⚡ electrical engineering

Robust Global-Local Behavior Arbitration via Continuous Command Fusion Under LiDAR Errors

This paper introduces a ROS2-native arbitration module that employs a PPO-trained policy to continuously fuse global Pure Pursuit and local LiDAR-based Gap Follow commands, demonstrating enhanced robustness against sensing impairments like noise, delay, and dropout in close-proximity driving scenarios compared to a sampling-based baseline.

Original authors: Mohamed Elgouhary, Amr S. El-Wakeel

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohamed Elgouhary, Amr S. El-Wakeel

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 driving a car that has two distinct "brains" working at the same time, but they have very different personalities.

The Two Brains:

The Navigator (Global Controller): This brain is like a strict GPS. It knows the perfect route, the fastest way to the finish line, and it wants to stay exactly on the painted line. It says, "Keep going straight, stay in the lane, don't deviate!"
The Guard Dog (Local Controller): This brain is hyper-alert and reactive. It only cares about what's right in front of the car's nose. If it sees a rock, a pothole, or another car, it screams, "STOP! SWERVE! AVOID!" It doesn't care about the destination; it only cares about not crashing right now.

The Problem:
In the real world, the car's sensors (specifically the LiDAR, which is like a high-tech laser eye) aren't perfect. Sometimes the laser gets confused by dust, rain, or a glitch. It might think a tree is right in front of the car when it's actually ten feet away, or it might miss a car entirely for a split second.

If you just let the "Guard Dog" take over whenever the sensors get a little noisy, the car might swerve wildly and never reach its destination. If you only listen to the "Navigator," the car might drive straight into a wall because the GPS didn't see the obstacle.

The Solution: The "Traffic Cop" (The Arbitration Module)
This paper introduces a new, smart "Traffic Cop" that sits between the Navigator and the Guard Dog. Instead of letting one brain completely take over the other, this Traffic Cop blends their commands together smoothly.

Here is how it works, using a simple analogy:

The "Volume Knob" Analogy

Imagine the Navigator and the Guard Dog are two musicians playing different songs.

The Navigator is playing a smooth jazz track (staying on the path).
The Guard Dog is playing a frantic drum solo (avoiding obstacles).

Usually, you only hear one or the other. But this new system uses a volume knob (called $\alpha$ ) that the Traffic Cop controls.

If the road is clear, the Traffic Cop turns the Navigator's volume up to 100% and the Guard Dog's down to 0%. The car cruises smoothly.
If a car suddenly appears, the Traffic Cop instantly turns the Guard Dog's volume up and the Navigator's down. The car swerves safely.
The Magic: The Traffic Cop doesn't just flip a switch. It turns the knob continuously. It might set the Navigator to 70% and the Guard Dog to 30%, creating a smooth, blended steering command that feels natural, not jerky.

The "Smart Filter" (Dealing with Bad Sensors)

The real genius of this paper is how the Traffic Cop handles bad sensor data.

In the past, if the sensors glitched (like a "hallucination" where the car thinks a ghost is in front of it), a simple rule-based system might panic and slam on the brakes.

The Old Way: "I see a ghost! STOP!" (Even if it's a glitch).
The New Way (This Paper): The Traffic Cop is trained using a special learning method (Reinforcement Learning). It has seen thousands of scenarios where the sensors lie. It learns to look at the big picture.
- It asks: "Is the Navigator still confident? Is the Guard Dog overreacting? Is this a glitch?"
- If the sensors are noisy, the Traffic Cop trusts the Navigator a bit more and ignores the frantic screaming of the Guard Dog, preventing the car from panicking over a false alarm.

The "Safety Net"

Even with a smart Traffic Cop, the authors added a "Safety Net." If the sensors are so broken that the car can't see anything, or if the data is too old (stale), the Safety Net immediately hits the brakes. It's like a parent grabbing the steering wheel if the child driver is about to make a fatal mistake.

The Results: Why It Matters

The researchers tested this in a simulation with a race car (F1TENTH) against a slower opponent. They intentionally broke the sensors by adding noise, delays, and fake "ghost" obstacles.

The Competitor: They also built a "math-heavy" system that tries to calculate every possible future path. This system was fast on average but would occasionally freeze or panic when the sensors got messy, causing it to miss the race or get too close to the opponent.
The New System: The "Traffic Cop" system was more consistent. Even when the sensors were lying to it, it kept the car moving forward safely. It didn't panic as easily, and it finished the race more often.

In Summary

This paper presents a new way to drive autonomous cars that is robust (doesn't break easily) and smooth. Instead of choosing between "follow the map" or "avoid the crash," it learns to blend the two perfectly, even when the car's eyes are blurry. It's like having a driver who is calm enough to ignore a false alarm but sharp enough to react instantly when danger is real.

1. Problem Statement

Modular autonomous driving systems face a critical challenge: coordinating global progress objectives (e.g., following a planned trajectory) with local safety reactions (e.g., avoiding obstacles) when sensor data is imperfect.

The Conflict: Global controllers (like Pure Pursuit) ensure path tracking but may ignore immediate obstacles. Reactive local controllers (like Gap Follow) ensure safety but may lack long-horizon progress goals.
The Sensing Challenge: Real-world LiDAR data is often corrupted by noise, latency, scan dropouts, and false short-range outliers. Traditional rule-based arbitration (switching between controllers based on thresholds) is highly sensitive to these artifacts, leading to premature maneuvers, oscillations, or overly conservative behavior that stalls progress.
The Goal: Develop a robust arbitration layer that fuses these two distinct behaviors into a single command, maintaining safety and progress even under significant sensing degradation, without replacing the underlying interpretable controllers.

2. Methodology

The authors propose a ROS2-native arbitration module that uses Reinforcement Learning (RL) to learn a continuous fusion policy while keeping the low-level controllers unchanged.

A. System Architecture

The system operates at 30 Hz and integrates two parallel low-level controllers:

Global Controller: Pure Pursuit (PP) for reference tracking.
Local Controller: Gap Follow (GF) for reactive collision avoidance based on 2D LiDAR.

Both controllers generate candidate Ackermann commands (steering $\delta$ and velocity $v$ ).

B. Continuous Command Fusion

Instead of switching between controllers, a learned policy outputs a continuous gate $\alpha \in [0, 1]$ . The final command $u_t$ is a linear blend:
$u_t = (1 - \alpha_t) u_{PP,t} + \alpha_t u_{GF,t}$
This approach preserves the stability and interpretability of the base controllers while allowing smooth transitions.

C. Learning-Based Arbitration (PPO)

Algorithm: Proximal Policy Optimization (PPO) with Generalized Advantage Estimation (GAE).
Observation Space: A compact feature vector ( $s_t$ $s_{t}$ ) rather than raw LiDAR scans. Features include:
- Forward-cone LiDAR statistics (e.g., minimum clearance).
- Reference geometry (curvature, lookahead).
- Relative opponent information (distance, angle, speed).
- Robustness Feature: During training, opponent features are randomly masked to simulate scenarios where opponent pose is unobservable.
Action Space: A scalar $\alpha$ mapped via a sigmoid function to ensure it stays within $[0, 1]$ .
Reward Design: Encourages forward progress, penalizes unsafe proximity (clearance violations), and discourages rapid switching (jitter) in the gate value.

D. Safety Mechanisms

To ensure safety without modifying the RL policy, the system includes:

Interaction-Mode Gating: The policy only activates during specific interaction scenarios (e.g., head-to-head overtaking).
Runtime Smoothing: Low-pass filtering of the gate $\alpha$ to prevent high-frequency oscillations.
Safety Override: A hard stop is triggered if forward clearance drops below a minimum threshold or if input data is stale (timeout).

E. Baseline Comparison

To validate the approach, the authors implemented a Sampling-Based Predictive Baseline (Sampling MPC).

It uses the same ROS inputs and control rate (30 Hz).
It evaluates a discrete set of constant-steering candidates using a kinematic bicycle model.
It screens candidates against LiDAR data for collisions and selects the minimum-cost trajectory.

3. Key Contributions

ROS2-Based Robustness Evaluation Protocol:
- Introduced a systematic impairment injection protocol for LiDAR data, including Gaussian noise, fixed delays (200 ms), scan dropouts (30%), and a sweep of false short-range outliers in the forward cone.
- Generated robustness curves quantifying success rates and runtime stability under increasing sensing stress.
Command-Level Continuous Fusion:
- Presented a lightweight arbitration module that fuses Pure Pursuit and Gap Follow via a single learned scalar gate.
- Demonstrated that learning a single decision variable is sufficient to coordinate complex behaviors while maintaining real-time constraints.
Matched Baseline Comparison:
- Provided a fair comparison against a model-based sampling MPC using identical inputs, control frequencies, and simulation environments, isolating the benefits of the fusion approach.

4. Experimental Results

Experiments were conducted in the F1TENTH Gym ROS simulator on the Levine track (training) and the unseen IMS track (evaluation).

Nominal Performance (Clean Sensing):
- Proposed Arbiter: Achieved 100% success rate (10/10 heats) with 0 collisions and 0 timeouts.
- Sampling MPC: Achieved 90% success rate, with one heat ending in a timeout.
- Runtime: The Arbiter had a slightly higher mean runtime (31.95 ms vs. 24.15 ms) but significantly lower worst-case runtime spikes (47.12 ms vs. 61.83 ms), indicating more consistent real-time stability.
Robustness Under LiDAR Impairments:
- As the probability of false short-range outliers ( $p_{out}$ $p_{o u t}$ ) increased from 0.0 to 0.4:
  - The Sampling MPC performance degraded sharply. It frequently rejected valid trajectories as "unsafe" due to false obstacles, leading to timeouts and unsafe proximity events.
  - The Proposed Arbiter maintained a higher safe success rate and degraded more gradually.
- Reasoning: The Arbiter relies on aggregated summary features (e.g., minimum clearance statistics) rather than checking every beam against every candidate trajectory. This makes it less sensitive to isolated outliers and prevents the "conservative paralysis" seen in the sampling baseline.

5. Significance and Conclusion

Practicality: The method offers a practical solution for modular autonomy stacks (like Autoware) where replacing existing, verified low-level controllers is often undesirable. It acts as a "plug-in" coordinator.
Robustness: The study demonstrates that learning a fusion policy is more robust to sensor noise and false positives than traditional sampling-based predictive control in interactive scenarios.
Real-Time Stability: By avoiding complex online optimization loops for every candidate, the fusion approach provides more predictable execution times, which is critical for safety-critical control loops.
Scope: The authors clarify that this is a command-level evaluation, not a replacement for high-level planning. It specifically addresses the coordination of existing behaviors under imperfect sensing, filling a gap between high-level planning and low-level control.

In summary, the paper validates that a continuous, learning-based fusion of interpretable controllers outperforms a matched sampling-based predictive baseline in terms of safety, success rate, and real-time stability when operating under realistic LiDAR sensor impairments.