Behavior Cloning Drift Simulator

A robot end-effector is trained purely by copying one human-demonstrated trajectory. Every run replays that trajectory twice in parallel: a naive behavior-cloning policy whose small per-step errors compound because it never saw off-path states during training, and a DAgger-style policy that receives periodic corrective guidance whenever it starts to drift off the demonstrated distribution. Tune the noise, compounding gain and correction strength, run the trajectory once or set it auto-running, and watch the cumulative-drift chart show how the two training strategies diverge over repeated attempts.