AWS DeepRacer Car Modifications

The AWS DeepRacer program has been one of the most exciting and engaging ways to introduce customers to machine learning in action. Whether at executive briefings, internal enablement sessions, or full-scale public races, few activities capture the collective energy of a room like a DeepRacer event. Since its launch in 2018, AWS DeepRacer has helped thousands of customers learn about reinforcement learning in an interactive and competitive format.

When a DeepRacer event is well-planned and executed, it becomes a rallying point—customers line up to race, ask questions, and dive deeper into how machine learning can apply to their business challenges. DeepRacer not only demystifies ML but also sparks broader conversations around AWS services and innovation.

Over the years and countless races, DeepRacer facilitators and pit crews have developed a deep understanding of what makes these events successful—and where friction can emerge. While the AWS platform and supporting software have evolved, the physical DeepRacer cars have remained largely unchanged since launch. As facilitators, we've identified a series of enhancement opportunities that could dramatically improve the customer experience during physical DeepRacer races.

This blog outlines those recommended enhancements—rooted in practical field experience—to help raise the bar on reliability, ease of use, and overall event impact.

Impact: Leads to confusion during race starts as well as troubleshooting if it is a bad model or a car issue which can interrupts the flow of a race.

One of the most common frustrations experienced by AWS DeepRacer pit crews and racers alike is a deceptively simple issue: the car won't drive straight. Even with a well-trained model, some DeepRacer cars zig-zag unpredictably down the straightaways. The root cause? Misaligned front wheels due to non-adjustable tie rods on the stock chassis.

The AWS DeepRacer uses a WLToys A979 4WD RC chassis, which incorporates a classic Ackermann steering geometry—a configuration widely used in both real and remote-controlled cars. Ackermann steering is designed so that the inside wheel turns more sharply than the outside wheel during a turn, compensating for the tighter turning radius of the inner path. This geometry is achieved by mounting the steering tie rods behind the pivot point of each front wheel on the “bellcrank” assembly.

However, in the DeepRacer’s stock configuration, both front wheels are set to a fixed "toe-out" alignment—angled outward from the centerline. This design exaggerates the effects of Ackermann geometry. It also introduces a serious limitation: the wheels cannot be adjusted, and as a result, the front wheels are never aligned with the rear wheels.

The current AWS SOP for steering calibration instructs the user to align one front wheel and the rear wheel flush with the edge of a table—considering this position "centered." But this results in the opposite front wheel pointing outward, often by 30 degrees. The DeepRacer’s onboard AI must then constantly overcorrect for this baked-in mechanical misalignment. The result is a car that constantly drifts, oversteers, and appears to "wander" down straight paths—even with precise models and calibration. This issue is especially noticeable during customer-facing races, where model performance is expected to be the primary differentiator—not mechanical defects.

Above is a stock DeepRacer car. The front wheels do not have adjustable tie rods. The closest to a “straight wheel” configuration is each wheel is turned outward 10+ degrees from centerline. This stock configuration exaggerates the Ackermann steering but negatively impacts the car when driving in a straight line.

The solution lies in replacing the fixed plastic tie rods with aftermarket adjustable tie rods, commonly available for the WLToys A979 chassis. These allow both front wheels to be properly aligned with the rear wheels, achieving zero toe-out when the wheels are centered.

 Above is the chassis with the adjustable tie rods installed and the front wheels aligned with the rear.Above is the chassis with the adjustable tie rods installed and the front wheels aligned with the rear.Above is the chassis with the adjustable tie rods installed and the front wheels aligned with the rear.

Above is the chassis with the adjustable tie rods installed and the front wheels aligned with the rear.

The calibration utility in the DeepRacer console / car GUI instructs users to use the table edge as a guide—but this method, when applied to a car with a stock "toe-out" configuration, teaches the system that a misaligned wheel is “center.” This has a cascading negative effect on how the reinforcement learning model controls the car.

Calibration with modified tie rods configuration:
 

The AWS DeepRacer utilizes the WLToys A979 chassis, renowned for its affordability and adaptability. However, this chassis wasn't originally designed to support the additional weight introduced by the DeepRacer's compute module, camera, and battery packs. This added weight can overwhelm the stock suspension system, leading to excessive body roll during high-speed maneuvers.

In its factory configuration, the A979's suspension employs soft coil-over shocks suitable for lightweight RC applications. When running with the additional weight of the DeepRacer additional components, the suspension's softness becomes a liability. During sharp turns or rapid directional changes, the vehicle experiences significant body roll, compromising stability and increasing the risk of veering off-track.

A practical and cost-effective remedy involves adding preload spacers to the existing shock absorbers. These spacers compress the springs slightly, increasing their stiffness and reducing unwanted suspension travel. The result is a firmer suspension setup that better manages the vehicle's weight, enhancing stability without the need for complete shock replacement.

For those interested in implementing this upgrade, a 3D-printable spacer model compatible with the WLToys A979 chassis is available:

Simple 3D print. Two sizes available for different amount of spring preload.

Above is the shock disassembled with the 3D printed spacer next to it.

To overcome this limitation, we designed a custom wire harness that allows three 1100 mAh 2S batteries to be connected in parallel. This upgrade offers several benefits:

Even with the added weight (approximately 3 ounces additional weight), testing shows no adverse effect on motor or ESC behavior. The extra current capacity is well within tolerance for the car’s stock Electronic Speed Controller (ESC).

In addition to combining battery capacity, the custom harness includes:

 

Above is the modular connect that allows for 3 batteries to be connected at once

· Voltage meter

· Step down transformer 7.4V to 5V used to run the DeepRacer compute module off car chassis batteries.

· LEDs PCB mounted to a metal bumper / LEDs power one when the car chassis is powered up.

In the stock DeepRacer configuration, there is no built-in way to visually monitor the chassis battery’s charge level. Pit crews must remove the car’s body shell and manually test voltage using an external meter—or worse, wait until the car exhibits signs of low power such as slow acceleration or delayed starts. This guesswork leads to inconsistent performance, unnecessary pit stops, and unequal racing conditions for participants.

To address this, we integrated a rear-mounted digital voltage meter into the custom wire harness. The meter is clearly visible during operation and provides a real-time readout of the battery’s voltage level, enabling at-a-glance diagnostics without interrupting the race or opening the chassis.

This simple yet effective enhancement empowers pit crews to make informed decisions about battery changes, eliminating unnecessary downtime, ensuring consistent car behavior, and keeping races flowing smoothly. It transforms battery management from reactive to proactive—an essential upgrade for competitive and customer-facing DeepRacer events.

Above is a 3 segment volt meter. This will display the total voltage of all 3 batteries connected to the car. Max voltage when fully charged for a 2S battery is 8.4V, when the voltage drops to 7.00 volts it is recommended to make a battery swap for best performance.
 

The stock AWS DeepRacer bumper is made from lightweight plastic and is prone to bending or cracking during real-world race conditions. Impacts with track barriers, walls, can quickly render the bumper ineffective. Additionally, the stock design lacks a stable mounting point for the LED PCB board mod used for chassis power status indicators, limiting both durability and functionality.

We upgraded the bumper to a rigid steel replacement, offering significantly improved structural integrity. This metal bumper can withstand repeated collisions without deforming and provides a reliable, flat surface for securely mounting an LED light PCB.

The steel bumper is stronger and more resilient able to withstand more than the ABS plastic bumper it replaces. The new bumper accommodates the mounting of LED PCB power indicator lights. This dual-purpose enhancement increases the car’s physical resilience and contributes to a more professional and event-ready setup—improving both the functionality and aesthetics of the DeepRacer platform.

Pictured above is the stock bumper (one of two styles available). It is made of ABS plastic and can brake or become flexable after multiple impacts
 

Above is the DeepRacer car with the steel bumper and LED board installed.

The original AWS DeepRacer design lacks a visual indicator to show when the chassis is powered on. The power switch is small, recessed behind the left front wheel, and difficult to access or see—especially during staging or in fast-paced environments. As a result, pit crews often fumble blindly to confirm power status. In some cases, the car is placed on the track, only for the participant to discover it won’t move—because the chassis was never turned on. This oversight wastes time and causes frustrating race delays.

We installed a custom LED board on the upgraded front metal bumper and wired it directly into the chassis power circuit. As soon as the power switch is activated, the LED lights up—providing a clear, immediate visual cue that the car is powered and ready to race.

This upgrade eliminates power status uncertainty, streamlines the staging process, and helps prevent race interruptions due to unpowered cars. It also adds a polished, professional aesthetic to the car while improving operational flow for pit crews and participants alike.

 

The DeepRacer’s front-facing camera is essential for real-time model inference, continuously capturing visual data from the track. However, stock configurations often perform inconsistently in variable lighting conditions. Overhead lighting reflecting off glossy vinyl tracks can introduce glare, while shadows and uneven lighting can confuse the model, leading to poor lane detection, erratic behavior, or performance degradation—especially in indoor environments.

To address these visibility issues, we enhanced the camera setup with two key modifications:

These upgrades significantly increase model reliability, enhance lap-to-lap consistency, and allow DeepRacer events to run confidently across a wide range of indoor environments. By ensuring the camera sees the track clearly regardless of lighting, participants experience smoother, more predictable races, and facilitators gain greater confidence in event performance.

As the original ASUS power banks used in DeepRacer cars become unavailable, the Dell 12,000 mAh power bank has become the preferred replacement. However, its proprietary USB-C cable is over 4 feet long—excessively long for the compact 1/18 scale chassis. This excess cable creates clutter, routing issues, and an increased risk of snagging or disconnection, especially when removing the car shell or adjusting components during a race.

To improve fit and functionality, we implemented a custom modification to shorten the Dell USB-C cable to an appropriate length for the DeepRacer chassis. The cable is now neatly routed and secured.

This upgrade results in a cleaner and more manageable internal layout, reduces the risk of cable interfering with the shell installation or coming in contact with a wheel. This modification improves overall usability especially during fast-paced race environments where quick access and consistent performance are essential.

The AWS DeepRacer experience continues to be one of the most compelling ways to introduce customers to machine learning, blending technical learning with hands-on excitement. But as with any maturing platform, there are always opportunities to refine and evolve the experience. The enhancements outlined in this blog—ranging from improved wheel alignment and suspension tuning to smarter battery management, lighting upgrades, and cable routing—are grounded in real-world experience from facilitators and pit crews who live and breathe DeepRacer events. These practical, low-cost modifications dramatically increase reliability, reduce downtime, and create a more polished and professional experience for both participants and organizers. By addressing long-standing hardware limitations, we can ensure that DeepRacer events continue to deliver the same "wow" factor while operating more smoothly and consistently than ever before.

About the Author
Todd Bevins is an AWS Technical Account Manager and hands-on AWS DeepRacer facilitator and event specialist. With over 30 years of experience in IT, Todd is passionate about machine learning, customer education, and driving hands-on engagement through technology. He brings a unique blend of technical expertise and real-world RC experience, including work with remote control cars and airplanes. In addition to his cloud and hardware knowledge, Todd is a certified flight and ground instructor and actively tutors students preparing for AWS and CompTIA certifications.