Questions: Imitation Learning and Learning from Demonstrations
1 questions to test your understanding
Score: 0 / 1
Question 1 Multiple Choice
A robot learns to pour water into a cup using behavioral cloning from 50 human demonstrations. The robot watches a human pour, records camera images and robot joint angles, and trains a neural network to predict joint velocities from images. After training, the robot begins pouring but makes small errors (hand slightly to the left) in the first second. Why does this small error often lead to catastrophic failure (water missing the cup entirely)?
AThe neural network has low accuracy and cannot learn pouring well
BThe human demonstrations were poor quality
CDistribution shift: small deviations push the robot into states unseen during training. The learned policy has no experience recovering from such states, so errors compound over time
DThe robot's actuators are imprecise
This is the standard industrial approach: use imitation learning for fast initialization (avoiding RL sample inefficiency), then refine with RL (for robustness and optimality). This hybrid strategy is used in robot manipulation at companies like DeepMind, OpenAI, and Boston Dynamics, achieving impressive real-world results by combining the speed of behavioral cloning with the robustness of learned rewards.