Intent Inference for Multi-Agent Coordination

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2175 papers190 conceptsfrozen 2026-10-04

In a world of autonomous actors—from self-driving cars to collaborative robots—sensors like LiDAR and cameras can tell us where an agent is, but they cannot tell us why it is there. A car slowing down at an intersection might be yielding to you, or it might be preparing for a sudden left turn. This "fog of war" in multi-agent environments is a fundamental hurdle: when internal motivations are hidden, planning becomes a guessing game.

To move beyond reactive behavior, researchers are developing frameworks that treat the internal states of others—their goals, aggressiveness, and even their level of "foresightedness"—as latent variables to be inferred in real-time.

The Problem of Hidden Intent

Most classical navigation algorithms treat other agents as moving obstacles with predictable trajectories. However, as Stephen J. Guy notes in Moving in a Crowd, real-world navigation is hindered by a lack of direct communication regarding the underlying logic of nearby agents. Without knowing a neighbor's destination, a robot cannot plan proactive maneuvers, often leading to the "frozen robot problem"—a state of indecision where the agent stops because it cannot guarantee a safe path through an unpredictable crowd.

This challenge is particularly acute in "mixed-traffic" environments. As Stuart Russell 0001 and colleagues argued in the BATmobile project, an autonomous system must maintain a "belief state" about whether a neighboring driver is distracted or planning a lane change to make safe, real-time decisions.

Modeling Agents as Games of Incomplete Information

The dominant approach to solving this is to frame interaction as a Bayesian Game—a game where players have incomplete information about each other's "type."

In Integrated Decision Making and Trajectory Planning for Autonomous Driving Under Multimodal Uncertainties, Zhenmin Huang and Jun Ma model human drivers as discrete types, such as "aggressive" or "conservative." By solving for a Bayesian Coarse Correlated Equilibrium, the autonomous vehicle can anticipate how different types of drivers will respond to its own maneuvers.

Similarly, LUCIDGames, developed by Simon Le Cleac'h and Mac Schwager, uses an Unscented Kalman Filter to "guess" a driver's hidden cost function—essentially their internal priority list—simply by observing their movement. This allows the robot to adapt its trajectory at 40 Hz, negotiating space in real-time.

Beyond Position: Inferring Strategy and Foresight

Recent work has expanded the scope of what can be inferred. It isn't just about where someone is going, but how they think.

  • Foresightedness: Cade Armstrong and Ryan Park demonstrated in Inferring Short-Sightedness in Dynamic Noncooperative Games that estimating an agent's "discount factor"—how much they value future costs versus immediate ones—can reduce trajectory prediction error by over 33%.
  • Equilibrium Alignment: In complex social interactions, there are often multiple "right" ways to coordinate (multiple Nash equilibria). Inference-Based Strategy Alignment proposes that robots should use Bayesian inference to detect which "unwritten rule" humans have settled on and align their strategy accordingly.
  • Collective Plans: In sports or team environments, agents are often coordinated by a "hidden plan." baller2vec++ by Anh Totti Nguyen uses a look-ahead Transformer to model these concurrent dependencies, inferring the group's intent even when individual movements seem ambiguous.

Active Inference and Persuasion

If watching is not enough, an agent can act to reveal information. This is known as Active Uncertainty Reduction.

In IDSMPC-SHARP, Haimin Hu and Jaime Fernández Fisac describe a system that "probes" other agents. By nudging slightly into a lane, an autonomous vehicle can force a human driver to react, thereby revealing whether they are cooperative or aggressive much faster than passive observation would allow.

This moves the field toward Bayesian Persuasive Driving (/work/bayesian-persuasive-driving), where Cheng Peng and Masayoshi Tomizuka suggest that an AV can strategically signal its own intentions to "persuade" humans into safer, more efficient coordination.

Why It Works: Closing the Performance Gap

The intuition behind this research direction is that the gap between a cautious, "hesitant" robot and an efficient, "human-like" one is defined by the quality of its intent inference. As Zachary N. Sunberg argues in Efficiency and Safety in Autonomous Vehicles Through Planning With Uncertainty, modeling hidden human states allows autonomous systems to approach the performance of an "omniscient" planner. By treating the "mind" of the other agent as a variable to be solved, we transform a chaotic multi-agent environment into a structured, predictable game.

Papers in this version

Showing 60 of 175 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 203 papers
  2. Version 1 · 2026-10-04 · 175 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_latent_intent_estimation_7aea1dd5,
  title = {latent intent estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/latent-intent-estimation/rdr_7aea1dd5a23a4e119c4ff0de577ed20f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_intent_inference_for_multi_agent_coordin_7aea1dd5_v1,
  title = {Intent Inference for Multi-Agent Coordination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intent-inference-for-multi-agent-coordination/rdr_7aea1dd5a23a4e119c4ff0de577ed20f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}