SA-1 Robotic Self-Perception and Body Scheme Learning Jürgen Sturm Christian Plagemann Wolfram Burgard University of Freiburg Germany
Dec 17, 2015
SA-1
Robotic Self-Perception and Body Scheme Learning
Jürgen SturmChristian PlagemannWolfram Burgard
University of FreiburgGermany
Motivation
Existing robot models are typically specified (geometrically) in advance calibrated manually
Motivation
Problems with fixed robot models: Wear-and-tear
wheel diameter, air pressure
Recovery from failure malfunctioning actuators
Tool use extending the model
Unknown modelre-configurable robots
Problems with fixed robot models: Wear-and-tear
wheel diameter, air pressure
Recovery from failure malfunctioning actuators
Tool use extending the model
Unknown modelre-configurable robots
Similar problems in humans/animals?
Motivation
Problems with fixed robot models: Wear-and-tear
wheel diameter, air pressure
Recovery from failure malfunctioning actuators
Tool use extending the model
Unknown modelre-configurable robots
Similar problems in humans/animals?
Motivation
growth, aging
injured body parts
writing
riding a bike
Related Work
Neuro-physiology Mirror neurons [Rizzolatti et al., 1996]
Body Schemes [Maravita and Iriki, 2004]
Robotics Self-calibration [Roy and Thrun, 1999]
Cross-modal maps [Yoshikawa et al., 2004]
Structure learning [Dearden and Demiris, 2005]
Problem motivation
Fixed-model approaches fail when parameters change over time geometric model is not available
Bootstrapping of the body scheme and Life-long adaptation using visual
self-observation
Our Contribution
Sense6D Poses
ActJoint angles
ThinkBootstrap, monitor, and maintaininternal representation of body
Problem Description
Problem Formulation
Visual self-perception of n body parts:
Actuators (m action signals):
Learn the mapping
p(X 1; : : : ;X n ja1; : : : ;am)
X 1; : : : ;X n 2 R4£ 4
Body pose Configuration
a1; : : : ;am 2 R
Existing Methods
Analytic model + parameter estimation
Function approximation Nearest neighbor Neural networks
Requires prior knowledge
High-dimensional learning problem
Requires large training sets
Body Scheme Factorization
Idea: Factorize the model
We represent the kinematic chain as a Bayesian network
Bootstrapping
Learning the model from scratch consists of two steps:
1. Learning the local models (conditionaldensity functions)
2. Finding the network/body structure
Learning the Local Models
Using Gaussian process regression Learn 1D 6D transformation function
for each (action, marker, marker) triple
p(¢ 12 j a1) = p(X ¡ 11 X 2 j a1)
Finding the Network Structure
Select the most likely network topology
Corresponding to the minimum spanning tree
Maximizing the data likelihoodp(M jD)
Model Selection
Model Selection
7-DOF example
Fully connected BN
Model Selection7-DOF example
Fully connected BN
Selected minimalspanning tree
Forward Kinematics
Purpose: prediction of end-effector pose in a given
configuration Approach:
integrate over the kinematicchain in the Bayesian network
by concatenating Gaussians approximate the result
efficiently by one Gaussianp(X n jX 1;a1; : : : ;am) =Z
:::Z
pM 1pM 2
: : :dX 2; : : : ;dX n¡ 1
Inverse Kinematics
Purpose: Generate motor commands for reaching a given target pose
Approach: Estimate Jacobian of end-effector using forward kinematics prediction
Use standard IK techniques Jacobian pseudo-inverse
r Xn(a) =
·@X n(a)
@a1; : : : ;
@X n(a)@am
¸
Experiments
Evaluation: Forward Kinematics
Fast convergence (approx. 10-20 iterations) High accuracy (higher than direct perception)
Evaluation: Inverse Kinematics
Accurate control using bootstrapped body scheme
Life-long Adaptation
Robot’s physical properties will change over time
Predictive accuracy of body scheme needs to be monitored continuously
Localize mismatches in the Bayesian network Re-learn parts of the network
Life-long Adaptation
Initial
Error is detected and is localized
Robot re-learns some local models
Life-long Adaptation
Evaluation
Quick localization of error Robust recovery
Summary
Novel approach learning body schemes from scratch using visual self-perception Model learning using Gaussian process
regression Model selection using data likelihood as
criterion
Efficient adaptation to changes in robot geometry
Accurate prediction and control
Future Work
Active self-exploration, optimal control, POMDPs
Marker-less self-perception
Moving robot
Tool use