Interactive Motion Planning for Autonomous Vehicles

This direction is split into other directions: decoupled prediction and planning. Its last version stays readable and citable here.

Version 1 of 160 papers61 conceptsfrozen 2026-10-04

In traditional autonomous driving architectures, the system follows a linear pipeline: perceive the environment, predict where others will go, and then plan a safe path. This "predict-then-plan" paradigm treats surrounding agents—human drivers, cyclists, and pedestrians—as moving obstacles with fixed future trajectories.

However, researchers like Jaime F. Fisac and Wenchao Ding have identified a fundamental flaw in this approach: it ignores mutual influence. In the real world, driving is a social negotiation. When you edge into a busy lane, the car behind you slows down to let you in. If an autonomous vehicle (AV) fails to account for how its own actions change the behavior of others, it falls victim to the "Frozen Robot Problem"—a state of paralysis where the vehicle remains stuck because it assumes no one will ever yield.

The Feedback Loop of Interaction

The core shift in recent research is moving from a decoupled, sequential architecture to an integrated, closed-loop framework. Instead of asking "Where will that car go?", the system asks "Where will that car go if I perform this specific maneuver?"

As Christoph Stiller and Maximilian Naumann argue, the trajectories of all agents in a scene should be treated as a single "ensemble." By optimizing a shared cost functional, the AV can anticipate how its presence creates safety margins or induces cooperation in others. This allows for fluid maneuvers in tight spots, such as narrowings or T-junctions, where a "selfish" planner would simply stall.

Game-Theoretic Reasoning

To model this "I think that you think" logic, many researchers have turned to game theory. Frameworks like GameFormer, developed by Zhiyu Huang and Chen Lv, use hierarchical game theory to iteratively refine trajectories.

The intuition behind these models is often Iterative Best Response:

  1. The AV proposes a plan.
  2. The system predicts how others would respond to that specific plan.
  3. The AV adjusts its plan based on those responses.
  4. The process repeats until a "negotiated" equilibrium is reached.

Aron Distelzweig and Joschka Boedecker take this further with BIBeR (Bayesian Iterative Best Response), which treats driving as a game of incomplete information. By maintaining different "types" or driving styles for other agents (e.g., aggressive vs. conservative), the AV can safely probe the environment to see how others react, effectively "guessing" their hidden objectives in real-time Mac Schwager.

Joint Neural Architectures

While game theory provides the mathematical rigor, Transformers and Diffusion models provide the predictive power to handle complex, high-dimensional scenes.

The Holistic Transformer by Hongyu Hu breaks the silo between prediction and decision-making by performing both tasks simultaneously within a single neural network. Similarly, Zhili Chen and Maosheng Ye introduced PPAD, which interleaves prediction and planning steps. This ensures that the ego-vehicle's intended path is "baked into" the attention mechanisms that forecast other agents' movements.

A more recent frontier involves World Models, which allow an AV to "imagine" the future evolution of a 3D scene. Wenzhao Zheng (OccWorld) and Tianze Xia (DriveLaW) have shown that by unifying video generation and motion planning, the vehicle can ensure its actions are physically and spatially consistent with the predicted environment. If the model "imagines" a car slowing down in response to a merge, the planner can confidently execute that merge.

Why Integration Works

The reason these integrated approaches outperform traditional ones isn't just about better accuracy—it's about information flow.

  • In a decoupled system, the planner is blind to the reasons behind a prediction. It sees a car moving forward and assumes it will continue, even if the AV's own path is about to cut it off.
  • In an integrated system, the prediction is conditional. As Fabian Konstantinidis demonstrates, using closed-loop simulations to forecast reactions allows the AV to recognize opportunities for cooperation.

By treating the environment as a reactive participant rather than a static backdrop, these frameworks enable autonomous vehicles to move beyond cautious obstacle avoidance and toward true social intelligence on the road.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_interactive_motion_planning_for_autonomo_c5d2ca3b,
  title = {Interactive Motion Planning for Autonomous Vehicles},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interactive-motion-planning-for-autonomous-vehicles/rdr_c5d2ca3bb1904b8cb0820a88da1b33f1},
  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_interactive_motion_planning_for_autonomo_c5d2ca3b_v1,
  title = {Interactive Motion Planning for Autonomous Vehicles},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interactive-motion-planning-for-autonomous-vehicles/rdr_c5d2ca3bb1904b8cb0820a88da1b33f1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}