benchmark_DS_GP_GMM01.m 9.7 KB
Newer Older
1
function benchmark_DS_GP_GMM01
Sylvain Calinon's avatar
Sylvain Calinon committed
2
% Benchmark of task-parameterized model based on Gaussian process regression, 
3
% with trajectory model (Gaussian mixture model encoding), and DS-GMR used for reproduction.
Sylvain Calinon's avatar
Sylvain Calinon committed
4
%
Sylvain CALINON's avatar
Sylvain CALINON committed
5
% If this code is useful for your research, please cite the related publication:
6
% @article{Calinon16JIST,
Sylvain Calinon's avatar
Sylvain Calinon committed
7
%   author="Calinon, S.",
8 9
%   title="A Tutorial on Task-Parameterized Movement Learning and Retrieval",
%   journal="Intelligent Service Robotics",
10 11 12 13 14 15
%		publisher="Springer Berlin Heidelberg",
%		doi="10.1007/s11370-015-0187-9",
%		year="2016",
%		volume="9",
%		number="1",
%		pages="1--29"
Sylvain Calinon's avatar
Sylvain Calinon committed
16
% }
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
% 
% Copyright (c) 2015 Idiap Research Institute, http://idiap.ch/
% Written by Sylvain Calinon, http://calinon.ch/
% 
% This file is part of PbDlib, http://www.idiap.ch/software/pbdlib/
% 
% PbDlib is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License version 3 as
% published by the Free Software Foundation.
% 
% PbDlib is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
% 
% You should have received a copy of the GNU General Public License
% along with PbDlib. If not, see <http://www.gnu.org/licenses/>.

Sylvain CALINON's avatar
Sylvain CALINON committed
35
addpath('./../m_fcts/');
36

Sylvain Calinon's avatar
Sylvain Calinon committed
37

38 39 40 41 42 43 44 45 46
%% Parameters
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
model.nbStates = 3; %Number of Gaussians in the GMM
model.nbFrames = 2; %Number of candidate frames of reference
model.nbVar = 3; %Dimension of the datapoints in the dataset (here: t,x1,x2)
model.dt = 0.01; %Time step
model.kP = 100; %Stiffness gain
model.kV = (2*model.kP)^.5; %Damping gain (with ideal underdamped damping ratio)
nbRepros = 4; %Number of reproductions with new situations randomly generated
Sylvain Calinon's avatar
Sylvain Calinon committed
47 48
nbVarOut = model.nbVar-1; %(here: x1,x2)
L = [eye(nbVarOut)*model.kP, eye(nbVarOut)*model.kV]; %Feedback gains
49 50 51 52 53 54 55 56 57


%% Load 3rd order tensor data
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
disp('Load 3rd order tensor data...');
% The MAT file contains a structure 's' with the multiple demonstrations. 's(n).Data' is a matrix data for
% sample n (with 's(n).nbData' datapoints). 's(n).p(m).b' and 's(n).p(m).A' contain the position and
% orientation of the m-th candidate coordinate system for this demonstration. 'Data' contains the observations
% in the different frames. It is a 3rd order tensor of dimension D x P x N, with D=3 the dimension of a
Sylvain Calinon's avatar
Sylvain Calinon committed
58 59
% datapoint, P=2 the number of candidate frames, and N=TM the number of datapoints in a trajectory (T=200)
% multiplied by the number of demonstrations (M=5).
60 61 62 63 64 65 66 67 68 69 70 71 72
load('data/DataLQR01.mat');


%% Transformation of 'Data' to learn the path of the spring-damper system
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
nbD = s(1).nbData;
nbVarOut = model.nbVar - 1;
%Create transformation matrix to compute [X; DX; DDX]
D = (diag(ones(1,nbD-1),-1)-eye(nbD)) / model.dt;
D(end,end) = 0;
%Create transformation matrix to compute XHAT = X + DX*kV/kP + DDX/kP
K1d = [1, model.kV/model.kP, 1/model.kP];
K = kron(K1d,eye(nbVarOut));
Sylvain Calinon's avatar
Sylvain Calinon committed
73
%Compute data with derivatives
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134
%Data = zeros(model.nbVar, model.nbFrames, nbD*nbSamples);
Data = s(1).Data0(1,:);
for n=1:nbSamples
	DataTmp = s(n).Data0(2:end,:);
	s(n).Data = K * [DataTmp; DataTmp*D; DataTmp*D*D];
	Data = [Data; s(n).Data]; %Data is a matrix of size M*D x T (stacking the different trajectory samples)
end

%% GPR with GMM encoding
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
fprintf('Parameters estimation of GPR with GMM encoding:');
%GMM encoding for each trajectory sample, learned at once by stacking the different trajectory samples in 'Data' matrix of size M*D x T
model.nbVar = size(Data,1); %Temporary modification of nbVar (for stacked data training)
model = init_GMM_timeBased(Data, model);
model = EM_GMM(Data, model);
model.nbVar = size(s(1).p(1).A,1); %Setting back the initial nbVar
for n=1:nbSamples
	id = (n-1)*2+2:n*2+1;
	s(n).Priors = model.Priors;
	s(n).Mu = model.Mu([1,id],:);
	s(n).Sigma = model.Sigma([1,id],[1,id],:);
	%Regularization of Sigma
	for i=1:model.nbStates
		[V,D] = eig(s(n).Sigma(:,:,i));
		U(:,:,i) = V * max(D,1E-3).^.5;
		s(n).Sigma(:,:,i) = U(:,:,i) * U(:,:,i)';
	end
	%Set query point vector (position and orientation of the two objects)
	s(n).DataIn = [s(n).p(1).b(2:3); s(n).p(1).A(2:3,3); s(n).p(2).b(2:3); s(n).p(2).A(2:3,3)];
	model.DataIn(:,n) = s(n).DataIn;
	%Set model output vector (Mu and Sigma)
	model.DataOut(:,n) = [reshape(s(n).Mu, model.nbVar*model.nbStates, 1); reshape(s(n).Sigma, model.nbVar^2*model.nbStates, 1)];
	%model.DataOut(:,n) = [reshape(s(n).Mu, model.nbVar*model.nbStates, 1); reshape(U, model.nbVar^2*model.nbStates, 1)];
end


%% Reproduction with GPR and DS-GMR for the task parameters used to train the model
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
disp('Reproductions with DS-GMR...');
DataIn = [1:s(1).nbData] * model.dt;
for n=1:nbSamples
	%Rebuild model parameters with GPR
	vOut = GPR(model.DataIn, model.DataOut, s(n).DataIn);
	
	%Re-arrange GPR output as GMM parameters
	r(n).Mu = reshape(vOut(1:model.nbVar*model.nbStates), model.nbVar, model.nbStates);
	r(n).Sigma = reshape(vOut(model.nbVar*model.nbStates+1:end), model.nbVar, model.nbVar, model.nbStates);
% 	U = reshape(vOut(model.nbVar*model.nbStates+1:end), model.nbVar, model.nbVar, model.nbStates);
% 	for i=1:model.nbStates
% 		r(n).Sigma(:,:,i) = U(:,:,i) * U(:,:,i)';
% 	end
	r(n).Priors = model.Priors;
	r(n).nbStates = model.nbStates;
	
	%Regularization of Sigma
	for i=1:model.nbStates
		[V,D] = eig(r(n).Sigma(:,:,i));
		r(n).Sigma(:,:,i) = V * max(D,1E-3) * V';
	end
	
% 	%Retrieval of attractor path through GMR
Sylvain CALINON's avatar
Sylvain CALINON committed
135
% 	currTar = GMR(r(n), DataIn, 1, [2:model.nbVar]); 
136 137 138 139 140 141
% 	
% 	%Motion retrieval with spring-damper system
% 	x = s(n).p(1).b(2:model.nbVar);
% 	dx = zeros(nbVarOut,1);
% 	for t=1:s(n).nbData
% 		%Compute acceleration, velocity and position
Sylvain Calinon's avatar
Sylvain Calinon committed
142
% 		ddx =  -L * [x-currTar(:,t); dx]; 
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
% 		dx = dx + ddx * model.dt;
% 		x = x + dx * model.dt;
% 		r(n).Data(:,t) = x;
% 	end
end


%% Reproduction with GPR and DS-GMR for new task parameters
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
disp('New reproductions with DS-GMR...');
load('data/taskParams.mat'); %Load new task parameters (new situation)
for n=1:nbRepros
	rnew(n).p = taskParams(n).p;
	%Query point vector (position and orientation of the two objects)
	rnew(n).DataIn = [rnew(n).p(1).b(2:3); rnew(n).p(1).A(2:3,3); rnew(n).p(2).b(2:3); rnew(n).p(2).A(2:3,3)];
	
	%Rebuild model parameters with GPR
	vOut = GPR(model.DataIn, model.DataOut, rnew(n).DataIn);
	
	%Re-arrange GPR output as GMM parameters
	rnew(n).Mu = reshape(vOut(1:model.nbVar*model.nbStates), model.nbVar, model.nbStates);
	rnew(n).Sigma = reshape(vOut(model.nbVar*model.nbStates+1:end), model.nbVar, model.nbVar, model.nbStates);
% 	U = reshape(vOut(model.nbVar*model.nbStates+1:end), model.nbVar, model.nbVar, model.nbStates);
% 	for i=1:model.nbStates
% 		rnew(n).Sigma(:,:,i) = U(:,:,i) * U(:,:,i)';
% 	end
	rnew(n).Priors = model.Priors;
	rnew(n).nbStates = model.nbStates;
	
	%Regularization of Sigma
	for i=1:model.nbStates
		[V,D] = eig(rnew(n).Sigma(:,:,i));
		rnew(n).Sigma(:,:,i) = V * max(D,1E-3) * V';
	end
	
	%Retrieval of attractor path through GMR
Sylvain CALINON's avatar
Sylvain CALINON committed
179
	[rnew(n).currTar, rnew(n).currSigma] = GMR(rnew(n), DataIn, 1, [2:model.nbVar]); 
180 181 182 183 184 185
	
	%Motion retrieval with spring-damper system
	x = rnew(n).p(1).b(2:model.nbVar);
	dx = zeros(nbVarOut,1);
	for t=1:nbD
		%Compute acceleration, velocity and position
Sylvain Calinon's avatar
Sylvain Calinon committed
186
		ddx =  -L * [x-rnew(n).currTar(:,t); dx]; 
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212
		dx = dx + ddx * model.dt;
		x = x + dx * model.dt;
		rnew(n).Data(:,t) = x;
	end
end


%% Plots
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
figure('PaperPosition',[0 0 4 3],'position',[20,50,600,450]);
axes('Position',[0 0 1 1]); axis off; hold on;
set(0,'DefaultAxesLooseInset',[0,0,0,0]);
limAxes = [-1.5 2.5 -1.6 1.4]*.8;
myclr = [0.2863 0.0392 0.2392; 0.9137 0.4980 0.0078; 0.7412 0.0824 0.3137];

%Plot demonstrations
plotPegs(s(1).p(1), myclr(1,:), .1);
for n=1:nbSamples
	plotPegs(s(n).p(2), myclr(2,:), .1);
	patch([s(n).Data0(2,1:end) s(n).Data0(2,end:-1:1)], [s(n).Data0(3,1:end) s(n).Data0(3,end:-1:1)],...
		[1 1 1],'linewidth',1.5,'edgecolor',[0 0 0],'facealpha',0,'edgealpha',0.04);
end
for n=1:nbSamples
	plotGMM(r(n).Mu(2:3,:),r(n).Sigma(2:3,2:3,:), [0 0 0], .04);
end
axis equal; axis(limAxes);
Sylvain Calinon's avatar
Sylvain Calinon committed
213
%print('-dpng','-r600','graphs/benchmark_DS_GP_GMM01.png');
214 215

%Plot reproductions in new situations
Sylvain Calinon's avatar
Sylvain Calinon committed
216
disp('[Press enter to see next reproduction attempt]');
217 218 219 220 221 222 223 224 225 226
h=[];
for n=1:nbRepros
	delete(h);
	h = plotPegs(rnew(n).p);
	h = [h plotGMM(rnew(n).currTar, rnew(n).currSigma,  [0 .8 0], .2)];
	h = [h plotGMM(rnew(n).Mu(2:3,:), rnew(n).Sigma(2:3,2:3,:),  myclr(3,:), .6)];
	h = [h patch([rnew(n).Data(1,:) rnew(n).Data(1,fliplr(1:nbD))], [rnew(n).Data(2,:) rnew(n).Data(2,fliplr(1:nbD))],...
		[1 1 1],'linewidth',1.5,'edgecolor',[0 0 0],'facealpha',0,'edgealpha',0.4)];
	h = [h plot(rnew(n).Data(1,1), rnew(n).Data(2,1),'.','markersize',12,'color',[0 0 0])];
	axis equal; axis(limAxes);
Sylvain Calinon's avatar
Sylvain Calinon committed
227
	%print('-dpng','-r600',['graphs/benchmark_DS_GP_GMM' num2str(n+1,'%.2d') '.png']);
Sylvain CALINON's avatar
Sylvain CALINON committed
228
	pause;
229 230 231 232 233 234 235 236 237
end

pause;
close all;
end

%Function to plot pegs
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function h = plotPegs(p, colPegs, fa)
Sylvain CALINON's avatar
Sylvain CALINON committed
238 239 240 241 242 243 244 245 246 247
	if ~exist('colPegs')
		colPegs = [0.2863 0.0392 0.2392; 0.9137 0.4980 0.0078];
		fa = 0.4;
	end
	pegMesh = [-4 -3.5; -4 10; -1.5 10; -1.5 -1; 1.5 -1; 1.5 10; 4 10; 4 -3.5; -4 -3.5]' *1E-1;
	for m=1:length(p)
		dispMesh = p(m).A(2:3,2:3) * pegMesh + repmat(p(m).b(2:3),1,size(pegMesh,2));
		h(m) = patch(dispMesh(1,:),dispMesh(2,:),colPegs(m,:),'linewidth',1,'edgecolor','none','facealpha',fa);
	end
end