Transcript Slides

Keyframe-based Learning from
Demonstration
Anthony Dubis
“Keyframe-based Learning from Demonstration –
Method and Evaluation”
– Akgun, Cakmak, Jiang, and Thomaz
What We’ll Cover
1. Learning from Demonstration & Its Types
2. The Proposed Framework
3. The Framework Results and Conclusions
Bonus: Video on extension
LEARNING FROM DEMONSTRATION
& ITS TYPES
Learning from Demonstration
• Teach a robot through successful examples
• Various options
– Teleoperation
– Motion capture
– Kinesthetic manipulation
• Paper’s focus: Kinesthetic teaching: Having a
human teacher physically guide the robot in
performing a skill
Two Kinesthetic Input Methods for
Demonstrations
• Draw Letters Using
a Mouse (2D)
• Teaching a robot:
– Scoop
– Pour
– Place
Learning from Demonstration
Introduction (Traditional)
• Learning from Demonstration (LfD)
• Continuous trajectory with two endpoints
• Trajectory Demonstration (TD) Example
Learning from Demonstration
Advantages
• Intuitiveness
• No correspondence problem
• No extra instrumentation
Learning from Demonstration
Disadvantages
• Users lack experience manipulating robots
• Noisy movements
Keyframe-based Learning from
Demonstration
• Keyframe-based Learning from Demonstration
(KLfD)
• Sparse set of consecutive poses, or critical
points
• Provide start, end, and several in-between
• Keyframe Demonstration (KD) Example
Keyframe-based Learning from
Demonstration - Advantages
• Intuitive for the user
• Pick poses with care
Keyframe-based Learning from
Demonstration - Disadvantages
• User lack of experience in manipulating robots
• Lack of timing information
• Complex and curvy movements are difficult to
express
Hybrid Learning from Demonstration
• Hybrid Learning from Demonstration (HLfD)
• Let the user choose whatever suits the
situation
• Hybrid Demonstration (HD) - Example
Demonstration Types
• Trajectory, Keyframe, or Hybrid
Demonstrations
• Convert this data into a Sequential Pose
Distribution for skill reproduction
KLfD – Proposed Framework
• Traditional LfD techniques are limited
• Goal: Create one that can take in TD, KD, HD
Implementation Overview
Can accept and process input from trajectory, keyframe, or hybrid
demonstrations.
KLfD – Framework Implementation
Overview
• Trajectory to Keyframe Conversion
• Temporal Alignment and Clustering
– Provides Sequential Pose Distribution (SPD)
• Skill Reproduction – determine parameters
Validity
• Requirements
– Handle trajectory input as well as conventional
methods
– “Lost” data
• Compare trajectory demos to baseline:
– Gaussian Mixture Model (GMM) to fit the data
– Gaussian Mixture Regression (GMR) to reproduce
the skill
– GMM + GMR
Drawing Letters
• 2D mouse gestures
• Allows TD, KD, and HD
• Skills: B, D, G, M, O, P
• Measurement: alignment
cost between generated
and goal trajectories
KFD
GMM
Validity - 2D Letters
Robot Skills
• Simon Robot
• 7 DOF arms
• 2 DoF torso
• 13 DoF head
• Scooping using TD
• Pouring using TD
• Placement using KD
Validity - Robot Skills
Validity Robot Skills
• Similar scooping
and pouring
weights
• KLfD framework
is on par with
GMM+GMR
FRAMEWORK RESULTS &
CONCLUSIONS
Results – Letter Drawing
Comparing Input Types
Letter O is all curved, trajectory is best. Letter M is straight, KFD is best
Conclusions - Advantages
• Framework seems to do its job
• Stacks up against conventional models
• Accept any of the three inputs to create
Sequential Pose Distributions (SPD)
Conclusions - Disadvantages
• Keyframe inputs -> missing velocity
parameters
• Zero velocity and acceleration assumption
Extensions
• Adding Queries by robot
• PR2