Skip to content
NewsletterNov 19, 2026

← Past issues

The Most Dangerous Way to Make Self-Driving Cars Safe

Sent Nov 19, 2026

GA2AD: An Experiment in Training Self-Driving Cars Alongside “Attacker Vehicles”

They conducted training by having an “attack vehicle” deliberately cut in front of the self-driving car and slam on the brakes. The collision rate was reduced by 55–70% in the simulation.

💡 This Week’s Newsletter Summary

  • Deliberately Creating Dangerous Opponents: GA2AD is a training method that uses vehicles surrounding an autonomous vehicle to create dangerous situations such as cutting in, sudden braking, and speeding.
  • Neither too weak nor too strong: Choose the right moment to attack, and alternate between training the self-driving car and the attacking vehicle based on how often collisions occur.
  • 55–70% Reduction in Collisions: In simulations, the collision rate decreased by 55–70%, and in tests where a virtual vehicle was superimposed on a real car, the collision rate estimated by computation was lower than that of AI trained using conventional methods (early-stage results).

1. What Makes GA2AD Different?

The method of deliberately creating challenging situations to train AI is called adversarial training.

This is an approach that has already been studied in the field of autonomous driving. However, there were two problems with the existing method.

First, there were many instances where we needed to understand how the autonomous driving AI performs its calculations internally.

However, the design of autonomous driving AI varies depending on the type. It was difficult to apply a training method developed for one AI directly to another.

Second, in many cases, self-driving cars were unable to respond appropriately unless the intensity of the attack was adjusted.

If the attack is too weak, the vulnerabilities of the self-driving car may not be exposed; if it is too strong, the self-driving car may not be able to handle it, causing the training itself to fail.

It’s similar to sparring. If your opponent is too weak, you won’t improve, and if they’re too strong, you’ll get knocked out before you even have a chance to learn.

📌 Adversarial training: A method of training AI by deliberately creating situations that challenge it. In this study, the attacking surrounding vehicle and the defending autonomous vehicle take turns learning.

GA2AD, developed by a research team led by Harbin Institute of Technology in China and published in Nature Communications, does not delve into the inner workings of autonomous driving AI. Instead, it trains a separate AI to control the surrounding vehicles.

The six vehicles surrounding the self-driving car drive according to pre-established traffic rules.

Then, the AI selects one of the vehicles as an “attack vehicle” and makes it cut in, slam on the brakes, and speed up.

Simply put, we’ve provided the self-driving car with an AI sparring partner.

**📌 Black box: ** This term does not refer to a video recording device installed in a car, but rather to something that is handled without knowing what is inside it. Since GA2AD treats the autonomous driving AI as a black box, we were able to test it by applying it to four types of reinforcement learning-based AI without modifying its architecture.

Normally, the six surrounding cars drive according to the rules (left); during an attack, one car selected by the AI engages in dangerous behavior (right) | Created based on the paper · Reference: Xu et al., GA2AD (2026)

2. There’s a knack to attacking, too.

So, do the attack vehicles charge at any time?

That’s not the case. For training to be effective, the attacks must be neither too easy nor too difficult. GA2AD adjusts this using three mechanisms.

**01 Assigns a risk score to each nearby vehicle. **

The paper examines how dangerous surrounding vehicles are to a self-driving car. It finds that the closer a vehicle is and the faster it is approaching, the more dangerous it is considered to be. The paper refers to this map-like representation of the risks surrounding a vehicle as a “hybrid risk field.”

Then, we assess how closely each car aligns with “Danger,” “Caution,” or “Safe” and summarize the results into a single risk score.

It also calculates a risk score for the entire surrounding area, based not on the average of the six vehicles but on the most imminent threat.

02 Choose only the right moments to attack.

Whether or not to launch an attack is determined by a rule referred to in the paper as the “activation function.” It takes into account the risk score of the vehicle selected as a candidate for attack (the candidate vehicle) and the risk score of the entire surrounding area, and an attack is permitted only when the result falls within a predetermined range.

If the risk is still low, or if the target vehicle’s risk score is significantly lower than that of the surrounding area, it is considered unsuitable as a target for an attack. Conversely, if the surrounding area is already too dangerous, an attack could lead directly to a collision, so we do not attack in such cases.

Ultimately, it skips both the moments when there is no threat and the moments when it’s already too dangerous.

03 Learn one side at a time.

If you teach both at the same time, the subject matter keeps changing, which may prevent effective learning.

So while one side is learning, the other stops learning.

This switching function is handled by the “adaptive switch module.”

The turn is determined by the interval between collisions. If the interval is short—that is, if collisions occur frequently—it is the autonomous vehicle’s turn to learn. Conversely, if the interval is long, it is the attacking vehicle’s turn to learn.

Criteria for selecting the moment to attack (top) and conditions for changing the order of turns (bottom) | Created based on the paper · Reference: Xu et al., GA2AD (2026)

How will the attack vehicle move once it has completed its training? Here is a scene described in the paper:

The attack vehicle behind on the left does not attack at first because its threat score is low.

When the distance closes to about 0.7 seconds and the danger score spikes, the attack begins.

It then picks up speed to catch up and rear-ends the self-driving car in 3.4 seconds.

Self-driving cars learn how to respond by repeatedly encountering situations like this.

3. Results: The number of collisions decreased significantly, while the speed decreased slightly.

The research team implemented a 3.2-kilometer-long road (four lanes, three signalized intersections) based on real-world geographic data within the SUMO traffic simulator. They then compared vehicles trained only in normal environments with those trained alongside attack vehicles.

First, the collision rate.

It indicates how often collisions occur while driving the same distance.

Vehicles that had trained with the attack vehicle had a 69.31% lower collision rate in standard tests without the attack vehicle.

Furthermore, it was 55.24% lower in the test involving attacking vehicles (the attacking AI we faced during training).

The ride has become smoother as well.

When you accelerate or brake suddenly, your body is thrown forward or backward. The measure of how abrupt this acceleration or deceleration is is called “jerk,” and this value decreased by 21.23% and 42.35%, respectively.

**📌 Jerk: ** Strictly speaking, this is a value that indicates how abruptly acceleration changes. The higher the value, the rougher the ride. The paper used the value that separates the top 5% of the roughest moments (the 95th percentile).

Could it be that they slowed down because they were being careful?

The speed did slow down slightly. The average speed decreased by 2.9 m/s (about 10 km/h, 9.18%) in the standard test and by 1.2 m/s (about 4 km/h, 4.24%) in the attack test.

The research team considers this to be an acceptable level, given the improvements in safety.

Even when we changed the autonomous driving AI, the results were similar. When we applied the same reinforcement learning method to four different AI models (DQN, TD3, SAC, and DDPG), the collision rate, jerk, and surrounding hazard scores improved for all four models.

The average speed decreased slightly for all four types, even though no changes were made to the structure.

**📌 Reinforcement Learning: ** A learning method in which an agent learns behaviors that yield high rewards through trial and error. DQN, TD3, SAC, and DDPG are all reinforcement learning algorithms.

How much the number of vehicles trained with attack vehicles has decreased compared to those that received only general training(reduction rate) | Source: Xu et al., GA2AD, Nature Communications (2026), doi:10.1038/s41467-026-75643-z, reconstructed from figures in the text

However, the general driving conditions in the simulation may have been less challenging than those in real life. The accident rate for vehicles trained and tested under general driving conditions was one accident per 1 million kilometers, which was lower than the actual figure recorded in China’s pilot operation zones (approximately 6.48 accidents per 1 million kilometers).

The research team believes this may be because the surrounding cars in the simulation moved more cautiously than people do.

In that case, was it really necessary to adjust the attack intensity?

When we mechanically changed the training sequence at fixed intervals, regardless of collisions, the self-driving car’s learning score (reward) remained lower and more erratic even after prolonged training.

On the other hand, when we removed the activation function, the attack became less effective.

Close calls have become as rare as they were during regular training. In the end, the GA2AD, equipped with both devices, managed to stay in a situation that, while not easy, was still manageable.

4. I placed a virtual car in the actual test site.

So, will this work in real-world vehicles? The research team conducted tests using an actual autonomous vehicle (the Dongfeng Chenglong H7) at the Heihe Autonomous Driving Test Site in Heilongjiang Province, China.

We created the surrounding cars using augmented reality. When we overlaid virtual vehicles onto the test car’s camera footage and LiDAR (a sensor that measures distance using lasers) data in real time, the test car recognized them as real cars and reacted accordingly.

As the name suggests, the road and the test vehicle are real, while the surrounding vehicles posing a threat are virtual. This is not a test conducted while driving among regular drivers.

This environment was much more challenging. In 34.97% of the recorded cases, there were less than 2 seconds remaining until a collision at the current speed—approximately 58 times higher than in the standard training environment (0.60%).

The results followed the same pattern as the simulation.

All four AI models trained using GA2AD had lower collision rates than those trained using standard methods.

However, this collision rate is not a direct measurement of the frequency of accidents on actual roads.

It is a “virtual-physical fusion collision rate” that combines the probability of a collision in a hazardous situation (measured at a test site) with the frequency of such situations occurring (estimated through simulation), and the research team describes this as evidence that is still in its early stages.

5. What My Sparring Partner Taught Me and What I Still Don’t Know

This study showed that, even without delving into the inner workings of the autonomous driving AI, training it in alternation with an attack vehicle that adjusts its speed can reduce collisions and aggressive driving in the tested attack scenarios.

But there’s still a lot I don’t know.

The fact that there are many instances where the time remaining until a collision is short simply means that the vehicle was frequently exposed to danger; it does not directly indicate the actual risk of an accident or the severity of potential damage. While the sequence of accident types caused by the autonomous vehicle was similar to actual U.S. accident statistics (NHTSA), the extent to which its detailed movements resemble those of a human driver has not yet been verified by numerical data. The duration and distance of the evaluation were also limited. Ultimately, whether the test track results—which showed the vehicle to be safer than a standard training car—will hold true on actual roads remains a question that must be answered through large-scale road testing.


You can also design training partners for self-driving cars.

Dangerous moments don’t often come just because you wait for them. Can self-driving cars trained on simulated attack vehicles also avoid unexpected drivers on real roads?

Leave feedback on this issue