Learning and Generalization of BehaviorGrounded Tool Affordances

1

Learning and Generalization of Behavior-Grounded
Tool Affordances
Jivko Sinapov and Alexander Stoytchev, Department
of Computer Science, Iowa State University
1. Tool Use in Animals
6. Action and Change Detection
8. Evaluation on a Novel Tool
4. Problem Statement and Notation
The robots action - grab the tool and slide the
gripper by x and y in the horizontal plane
The robot is trained on one tool but evaluated on
a different one.
Let the robots action at time t be defined as
Monkeys use tools in both natural and controlled
environments.
where
are the actions parameters
The sensory input at time t is a vector
Decision Tree model with a puck-centered frame of
reference is able to predict the affordances of
the new tool quite well.
Povinelli et al. (2000) conclude that chimpanzees
infer simple rules from their experience
regarding tool use, e.g. visual contact leads to
movement.
The robot can extract features from its sensory
input with a set of perceptual functions
such that
Ishibashi et al. (2000)
The robot explores the tool through
behavior-babbling, during which it stores data
points of the form
where
and k ltlt n
Generalization across tools is best achieved when
the novel tool shares local features with the
familiar tool.
Ishibashi et al. (2000) show that monkeys can
generalize tool-related knowledge from one tool
to a novel tool. Macaques were able to use novel
tools provided that they shared similar features
with familiar tools.
The robot observes changes over time with a
change detection function
8. Evaluation on a Larger Tool
The task of the robot is to learn a model,
, such that
The robot is trained on the L-hook tool, but
tested on a larger version of the tool.
2. Previous Work
Behavior-grounded representation of tool
affordances introduced by Stoytchev (2005)
Decision tree with puck-centric frame of
reference achieves the highest performance 83.8
good predictions.
5. Experimental Setup
The Robot is a 6-DOF arm, simulated using the
BREVE simulator.
The tool representation is grounded in robots
behavioral and perceptual repertoire, consistent
with the Verification Principle (Sutton 2001).
600 trials with each tool are performed. During
each trial the starting positions of the tool and
the puck, as well as the actions parameters are
randomly chosen.
The robot predicts the displacement of a small
cylindrical puck as a result of action with the
tool.
The model
Prediction errors are distributed around the
corners of both tools.
is trained and
evaluated by performing 3-fold cross validation
on the recorded data points.
7. Evaluation on a Familiar Tool
8. Conclusion and Future Work
6 tools T-Stick, L-Stick, L-Hook, Stick, T-Hook,
Paddle
The robot is trained and tested on the same tool
A compact decision tree model is capable of
representing the behavior-grounded tool
affordances introduced by Stoytchev (2005).
Stoytchev (2005)
Model Limitations
Two Learning Algorithms
The learned affordances are kept in a look-up
table, difficult to predict consequences of new
actions with the tool if the data is not already
included.
Decision Tree
Decision tree model has several advantages over
k-NN
k-Nearest Neighbor
It is compact and does not require storing all
data to make predictions.
Knowledge from experience with one tool cannot be
applied to a novel tool.
Sensory input is extracted from the camera
overlooking the robot. Each pixel labeled as
either the tool, the puck, the gripper, the arm
or the background.
It shows better generalization in new situations
such as novel tools, or scaled versions of
familiar tools.
This work presents a framework for overcoming
these limitations.
Decision tree model with puck-centric frame of
reference makes errors only when the pucks
starting position is near a corner of the tool,
i.e., the errors are quantifiable.
Five frames of reference camera image center,
the gripper, tool marker 1, tool marker 2 and
the puck.
3. Goals
Represent the affordances of a tool in a compact
predictive model.
Future Work
The five frames of are associated with five
perceptual functions
Model construction and selection how to
construct models for novel tools based on past
experience.
Study how models can generalize across novel
tools to which the robot has not been previously
exposed.
Active Learning of Affordances of Multiple Tools.
Identify frames of reference in the robots
visual input field that are useful for prediction.
Integrate models for tool affordances within a
planning framework in order to solve tool-using
tasks.