Trajectory Planning and Control in Mixed Autonomy Traffic

749 concepts490 papers9 proposals

The transition to fully autonomous transportation will not happen overnight. Instead, we are entering a prolonged era of mixed autonomy, where Connected and Automated Vehicles (CAVs) must share the road with human-driven vehicles (HDVs).

This environment introduces a fundamental informational and behavioral asymmetry. CAVs can communicate digitally via vehicle-to-vehicle (V2V) links, execute trajectories with millisecond precision, and follow deterministic control laws. Human drivers, by contrast, suffer from delayed reaction times, do not broadcast their acceleration profiles, and exhibit stochastic, non-cooperative behaviors driven by distraction, fatigue, or aggression.

How can an autonomous system plan safe, efficient trajectories when its closest neighbors are highly unpredictable? Rather than trying to solve the impossible puzzle of perfect human psychological prediction, recent research focuses on control frameworks that are mathematically robust to human variance, game-theoretic models that anticipate human adaptation, and cooperative multi-robot strategies that shield the system from human error.


Bounded Disturbances and Safety Invariants

When an autonomous vehicle executes a safety-critical maneuver, such as merging onto a highway or changing lanes, relying on a single, idealized prediction of human behavior is dangerous. If a human driver accelerates unexpectedly, a rigid trajectory plan can fail catastrophically.

To guarantee safety without being overly conservative, researchers often turn to Control Barrier Functions (CBFs). Instead of predicting exactly what a human will do, this approach treats the unshared intent of human drivers as a bounded disturbance within a dynamic system.

Chang Liu demonstrates this by combining CBFs with Control Lyapunov Functions (CLFs) to manage highway on-ramp merging. The intuition behind a CBF is simple yet powerful: it defines a forward-invariant safe set in the vehicle's state space (the positions and velocities where a collision is physically impossible).

   [ Safe Invariant Set (CBF Guardrail) ]
  ┌──────────────────────────────────────┐
  │  ▲                                   │
  │  │  AV Trajectory Plan (CLF)         │
  │  └───────────────►                   │
  │                   ▲                  │
  │                   │ [Human Sudden    │
  │                     Acceleration]    │
  └──────────────────────────────────────┘

The controller constantly optimizes for traffic efficiency using the CLF. However, if a human driver acts erratically, the state approaches the boundary of the safe set. The CBF acts as a mathematical guardrail, overriding the efficiency objectives and forcing the CAV to brake or accelerate to remain within the safe invariant zone.

On a broader scale, Jie Li et al. apply a similar philosophy using nonlinear $H_\infty$ robust control. By framing human behavioral deviations as worst-case disturbances, they use Sum of Squares (SOS) programming to design controllers that guarantee traffic stabilization even when real-world human braking patterns diverge completely from standard car-following models.


Branching Across Multi-Modal Intentions

Human behavior is not just noisy; it is multi-modal. When a CAV approaches an intersection, an oncoming human driver might yield, or they might aggressively accelerate to beat the light. Averaging these two distinct intentions results in a single "ghost" trajectory down the middle—a path the human will never actually take—which can cause the autonomous system to make dangerous or confusing decisions.

To handle this multi-modality, Yuxiao Chen et al. developed a framework called Branch Model Predictive Control (Branch MPC).

Instead of planning for a single outcome, the autonomous vehicle constructs a scenario tree representing the discrete possible intents of the human driver. The CAV then optimizes a corresponding trajectory tree:

                      ┌──► [Branch A: Human Yields]  ──► CAV Proceeds Safely
                      │
[Root: Current Action]┤
                      │
                      └──► [Branch B: Human Accelerates] ──► CAV Yields/Brakes

The key to why Branch MPC works lies at the root of the tree. The autonomous vehicle must choose a single, immediate control action right now that is safe for both branches. As time progresses and the human's true intent becomes clear through their real-world acceleration, the CAV seamlessly transitions down the corresponding branch of its pre-planned trajectory tree. This prevents the vehicle from freezing due to worst-case assumptions, allowing it to remain assertive yet prepared for any sudden behavioral shifts.


Exploiting Robot Cooperation to Buffer Human Error

If predicting human behavior is a bottleneck, an alternative strategy is to use the predictability of other autonomous vehicles to isolate and absorb human uncertainty.

Anni Li et al. demonstrate that during complex lane-changing maneuvers, a CAV does not need to solve the "human behavior puzzle" if it can cooperate with another digital ally. By leveraging V2V communication with a downstream or adjacent CAV, the lane-changing vehicle can coordinate a synchronized gap-creation maneuver. Because the two robotic vehicles move in perfect, predictable alignment, they create a protective buffer that mathematically insulates the lane-changing maneuver from whatever an uncoordinated human driver in the adjacent lane might do.

This concept of using a sparse network of robots to control a chaotic human environment extends to macro-level traffic flow. Human drivers naturally create "stop-and-go" traffic waves due to delayed braking reactions, which wastes energy and causes congestion.

Jonathan Lee et al. proved in a massive open-road field experiment that a small penetration rate of CAVs (as low as 4-5%) can act as Lagrangian actuators. By utilizing hierarchical control architectures, these sparse automated vehicles intentionally maintain large, elastic gaps in front of them. When a human-induced traffic wave hits, the CAVs slowly compress their gaps rather than slamming on the brakes, dampening the shockwave and smoothing the traffic flow for miles behind them.


Game-Theoretic Alignment and Human Adaptation

Early models of mixed autonomy treated human drivers like static obstacles or predictable physics particles. However, humans are intelligent, adaptive agents who actively change their behavior based on how the robots around them act.

Yiyun Wang analyzed real-world trajectory data from the Waymo Open Dataset and discovered a fascinating feedback loop: because autonomous vehicles are famously cautious, human drivers actively adapt by crowding them at red lights and cutting them off more aggressively. If an autonomous vehicle drives with rigid, hyper-conservative rules, it inadvertently provokes more dangerous, aggressive behaviors from the surrounding humans.

To resolve this interactive loop, Peng Hang et al. utilize noncooperative game theory—specifically Stackelberg games—to model the interaction.

In this formulation, the CAV acts as the "leader" and the human driver acts as the "follower." The CAV plans its trajectory by explicitly mathematicalizing how the human will respond to its movements. By optimizing its path through the human's expected reaction curve, the CAV can assertively claim its right-of-way or yield smoothly, reducing decision-making costs and preventing the awkward, jerky interactions that characterize naive automation.

Ultimately, safety in mixed autonomy requires that the robot's intent is clear to the human eye. Martin Lauer argues that trajectory generation algorithms must prioritize comprehensibility. When a CAV's path is mathematically optimized to match familiar human driving profiles, surrounding motorists can accurately predict the robot's next move. This mutual predictability closes the loop, transforming a chaotic mixed-autonomy environment into a stable, cooperative system.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 498 papers
  2. Version 1 · 2026-10-04 · 550 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_mixed_autonomy_coordination_2ba6a29c,
  title = {mixed-autonomy coordination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-autonomy-coordination/rdr_2ba6a29c97144c148bc5895d0fceb419},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_mixed_autonomy_coordination_2ba6a29c_v2,
  title = {mixed-autonomy coordination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-autonomy-coordination/rdr_2ba6a29c97144c148bc5895d0fceb419/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake