💻 computer science

Mosaic: An Extensible Framework for Composing Rule-Based and Learned Motion Planners

The paper introduces Mosaic, an extensible framework that integrates rule-based and learned motion planners through arbitration graphs to achieve transparent, safe, and state-of-the-art performance in autonomous driving without requiring retraining.

Original authors: Nick Le Large, Marlon Steiner, Lingguang Wang, Willi Poh, Jan-Hendrik Pauls, Ömer Şahin Taş, Christoph Stiller

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nick Le Large, Marlon Steiner, Lingguang Wang, Willi Poh, Jan-Hendrik Pauls, Ömer Şahin Taş, Christoph Stiller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a team to drive a self-driving car. You have two very different drivers:

The "Rulebook Driver" (PDM-Closed): This driver is like a strict, by-the-book police officer. They follow every traffic law perfectly, never speed, and are incredibly predictable. However, they can be a bit rigid. If the traffic gets weird or chaotic, they might freeze because they've never seen a situation quite like that before.
The "Intuition Driver" (FlowDrive):* This driver is like a seasoned, street-smart local who learned by watching thousands of videos of real traffic. They are great at handling chaos, cutting in smoothly, and reacting to unpredictable pedestrians. But, because they rely on "gut feeling" (AI), sometimes they make risky guesses or do things that are hard to explain to a human.

The Problem:
If you only use the Rulebook Driver, the car might get stuck in traffic. If you only use the Intuition Driver, the car might take a dangerous shortcut that gets it into an accident. You want the safety of the rulebook and the adaptability of the street-smart driver.

The Solution: Mosaic
The paper introduces Mosaic, which is like a Traffic Control Tower or a Referee that sits between these two drivers.

Instead of letting one driver take the wheel, Mosaic asks both of them to propose a path forward every second. Then, it uses a special decision-making process (called an Arbitration Graph) to pick the best one.

Here is how Mosaic works, step-by-step, using simple analogies:

1. The "Safety Inspector" (Verification)

Before the car moves, Mosaic acts as a strict safety inspector.

The Analogy: Imagine both drivers propose a route. The Rulebook Driver says, "I'll go left." The Intuition Driver says, "I'll go right."
The Check: Mosaic instantly checks both routes against a safety rulebook. If the Intuition Driver's "right" route would cause a crash, Mosaic immediately rejects it (like a referee blowing a whistle).
The Benefit: This is the paper's big innovation. Usually, the AI driver checks its own safety, but it might miss something. By having a central "Safety Inspector" check both drivers, the system catches mistakes the AI might have missed. If both drivers propose a crash, then the car slams on the brakes (Emergency Stop).

2. The "Scorekeeper" (Scoring)

If both routes are safe, Mosaic acts as a scorekeeper to pick the best one.

The Analogy: Imagine a game show. The Rulebook Driver gets points for being polite and following laws. The Intuition Driver gets points for being smooth and making progress.
The Decision: Mosaic calculates a score for both. If the Intuition Driver is making great progress without breaking rules, Mosaic picks them. If the Intuition Driver gets too risky, Mosaic switches back to the Rulebook Driver.
The Result: The car gets the "best of both worlds." It drives smoothly like a human but stays safe like a robot.

3. The "Emergency Brake" (Fallback)

What if both drivers are confused and suggest a crash?

The Analogy: This is the "panic button." If the Safety Inspector rejects both proposals, Mosaic ignores the drivers entirely and triggers a pre-programmed Emergency Stop.
Why it's smart: Because the Safety Inspector catches almost all errors, the Emergency Brake is rarely needed. It's a last-resort safety net that ensures the car never does something truly dangerous.

Why is this a big deal?

The researchers tested Mosaic on a massive driving simulation called nuPlan (think of it as a giant video game with thousands of tricky traffic scenarios).

The Result: Mosaic didn't just beat the Rulebook Driver or the Intuition Driver; it beat them combined.
The Stats: It reduced accidents by 30% compared to using either driver alone. It set a new "world record" for safe driving in these simulations.
The Magic: They didn't have to retrain the AI or teach the Rulebook Driver anything new. They just built a better "manager" (Mosaic) to organize them.

The Bottom Line

Mosaic is like a mosaic artwork. Just as a mosaic takes many different, colorful tiles (some rough, some smooth) and arranges them into a beautiful, coherent picture, this framework takes a rigid rule-based planner and a flexible AI planner and arranges them into a single, safe, and explainable system.

It proves that you don't have to choose between "safe but boring" and "smart but risky." With the right manager, you can have a car that is both.

1. Problem Statement

Autonomous driving motion planning faces a fundamental trade-off between explainability/safety and adaptability:

Rule-based planners (e.g., PDM-Closed) offer predictable, transparent, and verifiable behavior but often struggle with the complexity, uncertainty, and unstructured nature of real-world traffic.
Learning-based planners (e.g., FlowDrive) demonstrate strong adaptability and generalization to diverse scenarios but suffer from "black-box" decision-making, reduced transparency, and occasional safety violations.
Current Hybrid Approaches often combine these paradigms via ad-hoc post-processing or non-unified logic, lacking a structured mechanism to arbitrate between them, verify safety centrally, or explain the final decision.

The core challenge is to create a unified framework that leverages the strengths of both paradigms while ensuring safety, traceability, and explainability without requiring retraining of the underlying models.

2. Methodology: The Mosaic Framework

The authors propose Mosaic, an extensible framework based on Arbitration Graphs (AGs). This framework decouples trajectory generation from verification and scoring, structuring the decision-making process into a hierarchical graph.

A. Architecture Components

The framework integrates two complementary planners as "behavior components" within the AG:

FlowDrive*: A learning-based planner using rectified flow-matching (diffusion transformer) with moderated guidance for maneuver diversity.
PDM-Closed: A rule-based planner extending the Intelligent Driver Model (IDM) with model predictive control, generating multiple trajectory proposals based on lane selection and target speeds.
Emergency Stop: A deterministic fallback behavior (straight-line deceleration) acting as the ultimate safety layer.

B. The Arbitration Mechanism

The decision flow involves three key stages:

Centralized Verification (Safety Layer):
- Unlike individual planners that check safety internally, Mosaic moves verification to the Composer (a cost arbitrator).
- Before a trajectory is scored, it undergoes a domain-specific verification check (e.g., predicting at-fault collisions within 2.0s).
- Redundancy: If one planner proposes an unsafe trajectory, it is rejected, but the other planner's valid proposal is still considered. Emergency braking is only triggered if both planners fail verification simultaneously.
Unified Scoring (Cost Arbitrator):
- Valid trajectories are scored using a continuous function inspired by the nuPlan metrics, replacing binary pass/fail with a score in $[0, 1]$ .
- The total score ( $S_{total}$ $S_{t o t a l}$ ) is a multiplicative combination of:
  - Safety Gate: Collision score ( $S_{coll}$ ), Drivable area ( $S_{driv}$ ), and Driving direction ( $S_{dir}$ ). If any safety metric is zero, the total score collapses to zero.
  - Progress Gate: Penalizes standstill trajectories.
  - Performance Score: Weighted average of progress, Time-to-Collision (TTC), and comfort.
Priority Arbitration (Root Node):
- The root arbitrator (Mosaic) selects the final trajectory.
- It prioritizes the best valid trajectory from the Composer.
- If the Composer yields no valid trajectory (both planners failed), it falls back to the Emergency Stop behavior.

3. Key Contributions

Extensible AG Framework: Introduces Mosaic, a structured decision-making mechanism that can integrate any planning algorithm (rule-based, learned, or hybrid) without retraining.
Decoupled Verification and Scoring: Separates trajectory generation from safety checks and scoring. This makes every rejection decision transparent and traceable, allowing the system to explain why a specific planner was rejected.
Safety Redundancy: By centralizing verification, the framework reduces collisions significantly. It leverages the complementary strengths of planners; if one fails, the other often succeeds, minimizing the need for emergency braking.
State-of-the-Art Performance: Achieves new benchmarks on nuPlan and interPlan without additional data or model retraining.

4. Experimental Results

The framework was evaluated on the nuPlan (Val14 split) and interPlan benchmarks.

nuPlan Closed-Loop Benchmarks (Val14):
- CLS-NR (Non-Reactive): 95.48 (New SOTA).
- CLS-R (Reactive): 93.98 (New SOTA).
- Collision Reduction: Reduced at-fault collisions by 30% compared to either planner in isolation (14 collisions vs. 18 for FlowDrive* and 17 for PDM-Closed).
- Zero-Score Reduction: Reduced scenarios with a total score of zero by 33%.
interPlan Benchmark (Highly Interactive Scenarios):
- CLS-R Score: 54.30.
- Improvement: Outperformed the best constituent planner (FlowDrive*) by 23.3% and PDM-Closed by 31.7%.
Ablation Studies:
- Removing the verification module doubled collisions (from 14 to 35), proving its critical role in safety.
- The cost arbitrator selected FlowDrive* 18.7% of the time, PDM-Closed 28.8%, and chose based on a tie-breaker (favoring the learned planner) 52.5% of the time, demonstrating effective utilization of complementary strengths.
- Emergency braking was triggered in only 0.08% of time steps, confirming that the redundancy between planners effectively prevents unnecessary interventions.
Computational Overhead:
- The full Mosaic framework adds only ~13% wall-clock time compared to FlowDrive* alone.
- Since FlowDrive* runs on GPU and PDM-Closed/Arbitration logic runs on CPU, they can be executed in parallel, potentially reducing latency to near zero.

5. Significance

The Mosaic framework represents a significant step forward in autonomous driving by solving the "safety vs. adaptability" dilemma through structural integration rather than ad-hoc hybridization.

Explainability: It provides a clear audit trail for every decision, crucial for regulatory approval and public trust.
Scalability: The framework is planner-agnostic, meaning it can easily incorporate future advances in either rule-based or learning-based planning without architectural changes.
Safety Assurance: By treating safety verification as a distinct, centralized layer, it ensures that unsafe proposals are filtered out before selection, creating a robust safety net that outperforms isolated planners.

In summary, Mosaic demonstrates that combining heterogeneous planners via arbitration graphs yields a system that is safer, more robust, and more explainable than the sum of its parts, setting a new standard for motion planning in complex traffic environments.