Behavior Cloning on a Sim Arm

Key Insight

Behavior cloning (BC) is the foundational baseline for imitation learning, framing behavioral reproduction as standard supervised learning where neural networks predict motor actions from current sensory observations. By training a simple MLP policy on joint trajectories recorded during human teleoperation, the robot learns to reproduce the demonstration task without requiring a complex, hand-engineered reward function. However, because the policy only minimizes prediction errors on the training states, it remains highly vulnerable to covariate shift at test time, where small errors accumulate and steer the arm into unfamiliar configurations from which it cannot recover.