TinyLidarNet checklist: 4 checks before training

The article comes from a participant who reached the final of the End to End AI division of the 自動運転AIチャレンジ2026 using TinyLidarNet, a small model that takes LiDAR distance readings as input and outputs acceleration and steering. Rather than presenting a new model, the write-up collects the problems the author encountered during development into a checklist that later participants can use before starting their own training. The author notes that the model used in the final was run in a local AWSIM environment, not in an actual race.

The central theme is that many failures blamed on the model actually trace back to how data and settings are prepared. Labels from real driving data can fall outside the output layer's tanh range, acceleration multipliers can differ between training and inference, LiDAR scans can be normalized at a different distance than the sensor's actual range, and Hydra overrides can miss the keys the training script actually reads. The article ties these to specific fixes: PR #362 added a common configuration file and acceleration scales, PR #356 corrected the README's Hydra syntax, and PR #117 fixed the LiDAR specification documentation.

For teams preparing their own runs, the practical takeaway is less about the model itself and more about keeping the preparation chain consistent. The author explicitly says the acceleration scaling fix has not been shown to improve driving performance in a comparison, and that the sample normalization was left unchanged at 30 m when the author ran it. Whether these checks translate into fewer wasted training runs likely depends on how closely a team's own extraction and inference pipelines match the settings they record.