• Skip to main content
  • Skip to primary navigation
  • Skip to primary sidebar
Header Search Widget
Robot Open Autonomous Racing (ROAR™)
  • ROAR Home
  • Programs
    • AI Racing Tech
    • ROAR GO
    • ROAR Academy
    • Ambassadors Program
  • News
  • Simulation Racing
    • Maps
    • Past Results
    • Blog
  • Support Us
  • Contact

Combining Outcome Driven Optimization with LiDAR Collision Avoidance for ROAR Simulation Racing

Second Place Solution in the Spring 2026 Simulation Racing Series, written by Youssef Khalafalla on behalf of Kimball Jags

Table of Contents

Abstract

Previous Kimball Jags had two main developments. The Summer 2025 solution increased target speed, reduced braking, and introduced a Collision Avoidance System based on Time till Collision. The Fall 2025 solution expanded this controller with section and lap specific variables and an outcome driven optimization loop. Together, these methods reduced completion time while keeping the crash rate substantially below the rate produced by speed optimization alone.

The original Collision Avoidance System depended on track boundaries that were calculated before the race. The left and right boundaries were generated by offsetting the centerline according to an assumed road width. This paper presents the next stage of the solution. Live LiDAR measurements replace the stored track boundaries, allowing the vehicle to estimate nearby walls directly from its surroundings.

LiDAR could not be used reliably in the earlier solution because of a faulty implementation in the version of ROARpy available during development. This limitation can be avoided by accessing the raw CARLA library, spawning a CARLA LiDAR sensor, and processing its measurements separately from the ROARpy sensor wrapper. Future development will add radar and combine LiDAR geometry with radar distance and velocity measurements.

Introduction

The Robot Open Autonomous Racing program at the University of California Berkeley provides an environment in which students can develop and evaluate autonomous racing agents. The simulation racing series uses CARLA to model the vehicle, track, and driving physics.

The Kimball Jags Summer 2025 solution focused on increasing speed without accepting the higher crash rate that normally followed. Increasing the target speed and reducing brake activation lowered the average completion time, but the resulting lateral sliding raised the crash rate to approximately 30 percent. A Collision Avoidance System reduced this rate to 7 percent by predicting when the vehicle would reach a track boundary and temporarily steering away from it. 2025 solution preserved this system while adding a more detailed optimization method. The track remained divided into ten sections. Target speed, brake reduction, and collision avoidance thresholds could be adjusted according to the current section and lap. An external optimization loop evaluated completed runs through rewards for fast and stable performance and penalties for collisions or unnecessary collision avoidance activation. solution combines these developments while changing how the vehicle detects danger. Instead of predicting collisions against a stored geometric model, the controller uses live LiDAR returns to determine where the walls actually appear relative to the vehicle.

Background

The Summer 2025 Collision Avoidance System created two track boundary lines from the supplied centerline waypoints. Each centerline point was shifted to the left and right by half of the estimated track width. A ray was then projected from the vehicle in the direction of its velocity. The first intersection between that ray and either track boundary represented the predicted collision location. ance from the vehicle exterior to the predicted intersection was divided by vehicle velocity to calculate Time till Collision. When this value fell below a selected threshold, the controller temporarily replaced normal steering with a stronger command away from the wall.

This method was effective, but it depended on several assumptions. The calculated boundary width had to represent the physical track. The centerline had to follow the center of the road accurately. Every wall had to be reconstructed before the race. The method also had limited ability to represent irregular barriers, temporary objects, or differences between the geometric model and the visible simulation environment.

Live LiDAR sensing removes many of these assumptions. The controller no longer needs to know the complete shape of each wall before driving. It only needs to identify relevant points within the area that the vehicle may occupy during the next several simulation ticks.

Objectives

The updated solution has five main objectives.

First, preserve the target speed and brake optimizations developed in the Summer 2025 solution.

Second, preserve the section and lap specific parameters developed in the Fall 2025 solution.

Third, replace the precomputed track boundary model with live LiDAR measurements.

Fourth, calculate collision risk from the shape and position of current sensor returns.

Fifth, prepare the sensing system for future fusion with radar measurements.

Accessing LiDAR Through CARLA

The earlier paper treated LiDAR as unavailable because the sensor could not be used reliably through the ROARpy implementation available during development. The underlying CARLA simulator does provide a ray cast LiDAR sensor through the blueprint named sensor.lidar.ray_cast. It produces carla.LidarMeasurement data containing point coordinates and intensity measurements. approach bypasses the faulty ROARpy LiDAR wrapper. A sensor manager accesses the raw CARLA world, finds the LiDAR blueprint, configures its attributes, and attaches the sensor actor directly to the competition vehicle. A callback receives each LiDAR measurement and stores the newest point cloud for use by the controller.

The existing ROAR racing controller can continue managing waypoints, throttle, steering, and brakes. Only the sensing path needs to access CARLA directly. This keeps the previous controller structure while replacing the source of collision distance information.

The sensor configuration can be adjusted through the number of channels, detection range, rotation frequency, field of view, and points generated per second. CARLA returns the measurements in the local sensor coordinate system, which makes them suitable for immediate analysis relative to the vehicle. Point Cloud Processing

Each raw LiDAR measurement contains points surrounding the vehicle. Not every point should affect collision avoidance. Returns from the road surface, the vehicle body, distant scenery, and objects behind the car may not represent an immediate collision risk.

The first processing stage removes points below a selected height. This reduces returns from the road surface. A small region surrounding the sensor is also removed to prevent points from the vehicle body from being interpreted as obstacles.

The remaining points are converted into a vehicle centered coordinate system. One axis represents forward distance and the other represents left and right distance. Points far behind the vehicle or outside the relevant forward range are removed.

The controller then examines a projected corridor in front of the vehicle. The corridor follows the current velocity direction rather than only the orientation of the vehicle. This distinction is important during a slide because the vehicle may be facing toward the track while still moving laterally toward a wall.

The width of the corridor includes the dimensions of the vehicle and an additional safety margin. The controller searches for the nearest LiDAR returns inside this corridor. These points represent surfaces that the vehicle may reach if its current movement continues.

Live Time till Collision

The original system calculated collision distance using an intersection with a stored boundary line. The new system obtains the distance directly from LiDAR points.

A simplified Time till Collision value can be calculated as

The value represents the distance between the vehicle exterior and the nearest relevant LiDAR return. The value represents the component of vehicle velocity directed toward that return. The small value prevents division by zero.

Using closing velocity is more accurate than using total vehicle speed. A wall located beside the vehicle should not cause an immediate response when the vehicle is moving parallel to it. The same wall should produce a much lower Time till Collision when the vehicle begins sliding toward it.

The system also evaluates whether the dangerous returns are located on the left or right side. If points on the right create the greatest collision risk, the controller applies a temporary steering adjustment toward the left. If the dangerous points are on the left, the controller steers toward the right.

As in the previous solution, steering should remain the primary response. Earlier testing found that braking during a lateral slide could increase instability and reduce the vehicle’s ability to change direction. Steering alone produced a more effective avoidance response. ptual LiDAR Track View

The following figure demonstrates how accumulated LiDAR returns could outline the track. The centerline is included only to make the example easier to understand. It would not be required for detecting the nearby walls.

Local Collision Avoidance View

The following figure shows a local point cloud expressed in the vehicle coordinate system. The projected vehicle corridor follows the velocity direction. The nearest return within that corridor supplies the distance used for Time till Collision.

Figure 1 shows a conceptual local LiDAR scan used for collision avoidance

This representation does not require the controller to decide which points belong to a mathematically perfect left or right track boundary. It only needs to determine whether a surface enters the space the vehicle is expected to occupy.

Integration with the Previous Controller

LiDAR collision avoidance does not replace waypoint following. The normal steering controller still determines the desired racing path from upcoming waypoints. LiDAR acts as a temporary safety layer that modifies the steering command only when the predicted collision risk becomes too high.

The Fall 2025 solution used different values according to section and lap because vehicle stability changed throughout the race. The LiDAR system can preserve this structure. Sections with fast turns can use earlier activation, while slower or wider sections can require stronger evidence before interrupting normal steering.

The previous optimization variables can also remain in use. The target speed coefficient controls aggressiveness. The brake multiplier controls how strongly the vehicle slows before a turn. The collision threshold controls how soon the LiDAR safety layer activates.

Additional LiDAR variables can be optimized in the same loop. These may include the safety margin around the vehicle, the minimum number of nearby returns, the maximum sensing range used for collision prediction, and the amount of temporary steering correction.

The external reward model can continue rewarding completed sections, completed laps, and improved times. It can continue penalizing crashes and unnecessary collision avoidance activation. The goal is not simply to maximize speed. The goal is to identify settings that remain fast across repeated runs without depending on a single fortunate result.

Advantages of Live Sensing

The most important advantage is that collision avoidance becomes based on observation rather than assumption. The controller responds to the walls and objects detected in the current simulation frame.

The new method also reduces dependence on a constant track width. Irregular barriers can appear naturally in the point cloud. The controller can continue operating even when the vehicle moves far from the planned centerline.

LiDAR also provides information across many directions at the same time. The previous ray casting method checked a selected geometric direction against stored boundaries. A point cloud can represent several possible collision surfaces around the vehicle and can support a wider projected corridor.

The main challenge is measurement quality. LiDAR points may contain gaps, noise, and returns that do not belong to the track boundary. The controller should therefore require a minimum level of confidence before applying a strong steering correction. When confidence is low, it can preserve normal waypoint steering and temporarily reduce the target speed.

Future LiDAR and Radar Fusion

Future development will add CARLA radar and combine its measurements with LiDAR. CARLA radar provides detections containing distance, direction, and velocity relative to the sensor. well suited for describing geometry. It can show the shape of nearby walls and provide a dense set of surface positions. Radar provides fewer geometric details, but it directly measures relative velocity. Combining the two can improve closing speed estimates and help separate stationary track boundaries from moving vehicles or other objects

Figure 2 shows a conceptual combination of LiDAR geometry and radar motion

The measurements can be transformed into the same vehicle coordinate system and associated according to direction and distance. LiDAR points would provide obstacle position and shape. Radar returns would provide an additional estimate of how quickly the distance is changing.

The fused system could calculate Time till Collision using both vehicle motion and measured radar velocity. Repeated measurements could also be combined through a tracking filter to reduce sensor noise and improve consistency between simulation ticks.

Conclusion

The proposed controller combines the strongest parts of the Kimball Jags Summer and Fall 2025 solutions. It preserves aggressive speed optimization, section and lap specific control values, steering based collision avoidance, and the external outcome driven learning loop.

The main change is the replacement of precomputed track boundaries with live LiDAR measurements. The inability to use LiDAR in the earlier system resulted from the faulty ROARpy implementation available during development rather than a lack of LiDAR support in CARLA. Direct access to the raw CARLA sensor library provides a practical workaround.

This approach allows the Collision Avoidance System to respond to the observed environment instead of relying entirely on a stored model of the track. Future radar integration will add relative velocity measurements and provide a second source of information for calculating collision risk.

Acknowledgements

The author is grateful to the UC Berkeley ROAR program for providing the environment and guidance needed to develop these autonomous racing solutions. Special thanks are given to Dr. Allen Yang and Mr. Huo Chao Kuan for their mentorship.

Primary Sidebar

Recent News

  • The 2027 ROAR Ambassadors
  • AI Racing Tech Places Third in the Indy Autonomous Challenge Laguna Seca 2026
  • Fall 2026 ROAR Simulation Racing Series Competition due Dec. 11, 2026
  • Summer 2026 Simulation Racing Series Results​
  • Contact
  • FHL Vive Center
  • Berkeley Engineering
  • UC Berkeley
  • linkedin
  • youtube
  • facebook
  • instagram
  • Privacy
  • Accessibility
  • Nondiscrimination

© 2026 UC Regents  |  Log in