CSCI598A: Robot Intelligence. Apr. 23, 2015

Transcription

1 CSCI598A: Robot Intelligence Apr. 23, 2015

2 Reasoning Over Time Object recognition (static problem) We consider spatial relations with uncertainty We don t care about time Motion planning (dynamic problem) It is a dynamic problem with uncertainty Variable values change over time Locations, velocity, acceleration of joints Time must be modeled to estimate present status and probably predict future states High-Level Task Abstraction (static problem) We consider temporal relations of subtasks 2

3 Issues in Time Reasoning Temporal segmentation of streaming/times series data Alignment of time series data Reasoning high-level task abstraction

4 Temporal Segmentation A naïve, uniform segmentation

5 Temporal Segmentation A naïve, uniform segmentation Right: the standard deviation of the scores and its mean computed on a sliding window. The local minima of the standard deviation function are break points.

6 Temporal Segmentation A naïve, uniform segmentation After normalization: Blue dots are the break points computed that indicate the end of a segmentation and the beginning of a new one.

7 Temporal Segmentation Fuzzy segmentation Ground truth The uniform segmentation and many others assume break points and segments can be distinctly separated.

8 Temporal Segmentation Fuzzy segmentation Ground truth Frames Key Concept: Gradual Transition Write on Board Gradual Transition Answer Phone

9 Temporal Segmentation Fuzzy segmentation Ground truth Proposed approach Frames The fuzzy approach models each segment/event as a fuzzy set with fuzzy boundaries.

10 Alignment of Time Series Dynamic Time Warping (DTW)

11 Reasoning Task Abstraction Goal: reason about chronological order of subtasks

12 Reasoning Task Abstraction Goal: reason about chronological order of subtasks

13 Reasoning Task Abstraction

14 Reasoning Task Abstraction

15 Reasoning Task Abstraction

16 Reasoning Task Abstraction

17 Reasoning Task Abstraction Inference tasks:

18 Reasoning Task Abstraction Hidden Markov Model:

19 Reasoning Task Abstraction Hidden Conditional Random Field (HCRF): Learn a mapping from temporal data to a label Use latent variable to model underlying temporal structures y h 1 h 2 h 3 h 4 h 5 x 1 x 2 x 3 x 4 x 5 Frame 1 Frame 2 Frame 3 Frame 4 Frame 5

20 Reasoning Task Abstraction Example of HCRF y = tennis-serve h {Toss, Swing, Hit} Assuming identical h x HCRFs h: T T T S S S S S H H H H

21 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

22 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

23 Summary of the class A working definition of robot: Physical machine that generates intelligent connection between perception and action Robot intelligence: Robot intelligence includes recognizing patterns, comprehending ideas, plan, making decisions, and communicating

24 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

25 Summary of the class Robot perception Face examples Classification Result Off-line training Classifier Feature Extraction Representation Non-face examples Search for faces at different resolutions and locations 25

26 Summary of the class Bag of word models 1. Feature detection 2. Feature description 3. Dictionary learning 4. Bag-of-features representation

27 Summary of the class 3D object recognition

28 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

29 Summary of the class Learning from Demonstration Learning by watching: correspondence problem Learning by acting Gaussian mixture models and regressions

30 Summary of the class Learning from Demonstration Learning by watching: correspondence problem Learning by acting Gaussian mixture models and regressions Key issues in Learning from Demonstration Parameter learning: Expectation-Maximization Gaussian component estimation: Bayesian Information Criteria (BIC) Trajectory alignment: Dynamic Time Warping (DTW) Dimension reduction: Principal Component Analysis (PCA)

31 Summary of the class Reinforcement learning A learning approach that can adapt through interaction with the environment

32 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

33 Summary of the class Learning from Data Supervised learning Unsupervised learning K-means

34 Summary of the class Time modeling Temporal segmentation Sequence alignment Reasoning time orders of subtasks

35 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

36 Summary of the class Deep learning

37 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

38 Summary of the class Amazon Picking Challenge using the Baxter robot (named Zuko, the firelord)

39 Summary of the class Definition of robot and its intelligence Robot perception (perception) Sensing technologies Object recognition from 2D and 3D Learning from demonstration (action) Reinforcement learning Data and time modeling (reasoning) Tutorial of ROS, PCL, and deep learning A focus on the Amazon Picking Challenge

40 Examples of LfD and RL Work from Dr. Aude Billard 40

41 Examples of LfD and RL Work from Dr. Aude Billard 41

42 Examples of LfD and RL Work from Dr. Aude Billard 42

43 Additional Examples Work from Dr. Aude Billard 43