| # | Score | Max CTE | Laps | Description |
|---|---|---|---|---|
| 8 | 72.00 | 2.56 | 1 |
# Experiment: Compact SEResNet (~112K params)
Hypothesis: ## Hypothesis The current best model (run 8, score 153.53) uses a SEResNet with channels 24→48→96→128, totaling ~385K params. The suggestion is to make the model smaller. Smaller models can generalize better to unseen tracks by having less capacity to overfit to training track-specific features. This experiment keeps the same proven architecture (SE attention, residual blocks, GAP, anticipatory steering) but reduces channel widths by ~50%: 16→32→48→64. This yields ~112K params — roughly 3.5x sma
|
| 7 | 148.52 | 4.57 | 2 |
# Experiment: Deeper CNN with Global Average Pooling + Anticipatory Steering
Hypothesis: ## Hypothesis The current best model (run 4, score 99.55) used anticipatory steering labels with the baseline 3-conv NvidiaCNN architecture. That architecture uses a massive flatten operation (13K → 50) which loses spatial information and may limit generalization to unseen tracks. This experiment combines two ideas: 1. **Anticipatory steering labels** (K=5, proven in run 4) — teaches the model to steer preemptively 2. **Deeper CNN with Global Average Pooling** — replaces flatten+FC with adaptive
|
| 6 | 153.53 | 3.56 | 2 |
# Experiment: ResNet-style CNN with SE Attention
Hypothesis: ## Hypothesis The current best model (run 6, score 148.52) uses a plain 4-layer CNN with GAP. While effective, plain CNNs suffer from two limitations: 1. **No skip connections** — deeper layers must learn the full transformation, making optimization harder and potentially limiting generalization to unseen tracks. 2. **No channel attention** — all feature channels are weighted equally, even though some (e.g. road edges, lane markings) are much more informative than others (e.g. sky, background).
|
| 5 | 99.55 | 4.48 | 1 |
# Experiment: Anticipatory Steering Labels
Hypothesis: ## Hypothesis The current NN model trains on frame-action pairs where each frame is paired with the steering action taken at that exact moment. But effective driving requires anticipation — by the time you see a curve, you should already be starting to steer. By shifting steering labels forward by K frames (e.g. K=5 at 20fps = 250ms lookahead), each frame gets paired with the steering value that was applied ~250ms in the future. This teaches the model to steer preemptively: when it sees a curve
|
| 4 | 86.77 | 3.67 | 3 | Simple PD controller pilot |
| 3 | 55.11 | 4.18 | 2 | Simple NN pilot (NvidiaCNN) |
| 2 | 0.00 | 0.00 | 0 |
# Experiment: Recovery Data via Horizontal Translation Augmentation
Hypothesis: ## Hypothesis The current NN model (run_000003, score 55.11) trains only on "on-track" PID driving data where the car is centered on the road. When the model deviates even slightly during evaluation, it encounters visual states it never saw in training and doesn't know how to recover, leading to early out-of-bounds exits. The suggestion is to create training data of the car driving off-track and pair it with corrective steering labels. Rather than collecting this data live in the simulator (frag
|
| 1 | 8.84 | 4.12 | 0 | Go-fast-straight baseline |