deep learning helicopter dynamics models · 2017-03-12 · results on held-out test set stanford...
TRANSCRIPT
Deep Learning Helicopter Dynamics Models
Ali Punjani Pieter Abbeel
UC Berkeley EECS
Latent State: Airflow, Flexibility, Engine Dynamics etc.
Similar trajectories have similar dynamics
Similar trajectories have similar dynamics
acceleration state-control trajectory
Need similarity and local dynamics
acceleration state-control trajectory
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 dynamicsNo labels or annotation
Results on held-out test set
Stanford Autonomous Helicopter Data
0 2 4 6 8 10time (s)�20
�10
0
10
20
30
40
Up-
Dow
nA
cc.(
ms�
2)
circles ObservedLinear Acceleration ModelReLU Network Model
0 2 4 6 8 10time (s)�40
�30
�20
�10
0
10
20
Up-
Dow
nA
cc.(
ms�
2)
flips-loops ObservedLinear Acceleration ModelReLU Network Model
0 2 4 6 8 10time (s)�40�30�20�10
010203040
Up-
Dow
nA
cc.(
ms�
2)
freestyle-aggressive ObservedLinear Acceleration ModelReLU Network Model
Ground Truth Baseline Model Our Model
60% Improvement across all maneuvers
0 2 4 6 8 10
RMS up-down acceleration error (ms�2)
forward-sideways-flight
orientation-sweeps
vertical-sweeps
stop-and-go
turn-demos1
inverted-vertical-sweeps
orientation-sweeps-with-motion
dodging-demos4
circles
flips-loops
chaos
turn-demos2
dodging-demos3
tictocs
dodging-demos1
dodging-demos2
freestyle-gentle
freestyle-aggressive
turn-demos3
Up-Down Acceleration Error
Linear Acceleration ModelLinear Lag ModelQuadratic Lag ModelReLU Network Model
Thanks!
Similar trajectories have similar dynamics
Apprenticeship Learning (Abbeel, Coates, Ng 2010)