SidewalkPilot-v1.8b

SidewalkPilot-v1.8b is a PyTorch steering model for a small autonomous RC car. It takes a full camera frame as input and predicts a steering servo angle from 0 to 180 degrees.

The model is used for camera-based sidewalk/path following. In the full RC car stack, LiDAR runs above the model as a safety layer for obstacle avoidance, LiDAR override steering, and hard braking.

This is a best-checkpoint (b) variant from the same training version.

Model Details

  • Developed by: Ram Shreyas Naik Sabavat
  • Model type: CNN steering regression model
  • Library: PyTorch
  • License: Apache 2.0
  • Checkpoint: SidewalkPilot-v1.8b.pth
  • Checkpoint created: 2026-05-02 03:51 PM America/Los_Angeles
  • Input: Full camera frame, resized/normalized before inference
  • Evaluation input format: 200x66, OpenCV BGR
  • Output: Steering servo angle from 0 to 180

Specific Improvements

  • Saved the best checkpoint from the v1.8 shadow-fix run.
  • Kept the 20260502_12 shadow fix at 1.08 MAE.
  • Nearly tied v1.8 final while preserving the same raw-BGR Series 1 path.

Specific Issues Observed / Remaining

  • Slightly behind v1.8 final in full score.
  • Still weak on 20260502_19 hard-right/curb-smoothness cases.
  • Did not change the main remaining problem after v1.8: hard-right turns and curb hugging.

Output Meaning

Output Meaning
0 full left
90 straight
180 full right

Evaluation Setup

  • Eval set: 1,287 images
  • Failed samples: 0
  • Corrections included: 564
  • Input format: 200x66, OpenCV BGR
  • Output scale: servo angle 0..180
  • Error unit: servo degrees
  • Score formula: max(0, 100 * (1 - absolute_error / 180))

Version Update Categories

Version Main update category Data/status Result
1.0 Baseline steering initial mixed sidewalk/CARLA set complete
1.0b Best-checkpoint variant of baseline same v1.0 training run complete
1.1 Street-test correction pass manual street corrections complete
1.1b Best checkpoint from street-correction pass same v1.1 training run complete
1.2 Best early field behavior no-weather tuning + manual corrections complete
1.2b Best checkpoint from early field-behavior run same v1.2 training run complete
1.3 Smoothness / aggression tuning same dataset family complete
1.3b Best checkpoint from smoothness/aggression run same v1.3 training run complete
1.4 Curves + sloped shadow fixes photo_20260426 complete
1.4b Best checkpoint from curves/shadows run photo_20260426 complete
1.5 Curved curb interpreted as sidewalk photo_20260427 complete
1.5b Best checkpoint from curved-curb run photo_20260427 complete
1.6 Driveway failure fix photo_20260428 complete
1.6b Best checkpoint from driveway run photo_20260428 complete
1.7 Driveway + shadow field fixes photo_20260429 field validated
1.7b Known-good field rollback photo_20260429 0.6 mi, 0 overtakes
1.8 Mainly shadow fixes photo_20260502_12 offline leader before 1.9
1.8b Best checkpoint from shadow-fix run photo_20260502_12 near tie with 1.8 final

Evaluation Summary

Model Checkpoint Full Score MAE Median AE Max AE Signed Error Within 2° Within 5° Within 10° Within 20°
1.0 SidewalkPilot-v1.0.pth 88.339% 20.990 9.934 151.830 -7.262 202 / 1287 423 / 1287 648 / 1287 869 / 1287
1.0b SidewalkPilot-v1.0b.pth 88.292% 21.075 10.234 152.629 -6.691 179 / 1287 410 / 1287 640 / 1287 866 / 1287
1.1 SidewalkPilot-v1.1.pth 92.372% 13.730 3.886 144.676 -6.197 434 / 1287 746 / 1287 916 / 1287 1044 / 1287
1.1b SidewalkPilot-v1.1b.pth 92.372% 13.730 3.886 144.676 -6.197 434 / 1287 746 / 1287 916 / 1287 1044 / 1287
1.2 SidewalkPilot-v1.2.pth 93.634% 11.459 1.018 140.617 -5.314 748 / 1287 843 / 1287 945 / 1287 1058 / 1287
1.2b SidewalkPilot-v1.2b.pth 93.585% 11.548 1.281 140.920 -5.296 712 / 1287 849 / 1287 948 / 1287 1068 / 1287
1.3 SidewalkPilot-v1.3.pth 93.502% 11.697 1.765 148.944 -5.557 670 / 1287 845 / 1287 955 / 1287 1069 / 1287
1.3b SidewalkPilot-v1.3b.pth 93.223% 12.199 3.005 140.762 -5.423 483 / 1287 823 / 1287 961 / 1287 1070 / 1287
1.4 SidewalkPilot-v1.4.pth 94.079% 10.657 2.194 132.078 -4.997 613 / 1287 899 / 1287 1008 / 1287 1102 / 1287
1.4b SidewalkPilot-v1.4b.pth 93.840% 11.087 2.960 131.227 -5.101 513 / 1287 845 / 1287 1012 / 1287 1099 / 1287
1.5 SidewalkPilot-v1.5.pth 94.337% 10.193 3.711 112.289 -3.734 443 / 1287 766 / 1287 1011 / 1287 1123 / 1287
1.5b SidewalkPilot-v1.5b.pth 94.337% 10.193 3.711 112.289 -3.734 443 / 1287 766 / 1287 1011 / 1287 1123 / 1287
1.6 SidewalkPilot-v1.6.pth 94.540% 9.828 3.953 131.186 -2.877 442 / 1287 776 / 1287 1013 / 1287 1124 / 1287
1.6b SidewalkPilot-v1.6b.pth 94.522% 9.860 3.848 130.713 -3.025 436 / 1287 774 / 1287 1006 / 1287 1126 / 1287
1.7 SidewalkPilot-v1.7.pth 94.956% 9.079 3.523 108.732 -2.665 465 / 1287 801 / 1287 1007 / 1287 1151 / 1287
1.7b SidewalkPilot-v1.7b.pth 94.979% 9.039 3.470 109.148 -2.887 470 / 1287 804 / 1287 1007 / 1287 1150 / 1287
1.8 SidewalkPilot-v1.8.pth 95.702% 7.737 3.273 110.699 -3.228 510 / 1287 831 / 1287 1054 / 1287 1198 / 1287
1.8b SidewalkPilot-v1.8b.pth 95.697% 7.746 3.171 110.106 -3.220 514 / 1287 828 / 1287 1055 / 1287 1196 / 1287

Negative signed error means the model is left-biased on average.

Prediction Distribution

Model Pred Min Pred Max Pred Mean Pred Median Pred P05 Pred P25 Pred P75 Pred P95
1.0 8.459 174.146 89.579 89.676 37.520 68.904 108.229 145.649
1.0b 9.009 173.933 90.149 89.734 38.399 70.772 108.074 144.783
1.1 4.000 176.000 90.643 89.761 37.944 72.754 106.427 152.011
1.1b 4.000 176.000 90.643 89.761 37.944 72.754 106.427 152.011
1.2 4.000 176.000 91.526 91.271 41.237 72.686 108.045 149.390
1.2b 4.000 176.000 91.544 91.210 40.153 72.763 107.837 149.399
1.3 4.000 176.000 91.283 90.486 40.523 73.660 106.509 149.883
1.3b 4.000 176.000 91.417 90.950 41.317 74.564 105.508 149.390
1.4 4.000 176.000 91.843 90.920 40.737 74.986 106.295 154.155
1.4b 4.000 176.000 91.740 90.955 39.214 74.615 105.238 154.870
1.5 4.000 176.000 93.107 90.536 41.311 74.592 109.059 162.565
1.5b 4.000 176.000 93.107 90.536 41.311 74.592 109.059 162.565
1.6 4.000 176.000 93.963 92.289 40.245 73.353 109.929 167.733
1.6b 4.000 176.000 93.816 91.941 40.197 73.651 109.786 167.452
1.7 4.000 176.000 94.175 91.512 39.126 74.491 110.930 165.057
1.7b 4.000 176.000 93.954 91.184 38.861 74.424 110.936 164.965
1.8 4.000 176.000 93.613 90.857 38.486 73.404 109.949 165.694
1.8b 4.000 176.000 93.620 90.740 38.531 73.574 109.985 165.781

Ranking

Rank In This Card Model Checkpoint Score MAE Median AE Max AE Within 5° Within 10° Signed Error
1 1.8 SidewalkPilot-v1.8.pth 95.702% 7.737 3.273 110.699 831 / 1287 1054 / 1287 -3.228
2 1.8b SidewalkPilot-v1.8b.pth 95.697% 7.746 3.171 110.106 828 / 1287 1055 / 1287 -3.220
3 1.7b SidewalkPilot-v1.7b.pth 94.979% 9.039 3.470 109.148 804 / 1287 1007 / 1287 -2.887
4 1.7 SidewalkPilot-v1.7.pth 94.956% 9.079 3.523 108.732 801 / 1287 1007 / 1287 -2.665
5 1.6 SidewalkPilot-v1.6.pth 94.540% 9.828 3.953 131.186 776 / 1287 1013 / 1287 -2.877
6 1.6b SidewalkPilot-v1.6b.pth 94.522% 9.860 3.848 130.713 774 / 1287 1006 / 1287 -3.025
7 1.5 SidewalkPilot-v1.5.pth 94.337% 10.193 3.711 112.289 766 / 1287 1011 / 1287 -3.734
8 1.5b SidewalkPilot-v1.5b.pth 94.337% 10.193 3.711 112.289 766 / 1287 1011 / 1287 -3.734
9 1.4 SidewalkPilot-v1.4.pth 94.079% 10.657 2.194 132.078 899 / 1287 1008 / 1287 -4.997
10 1.4b SidewalkPilot-v1.4b.pth 93.840% 11.087 2.960 131.227 845 / 1287 1012 / 1287 -5.101
11 1.2 SidewalkPilot-v1.2.pth 93.634% 11.459 1.018 140.617 843 / 1287 945 / 1287 -5.314
12 1.2b SidewalkPilot-v1.2b.pth 93.585% 11.548 1.281 140.920 849 / 1287 948 / 1287 -5.296
13 1.3 SidewalkPilot-v1.3.pth 93.502% 11.697 1.765 148.944 845 / 1287 955 / 1287 -5.557
14 1.3b SidewalkPilot-v1.3b.pth 93.223% 12.199 3.005 140.762 823 / 1287 961 / 1287 -5.423
15 1.1 SidewalkPilot-v1.1.pth 92.372% 13.730 3.886 144.676 746 / 1287 916 / 1287 -6.197
16 1.1b SidewalkPilot-v1.1b.pth 92.372% 13.730 3.886 144.676 746 / 1287 916 / 1287 -6.197
17 1.0 SidewalkPilot-v1.0.pth 88.339% 20.990 9.934 151.830 423 / 1287 648 / 1287 -7.262
18 1.0b SidewalkPilot-v1.0b.pth 88.292% 21.075 10.234 152.629 410 / 1287 640 / 1287 -6.691

Field Case Comparison

Model D26 curves/shadows MAE D27 curved curb MAE D28 driveway MAE D29 driveway/shadow MAE 20260502_12 shadow MAE 20260502_19 hard/curb/smooth MAE
1.0 44.12 53.69 56.18 37.44 29.57 29.57
1.0b 42.85 54.06 57.52 37.39 29.16 29.76
1.1 38.93 36.60 83.08 33.44 21.42 26.13
1.1b 38.93 36.60 83.08 33.44 21.42 26.13
1.2 36.54 31.76 64.48 25.38 20.86 22.74
1.2b 35.95 32.11 61.94 25.46 21.01 22.78
1.3 35.00 34.06 74.95 25.26 19.51 22.92
1.3b 34.10 35.63 71.75 24.12 19.66 22.96
1.4 3.66 31.68 70.03 25.83 18.73 24.32
1.4b 3.82 31.91 71.01 26.65 19.21 24.41
1.5 3.98 0.99 87.18 25.74 18.59 21.06
1.5b 3.98 0.99 87.18 25.74 18.59 21.06
1.6 3.93 1.10 2.36 26.32 18.91 22.37
1.6b 4.00 1.10 2.03 26.22 19.02 22.32
1.7 4.22 1.23 1.82 0.93 17.51 22.41
1.7b 4.13 1.22 1.95 1.01 17.43 22.25
1.8 4.27 1.42 2.19 0.96 1.08 22.20
1.8b 4.28 1.42 2.27 0.97 1.08 22.17

Current Version Snapshot

  • Model: 1.8b
  • Checkpoint: SidewalkPilot-v1.8b.pth
  • Checkpoint created: 2026-05-02 03:51 PM America/Los_Angeles
  • Full score: 95.697%
  • MAE: 7.746 servo degrees
  • Median AE: 3.171 servo degrees
  • Rank in this card: 2 of 18 checkpoints

Intended Use

This model is intended for:

  • RC car autonomy experiments
  • Sidewalk/path steering research
  • Raspberry Pi robotics projects
  • Small-scale computer vision control systems
  • Testing steering regression from camera images

Out-of-Scope Use

This model is not intended for:

  • Real cars
  • Public road vehicles
  • Human transportation
  • Safety-critical systems
  • Fully autonomous deployment without external safety layers

System Context

The full RC car autonomy stack uses layered control:

camera frame
-> resize/normalize image
-> PyTorch steering model
-> predicted servo angle (0-180)
-> runtime decision logic
-> LiDAR safety override when triggered
-> final steering/throttle/brake command
-> servo + motor controller

LiDAR runs as a higher-priority safety layer:

LiDAR clear
-> use model steering

LiDAR obstacle
-> LiDAR override mode

LiDAR blocked/too close
-> hard brake

The model handles normal path following, while LiDAR handles obstacle avoidance and emergency behavior.

Training Data

The model was trained on camera images collected from the RC car driving in sidewalk-like environments. Labels represent steering servo angles from 0 to 180 degrees.

Detailed dataset composition belongs in the dataset README, not this model card.

Preprocessing

During inference/evaluation, the pipeline:

  1. Captures the full camera frame.
  2. Resizes the image to 200x66.
  3. Uses OpenCV BGR image ordering.
  4. Normalizes pixel values with (x / 255 - 0.5) / 0.5.
  5. Runs the PyTorch model.
  6. Clamps the output to 0..180.

The model sees the whole frame, not a cropped region.

Limitations

SidewalkPilot models can fail when lighting, sidewalk shape, camera angle, shadows, driveway cuts, curved curbs, hard turns, or curb-hugging cases differ from the training data.

The model does not understand obstacles by itself and is not a standalone safety system.

Use with external safety logic, manual override, and obstacle detection.

Safety Recommendation

Do not use this model alone to control a robot. In the original project, LiDAR has priority over the model and can override steering or trigger hard braking.

Model Card Contact

Ram Shreyas Naik Sabavat

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading