What was the main difference between the two simulators?

The self-built GPU simulator ran 256 environments and calculated 13.1k steps per second in total, while AWSIM ran up to 4 cars in real time at 20 steps per second. The custom simulator was used for data collection and candidate screening, but AWSIM was used for final acceptance.

Did the team fix the issues they found?

Yes, they identified five traps including steering multiplier, vehicle model, wall shape, opponent movement, and CPU contention. They fixed them by calibrating constants in AWSIM, using random courses, reverting to constant-speed opponents, and setting BLAS threads to 1.

What was the final result in the competition?

The team placed second (runner-up) in the End to End AI division of the 自動運転AIチャレンジ2026. Their final model completed six laps alone with zero penalties at 278.9 seconds, and in a four-car mixed run all four cars finished with zero car-to-car contacts.

RoboticsAutonomous DrivingQiita 機械学習14h agoOct 6, 2026, 01:00 JST

Sim-trained model with 90.2% success fails AWSIM, 0/4 finish

Send on LINE

3 Key Points

What happened
KSK's intermediate model, trained in a self-built GPU simulator, hit a 90.2% no-collision rate over 45 seconds, yet finished 0 of 4 cars in AWSIM's four-car, six-lap race. Another version with 83.6% finished 3 of 4 in one run.
Why it matters
This shows that a model's high scores inside a custom simulator do not guarantee it will work in the official simulator, so decisions based only on the custom simulator's numbers are risky.
What to watch
The outcome hinges on running multiple evaluation runs in AWSIM, since single-run results can be misleading. The article's final model completed six laps alone with zero penalties at 278.9 seconds, but its fastest single lap was about 5.6 seconds slower than the initial version trained on the actual course.

WHO IT HITSThis lands on machine learning engineers and robotics competition teams who rely on custom simulators for training and selection; they may need to validate their models in the target environment before committing.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The team, KSK, a single technical college student, trained a model called TinyLidarNet using only 2D LiDAR data. They built their own GPU simulator that ran 256 environments in parallel for fast data collection. However, when they transferred the model to the official simulator, AWSIM, it failed to finish a four-car, six-lap race. The article details five differences that caused the failure: steering multiplier, vehicle model, wall shape, opponent movement, and CPU contention.

The steering multiplier was a major issue. The custom simulator applied a factor of 0.448 to the model's output, but AWSIM sent the output directly. When tested, a multiplier of 0.448 gave a lap time of 93 seconds, while 0.80 gave 43 seconds. The vehicle model in the custom simulator was about 25% heavier, causing speeds to cap at 4-5 m/s. Wall shapes differed by an average of 1.28 m, so the team trained on 80 random courses to avoid memorizing specific walls. Opponent movement in the custom simulator assumed 25% of the time at 5 km/h, but actual AWSIM traffic showed only 2.4% and 0.0%. CPU contention was a hidden trap: BLAS threads were set to 24 per node, causing load average to reach 84. Setting threads to 1 improved finish rate from 8/12 to 12/12.

The team's final approach was to use the custom simulator for data collection and candidate screening, but to make acceptance decisions based on multiple AWSIM runs. This experience shows that sim-to-sim transfer can be as tricky as sim-to-real, and that validation in the target environment is essential.

FAQ

What was the main difference between the two simulators?: The self-built GPU simulator ran 256 environments and calculated 13.1k steps per second in total, while AWSIM ran up to 4 cars in real time at 20 steps per second. The custom simulator was used for data collection and candidate screening, but AWSIM was used for final acceptance.
Did the team fix the issues they found?: Yes, they identified five traps including steering multiplier, vehicle model, wall shape, opponent movement, and CPU contention. They fixed them by calibrating constants in AWSIM, using random courses, reverting to constant-speed opponents, and setting BLAS threads to 1.
What was the final result in the competition?: The team placed second (runner-up) in the End to End AI division of the 自動運転AIチャレンジ2026. Their final model completed six laps alone with zero penalties at 278.9 seconds, and in a four-car mixed run all four cars finished with zero car-to-car contacts.

Qiita 機械学習Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.