Choosing a navigation stack with limited labeled driving data
Case context: A small company is building a low-speed campus shuttle. It must choose between a system that maps front-camera video directly to steering commands and a modular system with separate detectors for lane markings, vehicles, and pedestrians that feed a planner. The team can easily obtain tens of thousands of labeled images showing lanes, cars, and people, but it has only a few short human-driving logs with synchronized camera frames and steering commands.
Question: Using the idea of data availability, which design is the better fit, and why? Explain how the available data changes the choice.
Sample answer: A modular pipeline is the better fit. A direct end-to-end model needs many paired examples of camera input and steering output, and those examples are costly to collect because they require driving demonstrations. In contrast, the team already has access to large amounts of labeled imagery for intermediate perception tasks such as detecting lanes, cars, and pedestrians. Since the easier data supports those intermediate modules, the modular design is more practical for the team’s budget and schedule.
Key points:
- Choose the modular pipeline with intermediate detectors.
- Paired camera-to-steering examples are scarce and expensive to gather.
- Labeled perception data is plentiful, making intermediate modules easier to build.
Rubric: The student should identify the modular pipeline as the better choice, noting that image-to-steering pairs are costly to collect while labeled lane/car/pedestrian images are readily available.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Why can limited data availability make an intermediate detector a better choice than a fully end-to-end driving system?
True or False: Gathering paired examples of street images and steering commands for an autonomous vehicle is typically no harder than labeling each image with objects such as pedestrians or trucks.
For a direct end-to-end warehouse navigation model, each training example must be an (image, _____) pair.
Why is it usually easier to gather data for an object detector than for an end-to-end control policy?
A fully end-to-end self-driving model is straightforward to train because labeled image-and-steering examples are cheap and quick to gather.
What paired examples are needed to train a direct driving policy from video?
Match each pipeline idea to the data situation it fits best in a factory inspection setting.
Order the steps for deciding whether to use a staged pipeline instead of a direct end-to-end model for a robot navigation task.
When direct pairs of sensor images and control commands are scarce, which model design is most strongly encouraged by the data situation?
It is hard to find labeled images of cars and pedestrians for training intermediate detectors.
If you have plenty of labeled data for the _____ parts of a workflow, a multi-stage design can be a sensible choice.
Match each data situation to the design lesson it suggests in an autonomous driving project.
Order the reasoning steps for why a staged perception pipeline can be easier to build than a pure end-to-end driving model.
How Data Collection Constraints Shape the Choice Between End-to-End and Modular Autonomous Driving Systems
Choosing a navigation stack with limited labeled driving data
Why modular driving systems are easier to train than fully end-to-end ones