Deep Learning Helicopter Dynamics Models Ali Punjani Pieter Abbeel - PowerPoint PPT Presentation

Deep Learning Helicopter Dynamics Models Ali Punjani Pieter Abbeel UC Berkeley EECS

Latent State: Airflow, Flexibility, Engine Dynamics etc.

Similar trajectories have similar dynamics

acceler acceleration ation stat state-contr e-control tr ol traject ajector ory Similar trajectories have similar dynamics

acceler acceleration ation stat state-contr e-control tr ol traject ajector ory Need similarity and local dynamics

Hierarchical Network Model Input raw 0.5 second trajectory; Output acceleration

Hierarchical Network Model

Hierarchical Network Model Jointly learn partitions of input space and local dynamics No labels or annotation

Stanford Autonomous Helicopter Data 40 Up-Down Acc. (ms − 2 ) circles Observed 30 Linear Acceleration Model ReLU Network Model 20 10 0 − 10 − 20 0 2 4 6 8 10 time (s) 20 Up-Down Acc. (ms − 2 ) flips-loops Observed 10 Linear Acceleration Model ReLU Network Model 0 − 10 − 20 − 30 − 40 0 2 4 6 8 10 time (s) 40 Up-Down Acc. (ms − 2 ) freestyle-aggressive Observed 30 Linear Acceleration Model 20 ReLU Network Model 10 0 − 10 − 20 − 30 − 40 0 2 4 6 8 10 time (s) Ground Truth Baseline Model Our Model Results on held-out test set

Up-Down Acceleration Error turn-demos3 freestyle-aggressive freestyle-gentle dodging-demos2 dodging-demos1 tictocs dodging-demos3 turn-demos2 chaos flips-loops circles dodging-demos4 orientation-sweeps-with-motion inverted-vertical-sweeps turn-demos1 Linear Acceleration Model stop-and-go Linear Lag Model vertical-sweeps Quadratic Lag Model orientation-sweeps ReLU Network Model forward-sideways-flight 0 2 4 6 8 10 RMS up-down acceleration error (ms − 2 ) 60% Improvement across all maneuvers

Thanks!

Apprenticeship Learning (Abbeel, Coates, Ng 2010) Similar trajectories have similar dynamics