decentralized agents can achieve high levels of coordination by generating internal simulations of their teammates' likely future trajectories alongside their own plans.
- Sequential World Models: Zijie Zhao et al. leverage an autoregressive chain of individual agent models in Empowering Multi-Robot Cooperation via Sequential World Models to predict future trajectories as intentions, enabling spontaneous role division and predictive adaptation without explicit rules.
- Decision Transformers: Even in ad-hoc settings with unknown partners, predicting intentions remains vital. Xinzhi Zhang et al. show in Ad Hoc Teamwork via Offline Goal-Based Decision Transformers that predicting future global states as sub-goals explicitly communicates team intent and direction, leading to robust collaboration without online exploration.
- Heuristic Maps: To facilitate this without explicit communication channels, Mehul Damani et al. utilize predictive maps of neighbors' likely future positions generated from single-agent pathfinding heuristics in PRIMAL$_2$: Pathfinding Via Reinforcement and Imitation Multi-Agent Learning - Lifelong.
Finally, Zixian Ma et al. introduce ELIGN: Expectation Alignment as a Multi-Agent Intrinsic Reward, where agents use their own internal dynamics models as a proxy for the expectations of others, rewarding themselves for acting predictably to facilitate decentralized coordination. While predicting intentions through learned dependency structures remains an open challenge to eliminate communication overhead entirely, as noted by Lipeng Wan et al. in Deep Hierarchical Communication Graph in Multi-Agent Reinforcement Learning, the shift from sharing what is to sharing what will be represents a fundamental leap in multi-agent intelligence.
Key Foundational Work
Increasing Coherence in a Distributed Problem-Solving Network (1985) — contributed 2 concept elements to this direction.
Local Search in the Coordination of Intelligent Agents (1994) — contributed 1 concept elements to this direction.