A.Levenchuk -- visuomotor learning in cyber-phisical systems
Cyber-physical systems architecture breakthrough:
learning of visuomotor coordination
107th meeting of INCOSE Russian chapter
• Cyber-Physical Systems or “smart” systems are co-
engineered interacting networks of physical and
• CPS include:
- Internet of Things (IoT)
- Industrial Internet
- Smart Cities
- Smart Grid
- "Smart" Anything (e.g., Cars, Buildings, Homes,
Manufacturing, Hospitals, Appliances)
• NIST CPS Public Working Group --
• NSF -- http://cps-vo.org/ 2
Draft Framework for Cyber-Physical Systems
This is all about
From «smart» to «intelligent»
How CPS perform it Decision?
Where is that «intelligence»?
Sensors Mechanics Actuators
• Decision is carefully programmed
• Example: robot-«butterfly»,
• Every type of movement should be programmed anew
• Non-adaptable to changes of environment and device
• The best science available up today!
• Perfect, if CPS perform only one or two movements. Not for
Goal: CPS capuchin-like
• Jurgen Schmithuber (July 2015): In order to pick a fruit at
the top of a tree, Capuchin monkey plans a sequence of
sub-goals (e.g., walk to the tree, climb the tree, grab the
fruit, …) effortlessly. We will have machines with animal-
level intelligence in 10 years.
• Needs planning
• Needs great visuomotor coordination!
• Impossible to program manually up to date.
New in «decisions»:
learn to decide!
Machine learning and reasoning:
• Symbolic (by induction)
• Evolutionary (by genetic programming)
• Bayes (by probability assesement)
• By analogy
• Connectivist (deep learning, artificial neuron nets)
The Master Algorithm: combine ‘em all!
Breakthrough: deep learning
Reinforcement learning + deep learning
Not only «visuo» but «motor» too – with a coordination!
Not a rocket science
• Open science
• ArXiv, GitXiv
• Open Source libraries
• GPU in all computer stores
• Conferences: ICML 2014 – 2500 participants, ICML
2015 – 4000 participants
• Multiple schools (summer schools, university
• Competitions (e.g. Kaggle)
Motor Visuo Motor
mediated perception behavior reflex
And everything in between!
• train a deep Convolutional Neural Network (CNN) using 12 hours of
human driving in a video game
• show that our model can work well to drive a car in a very diverse set of
• train another CNN for car distance estimation on the KITTI dataset,
results show that the direct perception approach can generalize well to
real driving images
• Open sourced
• Autopilot Driving is not a miracle now: Tesla X, Google car,
AVRORA/KAMAZ, Volvo trucks, and counting
• Body is an easy part.
• Manipulation is difficult!
• Non-prehensile manipulation is included.
• End-to-End Training of Deep Visuomotor Policies,
• Supersizing Self-supervision: Learning to Grasp
from 50K Tries and 700 Robot Hours,
• Fanuc: $7.5mln for 6%
in Preferred Networks
• ABB invested up to
$10mln in Vicarious
Osaro -- http://www.osaro.com/, learning
in environment, including cooperation
• Deep Spatial Autoencoders for Visuomotor
Learning to plan
Neurocognitive Architecture for Autonomous Task Recognition, Learning
and Execution (NARLE)
Quantum computer: waiting, but promising – currently 100mln times faster
than classical desktop (http://googleresearch.blogspot.com/2015/12/when-can-quantum-annealing-
• Computational multi-lens optics (near future)
• Solid state LIDARs for driving (below $100 in five
years, now $1000) -- http://www.quanergy.com/,
http://velodynelidar.com/, includes processor and
• But Elon Musk tell that lidar not needed, only
optical cameras and radar
Voice interface for visuomotor goal settings
• Deep learning is a leading technology for a voice recognition
• Voice command interface is not a problem today
• General intelligence of a CPS is a problem!
Company Name of Personal assistant
INCOSE Russian chapter, research director
http://ailev.ru (in Russian)
INCOSE Russian chapter