SidewalkPilot-v3.3b
SidewalkPilot-v3.3b maps one 320x180 OpenCV BGR frame to steering control for the SidewalkPilot RC car. It is the best-validation checkpoint from the stronger tree-shadow augmentation run. Physical testing found a severe regression relative to v3.2b, so it is preserved as evidence that validation selection did not produce a field-ready checkpoint.
Model Details
- Developer: Ram Shreyas Naik Sabavat
- Series: 3.x
- Architecture: SidewalkPilotV3 CNN, approximately 5.53 million parameters
- Model type: 9-class steering classification plus within-bucket offsets and throttle
- Checkpoint role: best-validation
- Field status: tested July 13, 2026; not selected
- Artifact:
SidewalkPilot-v3.3b.onnx(FP32 ONNX) - Checkpoint created: 2026-07-11 13:50 PDT
- Input:
[batch, 3, 180, 320], OpenCV BGR - Preprocessing: resize to
320x180, then(x / 255 - 0.5) / 0.5 - Output:
[batch, 19]: 9 class logits, 9 offsets, and throttle
Output Decoding
class_id = argmax(output[0:9])
fraction = sigmoid(output[9 + class_id])
steering_deg = bucket_low[class_id] + fraction * bucket_width[class_id]
throttle = sigmoid(output[18]) # not used by the current runtime
Logical steering uses 0 for full left, 90 for straight, and 180 for full right. The throttle output is not deployed because the collected throttle labels do not contain enough variation to train useful throttle control.
Evaluation Setup
- Evaluation set: 81,237 current Series 3 real field images
- Purpose: training-set fit check, not held-out generalization proof
- Error unit: logical steering degrees
- Selection priority: balanced 9-bucket accuracy and turn recall first; MAE is supporting evidence
- Cross-series warning: do not compare these numbers directly with Series 1/2, which use their original 2,224-image dataset
Chronological Metrics Through This Version
| Model | Bal9 | Turn exact | Turn +/-1 | Straight exact | MAE | Median AE | Signed error |
|---|---|---|---|---|---|---|---|
3.0 | 16.0% | 15.3% | 42.9% | 23.4% | 18.971 | 13.623 | -3.594 |
3.0b | 15.9% | 14.8% | 42.3% | 24.6% | 18.450 | 13.129 | -3.944 |
3.1 | 28.1% | 26.8% | 53.3% | 55.2% | 22.647 | 9.729 | -2.353 |
3.1b | 27.4% | 25.8% | 52.3% | 56.7% | 20.958 | 9.591 | -3.769 |
3.2 | 34.0% | 31.1% | 57.7% | 52.4% | 16.776 | 9.648 | -1.474 |
3.2b | 25.1% | 19.8% | 46.0% | 67.2% | 14.457 | 5.909 | -4.691 |
3.3 | 27.1% | 21.2% | 48.1% | 68.5% | 14.630 | 5.204 | -5.351 |
3.3b (this model) | 19.2% | 10.6% | 36.5% | 78.0% | 16.460 | 2.183 | -8.287 |
Class-Balanced Ranking Through This Version
| Rank | Model | Bal9 | Turn exact | Turn +/-1 | MAE | Median AE | Signed error |
|---|---|---|---|---|---|---|---|
1 | 3.2 | 34.0% | 31.1% | 57.7% | 16.776 | 9.648 | -1.474 |
2 | 3.1 | 28.1% | 26.8% | 53.3% | 22.647 | 9.729 | -2.353 |
3 | 3.1b | 27.4% | 25.8% | 52.3% | 20.958 | 9.591 | -3.769 |
4 | 3.3 | 27.1% | 21.2% | 48.1% | 14.630 | 5.204 | -5.351 |
5 | 3.2b | 25.1% | 19.8% | 46.0% | 14.457 | 5.909 | -4.691 |
6 | 3.3b | 19.2% | 10.6% | 36.5% | 16.460 | 2.183 | -8.287 |
7 | 3.0 | 16.0% | 15.3% | 42.9% | 18.971 | 13.623 | -3.594 |
8 | 3.0b | 15.9% | 14.8% | 42.3% | 18.450 | 13.129 | -3.944 |
Current Version Snapshot
- Balanced 9-bucket exact: 19.2%
- Turn exact: 10.6%
- Turn within one bucket: 36.5%
- Straight exact: 78.0%
- MAE: 16.460 degrees
- Median absolute error: 2.183 degrees
- Signed error: -8.287 degrees
Field Evaluation
v3.3b was tested on the physical car on July 13, 2026 in the same four-model comparison as v3.3, v3.4, and v3.4b. The comparison included harsh-shadow cases and normal left/right turns. The operator reported that v3.3b was much worse than v3.2b and did not promote it.
The offline table is consistent with that concern: v3.3b has very high straight exact recall but low balanced and turn recall, plus a large negative signed error. The field result is still qualitative because the exact route, weather, duration, manual-takeover count, and linked clip identifiers were not preserved.
Intended Use
- Small RC-car autonomy experiments
- Sidewalk/path steering research
- Jetson ONNX/TensorRT deployment experiments
Limitations and Safety
This model does not identify obstacles, prove a clear path, estimate confidence, or detect out-of-distribution scenes. Shadows, lighting changes, curb geometry, driveways, camera movement, and underrepresented turns can cause unsafe steering. Keep independent LiDAR emergency braking, manual takeover, conservative speed limits, and bounded operating conditions above model output. Its July 13 field regression is a reason not to use it as the current default.
Reproducibility
artifact_manifest.json records the artifact SHA-256, tensor signature, checkpoint role, evaluation set, and current report metrics.