Proactive Autonomous Planning via Leader Follower Games

23 concepts22 papers1 proposal

When autonomous vehicles (AVs) and robots operate alongside humans, they frequently suffer from a form of behavioral paralysis. Traditional navigation pipelines treat surrounding human agents as dynamic obstacles—predicting their future trajectories in a isolated "black box" module and then planning a collision-free path around them. As Sarkar notes, this one-size-fits-all assumption fails to capture the interactive nature of shared spaces. It often results in robots that are "too polite," freezing at busy intersections or getting trapped because they assume humans will maintain their current paths regardless of what the robot does (Sastry et al.).

To break this paralysis, a growing body of research shifts the paradigm from passive prediction to game-theoretic interaction modeling. By framing multi-agent coordination as a hierarchical, leader-follower (Stackelberg) game, autonomous systems can explicitly reason about how their immediate actions will alter the future decisions of others.

The Geometry of Influence: Why Stackelberg Games Work

In a standard simultaneous-move game (like a classic Nash equilibrium framework), agents choose their strategies at the same time, assuming the other's strategy is fixed. While mathematically elegant, this fails to capture the sequential, reactive nature of physical interaction. It also introduces a computational nightmare: infinite recursive reasoning ("I think that you think that I think...").

The leader-follower framework bypasses this recursion by imposing a strict optimization hierarchy. The autonomous agent designates itself as the leader and models the human or surrounding agent as the follower.

Mathematically, this is structured as a bi-level optimization problem:

$$\min_{u_{\text{leader}}} J_{\text{leader}}(x, u_{\text{leader}}, u_{\text{follower}})$$

$$\text{subject to } u_{\text{follower}} = \arg\min_{u} J_{\text{follower}}(x, u_{\text{leader}}, u)$$

Here, the leader optimizes its own objective function ($J_{\text{leader}}$) subject to the constraint that the follower will always respond rationally by minimizing their own cost ($J_{\text{follower}}$). Because the leader embeds the follower's reaction function directly into its own trajectory planning, it can choose actions that intentionally guide the follower toward a mutually beneficial state.

This structural simplification yields massive practical benefits. Hang et al. demonstrated that modeling interactions as a leader-follower game rather than a simultaneous-move game reduces decision-making costs by over 20%. Furthermore, because the leader anticipates reactions rather than reacting abruptly to surprises, this formulation naturally produces smoother velocity profiles during high-conflict scenarios like unsignalized intersections (Hang et al.).

Nudging and Implicit Communication

When a robot understands its influence over a follower, it can use its physical trajectory as a form of implicit communication. Instead of waiting for a gap in traffic, an AV can proactively "nudge" a human driver—for instance, by slightly asserting its position into a lane to signal an intent to merge, safely prompting the follower to decelerate (Sastry et al.; Vazquez Espinoza et al.).

[Traditional Planning]
Human Trajectory Predicted ──> Robot Plans Around It ──> Passive/Frozen Behavior

[Leader-Follower Planning]
Robot Trajectory Planned ──> Anticipates Human Response ──> Proactive "Nudge" (Implicit Communication)

This proactive signaling has been shown to eliminate the "awkward dances" often seen when robots and humans cross paths in tight corridors. By combining physical movement with explicit signals like haptic pulses, frameworks designed by Okamura et al. allow robots to establish clear leader-follower dynamics, significantly increasing human trust and navigation efficiency.

This mechanism extends beyond autonomous driving into collaborative robotics. For example, Zhu et al. developed a framework where a more capable "leader" robot strategically guides a less capable "follower" robot during complex manipulation tasks, vastly outperforming greedy, non-game-theoretic coordination strategies.

Modeling the Follower's Internal State

The primary engineering challenge of the leader-follower framework is capturing the follower's objective function ($J_{\text{follower}}$). Humans are boundedly rational, diverse, and change their behavior over time. Researchers employ several distinct strategies to approximate this internal state:

  • Inverse Reinforcement Learning (IRL): By observing human data, the leader learns a reward function that mimics typical human responses, treating them as rational utility-maximizers (Okamura et al.; Sun et al.).
  • Data-Driven Linearization: Zhao et al. leverage Koopman operator theory to lift nonlinear follower dynamics into a higher-dimensional linear space. This allows the leader to compute optimal guidance strategies rapidly without knowing the follower's exact internal decision model.
  • Meta-Learning: To handle variations in human driving styles (e.g., aggressive vs. conservative), Zhu proposes a Stackelberg meta-learning approach that allows the leader to quickly adapt its guidance strategy to diverse, unknown follower behaviors.
  • Strategic Unpredictability: In long-term interactions, humans can learn to game the robot's predictable strategies. To counter this, Sagheb et al. demonstrate that leaders can maintain long-term influence by optimizing for "belief entropy," intentionally keeping their exact coordination strategies slightly obscured.

Defining and Inferring Leadership

Who gets to be the leader? Leadership in these systems is not merely a spatial heuristic (i.e., the vehicle in front is the leader). Instead, Fridovich-Keil et al. demonstrate that leadership is a structural property of optimization. It is defined by whose strategy directly enters into the best-response function of the other. By running real-time filters, autonomous systems can infer who is currently driving the interaction's objective function and dynamically adjust their roles.

Ultimately, by structuring multi-agent interactions as hierarchical games, autonomous agents transition from passive observers to active coordinators. Whether navigating an intersection (Rus et al.), executing a robust lane change (Li), or managing macro-level traffic networks via targeted tolling (Lazar et al.), the leader-follower framework provides a mathematically rigorous foundation for proactive, human-centric autonomy.

Key Foundational Work

Adaptive Game-Theoretic Decision Making for Autonomous Vehicle Control at Roundabouts (2018) — contributed 2 concept elements to this direction.
Planning and Decision-Making for Autonomous Vehicles (2018) — contributed 1 concept elements to this direction.
Efficient and Trustworthy Social Navigation via Explicit and Implicit Robot-Human Communication (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_leader_follower_interaction_modeling_4fa2e888,
  title = {leader-follower interaction modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/leader-follower-interaction-modeling/rdr_4fa2e888ebc346e6b5eb1b900621ac57},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_leader_follower_interaction_modeling_4fa2e888_v1,
  title = {leader-follower interaction modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/leader-follower-interaction-modeling/rdr_4fa2e888ebc346e6b5eb1b900621ac57/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake