SidewalkPilot-v4.0c
SidewalkPilot-v4.0c is a Series 4.0 PCF steering model for the SidewalkPilot RC car. Its training contract uses image plus three previous targets with current and three future targets used as training supervision. It is the best-validation (epoch 7) checkpoint paired with v4.0a.
Model Details
- Developer: Ram Shreyas Naik Sabavat
- Series: 4.0 experimental temporal steering
- Architecture: SidewalkPilotV4 CNN, 5,572,696 parameters
- Contract: PCF
- Checkpoint role: best-validation (epoch 7)
- Field status: tested; slightly worse than v4.0a and rejected for steering echo
- Artifact:
SidewalkPilot-v4.0c.onnx(FP32 ONNX) - Checkpoint created: 2026-07-14 PDT
- Inputs:
image[batch,3,180,320]andtarget_history[batch,3]in logical steering degrees - History behavior: the model normalizes the three absolute prior targets internally
- Image preprocessing: OpenCV BGR, resize to
320x180, then(x / 255 - 0.5) / 0.5 - Output:
[batch,4,18]; each horizon contains 9 class logits and 9 class-local offsets - Live steering horizon: horizon 0
- Throttle output: none
Output Decoding
class_id = argmax(softmax(horizon[0:9]))
fraction = sigmoid(horizon[9 + class_id])
steering_deg = bucket_low[class_id] + fraction * bucket_width[class_id]
Logical steering uses 0 for full left, 90 for straight, and 180 for full right. Future targets are training labels only. They are never runtime inputs.
Common Evaluation
- Evaluation set: 6,952 frozen, sequence-valid anchors from the 81,237-image real Series 3/4 dataset
- Purpose: common open-loop comparison across model families
- Error unit: logical steering degrees
- Selection priority: Bal9 and turn recall first; MAE is supporting evidence
- Important limit: PC/PCF evaluation uses ground-truth prior targets, while live driving feeds prior predictions
| Model | Bal9 | Turn exact | Turn +/-1 | ST exact | MAE | Median AE | Signed error |
|---|---|---|---|---|---|---|---|
4.0p | 34.5% | 32.1% | 65.9% | 67.7% | 12.396 | 2.967 | +0.120 |
4.0r | 32.9% | 27.4% | 62.6% | 77.6% | 11.636 | 1.846 | -1.136 |
4.0f | 25.4% | 23.5% | 56.4% | 62.8% | 15.623 | 6.723 | +1.057 |
4.0g | 20.4% | 17.1% | 46.4% | 76.0% | 14.116 | 2.114 | -1.864 |
4.0a | 33.5% | 30.9% | 65.3% | 68.1% | 12.379 | 3.115 | +0.290 |
4.0c (this model) | 32.0% | 29.4% | 62.9% | 75.5% | 11.321 | 1.825 | -0.981 |
Field Evaluation
v4.0c repeatedly held earlier steering predictions and was slightly worse than v4.0a within the PCF pair. Its low offline MAE did not translate into usable closed-loop driving.
The comparison was supervised and video was captured, but it was not a formal route-controlled benchmark with a preserved per-case score sheet. The claim is limited to the cases presented.
Intended Use
- Small RC-car steering experiments
- Temporal behavioral-cloning research
- Jetson Orin Nano ONNX deployment experiments
- Reproducing the documented difference between open-loop metrics and closed-loop behavior
Limitations and Safety
This model does not identify obstacles, prove a clear path, estimate confidence, or detect out-of-distribution scenes. The Series 4.0 history models can amplify their own earlier predictions during closed-loop driving. Do not treat an offline score or successful ONNX load as approval for autonomous use. Keep independent LiDAR emergency braking, manual takeover, conservative speed, and bounded supervised testing above model output.
Reproducibility
artifact_manifest.json records the model SHA-256, tensor signature, checkpoint role, field status, and common-set metrics.