Deep Learning in robotics - Tecnalia · PDF fileDeep Learning in robotics JORNADA DEEP LEARNING: LA REVOLUCIÓN TECNOLÓGICA DE LA INTELIGENCIA ARTIFICIAL Jon Azpiazu Jon.azpiazu@

Deep Learning

in robotics

JORNADA DEEP LEARNING: LA REVOLUCIÓN TECNOLÓGICA DE LA INTELIGENCIA ARTIFICIAL

Jon Azpiazu [email protected]

ROBOTICS IN TECNALIA

INDUSTRY AND TRANSPORT Division HEALTH Division

ROBOT as a PRODUCT ROBOT as a Tool to automate process

ROBOT Autonomy as a key to Flexibility

Deep Learning in Robotics

3 ▌

Deep Learning in Robotics

Deep Learning in Robotics

ImageNet Error Rate 2010-2014

Source: ClarifAI

Deep Learning in Robotics

Source: ClarifAI

Deep Learning in Robotics

ImageNet Error Rate 2010-2014

Pixels Hand-engineered Interest Points (SIFT, SURF, Fast, …)

Matching Distance Classifier (SVM)

Label “cat” “screw” …

Audio Acoustic Model Phonetic Model Language Model

Hand-engineered Descriptors (SIFT, SURF, HOG, …)

Text

Deep Learning in Robotics

Speech Recognition Computer Vision

Pixels

Label “cat” “screw” …

Audio Deep Neural Network Text

Deep Learning in Robotics

Speech Recognition Computer Vision

Deep Neural Network

Pixels

Label “cat” “screw” …

Audio Deep Neural Network Text

Deep Learning in Robotics

Speech Recognition Computer Vision Robotics

Deep Neural Network

Sensors Perception State estimation

Planning Control Motor commands

Pixels

Label “cat” “screw” …

Audio Deep Neural Network Text

Deep Learning in Robotics

Speech Recognition Computer Vision Robotics

Deep Neural Network

Sensors Motor commands

Deep Neural Network

Deep Learning in Robotics

Goal

Reward Actions

Observations

Environment Agent

Additional challenges: • Stability • Credit assignment • Exploration

Deep Learning in Robotics

Deep Learning in Robotics

• End-to-end learning of values Q(s,a) from pixels s • Input state s is stack of raw pixels from last 4 frames • Output is Q(s,a) for 18 joystick/button positions • Reward is change in score for that step

Deep Learning in Robotics

Deep Learning in Robotics

Objective: full Autopilot by 2018 • 780 million miles in 18 months

(2014) • 1 million miles every 10 hours • Google: 1.5 million miles (2009)

• Simulation: 3 million miles a day

Deep Learning in Robotics

• Robot Bin Picking with Deep Learning

• Learning Contact-Rich Manipulation Skills with Guided Policy Search

• Learning Hand-Eye Coordination for

Robotic Grasping with Deep Learning

Deep Learning in Robotics

Open challenges: • Data • Transfer learning / Shared learning • Memory • Decision making (goals)

References

Pieter Abbeel (UC Berkeley) - DL for Robotics ["DeepLearning in Robotics", RSS 2016] https://www.youtube.com/watch?v=mT7HMTTCI1k Mnih, Volodymyr, et al. "Playing atari with deep reinforcement learning." arXiv preprint arXiv:1312.5602 (2013). Raia Hadsell, “Deep Learning for Robots”, European Robotics Forum 2017 (ERF2017) Levine, Sergey, et al. "Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection." arXiv preprint arXiv:1603.02199 (2016). Sergey Levine, “Deep Robotic Learning”, 4th International Conference on Learning Representations (ICLR 2017)

Visit our blog:

http://blogs.tecnalia.com/inspiring-blog/

www.tecnalia.com