Undergraduate Research · 2026

Pathfinder: Autonomous Bicycle

Embedded lane keeping, road perception, and steer-by-wire control for a low-speed autonomous bicycle.

Aman Nindra·University of California, Merced

Undergraduate Researcher · May 2026 – Present

DemoMethodImages
Roadway inference output showing lane boundaries, detected vehicles, target heading, steering angle, lateral offset, and confidence
Road-guidance output showing detected vehicles, fitted lane boundaries, target heading, steering command, lateral offset, and confidence.

Abstract

An embedded perception and steering research prototype for low-speed autonomous bicycle navigation, combining lane and road understanding, object detection, depth estimation, and Stanley-style steering guidance in an offline video pipeline.

  • LaneATT ego-lane detection reached a best validation F1 of 78.52% after fine-tuning on CULane.
  • YOLOv11 was reduced to four driving-relevant classes: person, vehicle, traffic light, and stop sign.
  • ONNX Runtime and TensorRT deployments were benchmarked on a Jetson Orin Nano Super, with TensorRT cutting inference latency by up to 4.55x, while ROS 2 and Gazebo support path-following tests.

Measured results

Evaluation results recorded in the project's training and validation logs.

LaneATT best validation F1
78.52%
YOLOv11 mAP@50
73.23%
HybridNets road IoU
83.88%

Deployment benchmarks

ONNX Runtime vs. TensorRT and power-mode results measured on an NVIDIA Jetson Orin Nano Super.

TensorRT speedup — LaneATT
4.55×
TensorRT speedup — YOLOv11n
1.83×
Throughput gain from power mode
1.92×
Quantized-model output correlation
≥0.999999

Method overview

A four-stage offline pipeline connects visual perception to a steering estimate and deployment experiments.

  1. 01

    Vision & perception

    LaneATT and HybridNets identify lane and road structure while YOLOv11 detects four driving-relevant object classes.

  2. 02

    Path extraction

    Temporal confidence, ego-lane pairing, missing-edge recovery, and road-center fitting produce a stable path estimate.

  3. 03

    Steering guidance

    Path-heading and cross-track errors feed a Stanley-style controller to produce an offline steering estimate.

  4. 04

    Deployment research

    ONNX Runtime and TensorRT tooling target Jetson-class inference, with ROS 2 and Gazebo used for path-following experiments.

Quick preview

Play the short offline clips to inspect road-guidance and relative depth outputs frame by frame.

Road guidance. Lane and object inference with live target-heading, steering, lateral-offset, and confidence overlays.
Depth estimation. Monocular depth-estimation output separates nearby roadway structure and vehicles from the distant scene.

Implementation

The research repository includes training and inference work for lane detection, road segmentation, YOLOv11 object detection, depth estimation, and Jetson-oriented ONNX/TensorRT execution. The road-guidance runtime extracts a center path and computes a Stanley steering command from path-heading and cross-track error.

LaneATT, YOLOv11, HybridNets, and Depth Anything V2 are published research systems adapted here for training, integration, guidance, and edge-deployment experiments; their base architectures are not claimed as original work.

LaneATTYOLOv11HybridNetsDepth Anything V2PyTorchONNX RuntimeTensorRTROS 2CUDAModel Quantization

Current limitations

  • The displayed videos are offline pipeline outputs, not evidence of closed-loop autonomous bicycle operation.
  • Depth Anything V2 currently provides relative depth; metric distance and vehicle-speed estimation are not yet implemented.
  • Steering estimates depend on camera and lane-width assumptions and have not been safety validated for riderless operation.

Continue exploring

Return to the portfolio experience section for the full research role and related work.

Back to experience