Scalable Robust	Learning	from	
Demonstration	with	Leveraged	
Deep	Neural	Networks
Sungjoon	Choi,	Kyungjae Lee,	and	Songhwai Oh
Seoul	National	University
Contents
2Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
§ Leverage Optimization
§ Leveraged Gaussian	Process
§ Leveraged Deep	Neural	Network
§ Conclusion
§ Learning	from	Demonstration
§ Experiment
§ Preliminaries
Learning	from	Demonstration
3
Human	
Expert
Learning	from	
Demonstration
Execute	in	Unseen	
Environments
Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Existing	Limitations
4Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
We	want	to	incorporate demonstrations	with	
mixed qualities without	labeling.
Most	LfD	methods	assume	the	optimality	of	actions.
Existing	Limitations
5Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
We	want	to	incorporate large-scale demonstrations.
Some	LfD	methods	are	not	be	scalable.
Leveraged Gaussian	Process
6Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Choi et.al.,"Leveraged Non-Stationary	Gaussian	Process	Regression	for	Autonomous	Robot	Navigation”,	ICRA	2015
For	example,	
𝛾 = 1.0 𝛾 = 0.8 𝛾 = −0.5 𝛾 = 0
	𝜋 	𝜉,
𝑗 = 1, …, 𝑁
Gaussian	process
	𝜋1 	𝜉1,
𝑖 = 1, … , 𝑀
𝑗 = 1, …, 𝑁
	𝛾1
where		the	correlation	between	𝜋1 and	𝜋14 is	defined	
by	cos
8
9
𝛾1 − 𝛾14 (𝛾1 is	between	-1	and	+1).	
Leveraged	Gaussian	process
Leveraged Gaussian	Process
7Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Positive	definiteness	can	be	shown	with	the	Bochner’s	theorem.
(An	isotropic	kernel	function	is	PSD	iff.	its	Fourier	coefficients	are	non-negative.)	
Leveraged	Kernel	Function Leverage	Gaussian	Process
However, Gaussian	process	regression	is	not	good	for	large	training	data.	
Choi et.al.,"Leveraged Non-Stationary	Gaussian	Process	Regression	for	Autonomous	Robot	Navigation”,	ICRA	2015
Leverage	Optimization
8Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Dataset	with	
Mixed	Qualities
Choi et.al.,"Robust Learning	from	Demonstration	Using	Leveraged	Gaussian	Processes	and	Sparse-Constrained	Optimization”,	ICRA	2016
Leverage	Optimization
9Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Dataset	with	
Mixed	Qualities
leverage:	1 leverage:	0.9 leverage:	0.7 leverage:	-1 leverage:	0
Leverage	Optimization
Choi et.al.,"Robust Learning	from	Demonstration	Using	Leveraged	Gaussian	Processes	and	Sparse-Constrained	Optimization”,	ICRA	2016
Leverage	Optimization
10Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Choi et.al.,"Robust Learning	from	Demonstration	Using	Leveraged	Gaussian	Processes	and	Sparse-Constrained	Optimization”,	ICRA	2016
The	key	intuition	behind	the	leverage	optimization is	that	we	cast	the	
leverage	optimization	problem	into	a	model	selection	problem	in	
Gaussian	process	regression.	
However,	the	number	of	leverage	parameters is	equivalent	to	the	
number	of	training	data.	
To	handle	this	issue,	a	sparse	constrained	leverage	optimization where	
we	assume	that	the	majority	of	leverage	parameters	are	+1	is	
presented.	
where	−L ⋅ is	the	negative	log	likelihood	and	𝛾̅ = 𝛾 − 1.	
Resulting	optimization	problem	becomes:
Leveraged	Neural	Network
11
• We	propose	a	leveraged	deep	neural	network	by	proposing	a	
leveraged	cost	function	by	interpreting	the	objective	function	of	
the	leveraged	Gaussian	processes	using	the	representer theorem.	
Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Representer theorem	+
(Assume 𝜸𝒊 is	either	−1 or	+1.)
Empirical	risk	term
Leveraged	Neural	Network
12Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Input New	TargetLeverage
Parameterized	
Estimator Regularizer
Estimator
Leveraged	
Cost	Function:
New	
Target
Experiment-setting
13Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Input:			 𝑑A
B,𝑑A
C, 𝑑D
B,𝑑D
C, 𝑑E
B,𝑑E
C, 𝑑FGH ∈ 𝑅K
Output:	 𝜃FGH ∈ 𝑅
*Two	hidden	layers	(512	units)	with	tanh activation	functions.
Experiment-data	collection
14Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Experiment-leverage	optimization
15Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Dataset	
with	
Mixed	
Qualities
Safe:Inexp:Suicidal =	70:15:15
Experiment-driving	results
16Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Gaussian	process	
without	optimization
Leveraged	deep	neural	
network	(20,000	demos)
Leveraged	Gaussian	
process	(5,000	demos)
Safe:Inexp:Suicidal =	70:15:15
Conclusion
17Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
§ Robust	and	scalable	learning	from	demonstration	is	presented.	
§ Robustness	comes	from	the	leverage	optimization	[1].	
§ Scalability	comes	from	the	leveraged	deep	neural	network	using	
the	proposed	leveraged	cost	function.	
§ The	proposed	method	is	successfully	applied	to	a	track	driving	
task	where	the	demonstrations	are	collected	from	multiple	modes	
with	different	proficiencies.	
§ Further	work	will	focus	on	incorporating	the	uncertainty	information	
of	a	model	prediction	using	a	Bayesian	network	where	the	initial	
results	can	be	found	in	[2].
[1]	Choi et.al.,"Robust Learning	from	Demonstration	 Using	Leveraged	Gaussian	Processes	 and	Sparse-Constrained	Optimization”,	ICRA	2016
[2]	Choi	et.	al.	‘Uncertainty-Aware	Learning	from	Demonstration	using	Mixture	Density	Networks	with	Sampling-Free	Variance	Modeling’,	ArXiv1709.02249,	2017
18Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Thank	you	for	your	attention.
Contact	information	(sungjoon.choi@cpslab.snu.ac.kr)
19Scalable	Robust	Learning	from	Demonstration	with	Leveraged	Deep	Neural	Networks
Choi	et.	al.	‘Uncertainty-Aware	Learning	from	Demonstration	using	Mixture	Density	Networks	with	
Sampling-Free	Variance	Modeling’,	ArXiv 1709.02249,	2017
Uncertainty-Aware	LfD

LevDNN