Bridging the Gap: From Fundamental AI Research to Real-World … · 2017. 10. 27. · Bridging the...

Bridging the Gap:

From Fundamental AI Research to Real-World Challenges

Talk ID: 23355

Yasser Jadidi, Global Head of AI Research

Bosch Center for Artificial Intelligence

GPU Technology Conference | Munich, Oct., 10-12, 2017

GPU Technology Conference 2017 | BOSCH

CR/PJ-AI-R | 11.10.2017

© Robert Bosch GmbH 2017. All rights reserved, also regarding any disposal, exploitation, reproduction, editing, distribution, as well as in the event of applications for industrial property rights.2

Strong Worldwide Cross-Domain Reach

Automotive

PowertrainAutonomous

Driving Perception

Industry 4.0

Logistics

Bosch

Center

for AI


CR/PJ-AI-R | 11.10.2017


Welcome to the Real World

Automotive Powertrain System

Hysteresis Effects

Strong Nonlinearities

Real-time online execution

Limited Computational Resources

Dirty DataDynamic System Behavior


CR/PJ-AI-R | 11.10.2017


AI-driven Powertrain Solution

I. Safe Active Learning

III. Hardware Acceleration

II. (Transient) Gaussian Processes

Sparsification

Numerical Stability

file://bosch.com/dfsrb/DfsDE/DIV/CR/AE2/Projects/AE2-033/40_Work_Packages/ODCM/20140919_OnlineDoE_Inbetriebnahme.avifile://bosch.com/dfsrb/DfsDE/DIV/CR/AE2/Projects/AE2-033/40_Work_Packages/ODCM/20140919_OnlineDoE_Inbetriebnahme.avi


CR/PJ-AI-R | 11.10.2017



CR/PJ-AI-R | 11.10.2017


Introducing Control Theory Standards into Reinforcement Learning

Policies

T3-model simplified

≤Threshold

λ map

1.0λ

T3 map

T

E IV I

x

T3 curve

2-point controler

A. Dörr, D. Nguyen-Tuong, M. Alonso, S. Schall, S. Trimpe, “Model-Based Policy Search for Automatic Tuning of

Multivariate PID Controllers”, ICRA, 2017.

Exemplary visualization of

stability regionIndustry standard:

hand-tuned two-level

controller based on 2D

control maps

Bosch AI solution:

RL learned controller

based on GP

dynamics model

Engine exhaust gas

temperature control

Stabile Reinforcement

Learning

Robust control learning

computation of regions of

guaranteed stability,

learning controllers while

ensuring stability

J.Vinogradska, B. Bischoff, D. Nguyen-Tuong, J. Peters, “Stability of Controllers for Gaussian Process Dynamics”,

JMLR, 2017.

J. Vinogradska, J., Bischoff, B., Nguyen-Tuong, D., Romer, A., Schmidt, H., Peters, J., “Stability of Controllers for

Gaussian Process Forward Models”, ICML, 2016.


CR/PJ-AI-R | 11.10.2017


Autonomous Agent Decision Making under Constraints

P. Schillinger, M. Bürger, D. V. Dimarogonas, “Simultaneous Task Allocation and

Planning for Temporal Logic Goals in Heterogeneous Multi-Robot Systems”, SAGE, 2016.

Linear Transient Logic (LTL)

Policy Synthesis

Goals & constraintsFormulation as Directed Graph +

Decomposition for Multiple Agents

𝓜𝟏 = ◊𝒊∈ 𝟏,𝟐,𝟑,𝟒 𝒉𝒊 ∧𝒄∧○¬𝒄

Deliver drinks to rooms h1…h4.

Consider limited drinks and battery consumption.

Service Robotics Application


CR/PJ-AI-R | 11.10.2017


The Marginal Difference between Cats and Modems

Autonomous Driving Perception:

Adversarial Threats

Original Picture

„Persian Cat“

Disturbed Picture

„Modem“

Perturbation

10x amplified

J. H. Metzen, K. C. Mummadi, T. Brox, V. Fischer, “Universal Adversarial Perturbations Against Semantic Image Segmentation”,

The IEEE International Conference on Computer Vision (ICCV), 2017.

Bosch Center for Artificial IntelligenceHQ at Bosch Research Campus Renningen, Germany

Bridging the Gap: From Fundamental AI Research to Real-World … · 2017. 10. 27. · Bridging the...

Documents

Transcript of Bridging the Gap: From Fundamental AI Research to Real-World … · 2017. 10. 27. · Bridging the...