Section 7.5 Glossary

End-to-End Learning

🔗

An approach that replaces a traditional pipeline of hand-crafted stages (object detection, path planning, motor control) with a single machine learning model that maps raw sensor input directly to motor output.

🔗

Behavioral Cloning

🔗

A form of supervised learning where a robot learns to mimic a human expert’s recorded demonstrations, training on paired (sensor input, control output) examples.

🔗

Data Collection Loop

🔗

The four-step behavioral cloning workflow: record human driving demonstrations, preprocess the data, train a regression model on it, and evaluate the trained model before deployment.

🔗

Regression (Control Estimation)

🔗

The supervised learning task of predicting a continuous numerical output, such as linear velocity \(v\) or angular velocity \(\omega\text{,}\) rather than a discrete category.

🔗

Mean Squared Error (MSE)

🔗

An evaluation metric computed as the average of squared differences between predicted and target values.

🔗

Mean Absolute Error (MAE)

🔗

An evaluation metric computed as the average of the absolute differences between predicted and target values, used to check whether a trained model’s error is small enough for safe deployment.

🔗

Dataset Bias

🔗

A systematic imbalance in training data that causes a behavioral cloning model to perform poorly on underrepresented conditions or actions.

🔗

Right-Turn Trap

🔗

A classic dataset bias failure mode where a training set dominated by right-turn demonstrations (e.g., from driving one direction around a loop) teaches a model to always favor turning right, even when a left turn is required.

🔗

Data Augmentation

🔗

Generating additional synthetic training examples from existing data, such as horizontally flipping a camera image and negating its steering command (\(\omega_{\text{new}} = -\omega\)) to create a balanced left-turn example from a right-turn one.

🔗

Recovery Trajectories

🔗

Training data intentionally recorded from off-center or near-obstacle positions, demonstrating how to steer back toward the center, so the model learns to recover from drift rather than only imitating perfect driving.

🔗