File preview
Active Safety Control in A ti S f t C t l i Automotive Cyber‐Physical Systems Automotive Cyber‐
PI: Francesco Borrelli PI: Francesco Borrelli
Email: fborrelli@me.berkeley.edu Department of Mechanical Engineering University of California y f f Berkeley, USA www.mpc.berkeley.edu
Co‐PI Karl Hedrick, Co‐PI Karl Hedrick Ruzena Bajcsy PI Karl Hedrick, Ruzena
Automotive Cyber‐Physical System Automotive Cyber‐
Vehicle Actuators
Vehicle and Tire Sensor Data
Intelligence I t lli
Driver Model/Intent
Safety
Comfort
Efficiency
Lane Departure A14 Highway – Lane Departure A14 Highway – June 2009
www.mpc.berkeley.edu
2008 US Statistics
www.mpc.berkeley.edu
Automotive Cyber‐Physical System Automotive Cyber‐
Vehicle Actuators
Vehicle and Tire Sensor Data
Intelligence I t lli
Driver Model/Intent
Safety
Comfort
Efficiency
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
Human‐Vehicle Interaction
Steering Angle
Actuation System
Wheels Traction Torque Wheels Braking Torques ……..
Human‐Vehicle Interaction
Steering Angle
Actuation System
Wheels Traction Torque Wheels Braking Torques ……..
Hydraulic Brake Unit
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
Vehicle‐ Vehicle‐Road Interaction FEM Simulation
www.mpc.berkeley.edu
Vehicle‐ Vehicle‐Road Interaction Simplified Models
Longitudinal Force Dry Asphalt
Maximum Braking B ki
Maximum Acceleration A l i Steer Angle
Tire Slip Tire Slip
Longitudinal Force
Lateral Force
www.mpc.berkeley.edu
Vehicle‐ Vehicle‐Road Interaction Simplified Models
Longitudinal Force Dry Asphalt S o Snow
Maximum Braking B ki
Maximum Acceleration A l i Steer Angle
Tire Slip Tire Slip
Longitudinal Force
Lateral Force
www.mpc.berkeley.edu
Vehicle‐ Vehicle‐Road Interaction p Simplified Models
Longitudinal Force Dry Asphalt S o Snow Ice
Maximum Braking B ki Maximum Acceleration A l i Steer Angle
Tire Slip Tire Slip
Longitudinal Force
Lateral Force
www.mpc.berkeley.edu
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
Environment‐ Environment‐Human Interaction
Sensory Inputs
Brain
Spinal Cord
Muscles
http://www.nature.com/nrn/journal/v5/
CPS‐ CPS‐Synoptic Scheme Human
Vehicle Model Environment E i t Model
Vehicle
Environment
``We know that a lot of the brain has an internal neural simulator”… We know that a lot of the brain has an internal neural simulator “to anticipate or predict the future for a given a input” Eric Kandel (Charlie Rose interview, 2008)
CPS‐ CPS‐Synoptic Scheme Human
Vehicle Model Environment E i t Model
Vehicle
Environment
CPS with Driver Assistance System (DAS) Human
Vehicle Model Environment E i t Model
Vehicle DAS
Environment
Vast Majority of DAS systems Human
Vehicle Model Environment E i t Model
Vehicle DAS
Environment
AntiAnti-lock Braking Systems
Longitudinal Force
Maximum Braking g
Maximum Acceleration Steer Angle
Long. Slip g p
Longitudinal Force
Lateral Force
CounterCounter-Steering and Over-Steering Over-
Vast Majority of DAS systems Human
Vehicle Model Environment E i t Model
Vehicle DAS
Environment
Vehicle CPS Main Issues
• • • • Complexity/Compositionality… y y “Shy” Autonomy No Guarantees/Heuristic Tunings u a o o e cep o / og o g o ed Human MotionPerception/Cognition Ignored
CPS Main Issues
• • • • Complexity/Compositionality y y “Shy” Autonomy No Guarantees/Heuristic Tunings u a o o /pe cep o /cog o g o ed Human Motion/perception/cognition ignored
Research Goals
Provably Safe and Adaptive Autonomy
Quantifying Uncertainties (Human and Environment)
Quantifying Uncertainties: Tire/Road Interaction Tire/Road Interaction
CYBER™TYRE
Quantifying Uncertainties: Tire/Road Interaction Tire/Road Interaction
Tangential and Lateral Sidewall Tangential and Lateral Sidewall Tange ent
Lateral
Quantifying Uncertainties: Driver Models
Objective: Exploit real‐time driver state (joint angles) detection to extract compatible set of trajectories
Methodology: 1‐Stereo cameras for 3D reconstruction of the scene 2‐ Articulated tracking, use body sensor network for validation
Real‐ Real‐Time 3D Reconstruction and the Middlebury Stereovision Dataset Middlebury Stereovision Dataset
Real‐ Real‐Time 3d Reconstruction Driver
Provably Safe Adaptive Autonomy
Vehicle Actuators Vehicle and Tire Sensor Data
Adaptive and Predictive p Autonomy
Driver Intent
www.mpc.berkeley.edu
Set‐ Set‐Based Control Design
• Compute N‐steps controllable set Compute N steps controllable set
…
• Given compute such that
…
Set‐ Set‐Based Control Design
• Compute N‐steps controllable set Compute N steps controllable set
…
Benefit in Autonomy Concept
At time j • Driver intent • Among all possible actuations choose the one that solves
Adaptive and Predictive Autonomy p y
www.mpc.berkeley.edu
Main Limitation: Real‐Time Computation Main Limitation: Real‐ ``Predictive Control”: Borrelli, Bemporad Morari
• PWA Model
www.mpc.berkeley.edu www mpc berkeley edu
• One‐step robust controllable set
‐For linear (A,B) system: ‐For linear systems, result is polytope For linear systems result is polytope ‐For PWA systems, result is union of polytopes
Process Industry 4Ghz, 4Ghz 1 Terabyte
Automotive 50Mhz, Mbytes 50Mh 2 Mb t
Safe Control Design Through Simplified Models Simplified Models
“True” System Simplified System
……
XN −2 ZN−2
XN −1 ZN −1
: k‐step controllable set for “True” System when control law p y designed for simplified system is applied.
2D Example
States: Inputs: Disturbance: Assume constant: A
y ,
M
f ,Vx
PWA approximation of Pacejka tire model PWA approximation of Pacejka tire model
2D Example – Robust Set Computation 2D Example –
0.6 0.4 0.2
0 x2 -0.2 -0.4 -0.6 -0.8 -2
0.6 0.4 0.2 0 x2 -0.2 -0.4 -0.6 -0.8 270 272 274 276 x1 278
-1.5
-1
-0.5
0 x1
0.5
1
1.5
2
280
282
From 2‐D to 12‐ From 2‐D to 12‐D Example Experimental results @ 72 Kph on Ice Experimental results @ 72 Kph on Ice Experimental results @ 72
Conclusions/Outlook
• Developing concept and methods for Provably Safe Adaptive Autonomy – Quantifying uncertainties in • Vehicle/Road Interaction • Human/Car Interaction – Real‐time Computation of Controllable Sets p with different level of granularity
The Team
UC Berkeley
• • • • • • • • • Gurkan Erdogan Ramanarayan Vasudevan Ricardo Cervera Ricardo Cervera Navarro Sanghyun Hong Theresa Lin Ye Zhuang Yiqi Gao Ruzena Bajscy Karl Hedrick
Ford Research Labs (Dearborn,USA) • Jahan Asgari, Eric Tseng, Davor Hrovat Pirelli Research Labs (Milano,Italy) ( , y) • Federico Mancosu, Giorgio Audisio
PI: Francesco Borrelli PI: Francesco Borrelli
Email: fborrelli@me.berkeley.edu Department of Mechanical Engineering University of California y f f Berkeley, USA www.mpc.berkeley.edu
Co‐PI Karl Hedrick, Co‐PI Karl Hedrick Ruzena Bajcsy PI Karl Hedrick, Ruzena
Automotive Cyber‐Physical System Automotive Cyber‐
Vehicle Actuators
Vehicle and Tire Sensor Data
Intelligence I t lli
Driver Model/Intent
Safety
Comfort
Efficiency
Lane Departure A14 Highway – Lane Departure A14 Highway – June 2009
www.mpc.berkeley.edu
2008 US Statistics
www.mpc.berkeley.edu
Automotive Cyber‐Physical System Automotive Cyber‐
Vehicle Actuators
Vehicle and Tire Sensor Data
Intelligence I t lli
Driver Model/Intent
Safety
Comfort
Efficiency
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
Human‐Vehicle Interaction
Steering Angle
Actuation System
Wheels Traction Torque Wheels Braking Torques ……..
Human‐Vehicle Interaction
Steering Angle
Actuation System
Wheels Traction Torque Wheels Braking Torques ……..
Hydraulic Brake Unit
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
Vehicle‐ Vehicle‐Road Interaction FEM Simulation
www.mpc.berkeley.edu
Vehicle‐ Vehicle‐Road Interaction Simplified Models
Longitudinal Force Dry Asphalt
Maximum Braking B ki
Maximum Acceleration A l i Steer Angle
Tire Slip Tire Slip
Longitudinal Force
Lateral Force
www.mpc.berkeley.edu
Vehicle‐ Vehicle‐Road Interaction Simplified Models
Longitudinal Force Dry Asphalt S o Snow
Maximum Braking B ki
Maximum Acceleration A l i Steer Angle
Tire Slip Tire Slip
Longitudinal Force
Lateral Force
www.mpc.berkeley.edu
Vehicle‐ Vehicle‐Road Interaction p Simplified Models
Longitudinal Force Dry Asphalt S o Snow Ice
Maximum Braking B ki Maximum Acceleration A l i Steer Angle
Tire Slip Tire Slip
Longitudinal Force
Lateral Force
www.mpc.berkeley.edu
CPS‐ CPS‐Synoptic Scheme Human Vehicle
Environment
Environment‐ Environment‐Human Interaction
Sensory Inputs
Brain
Spinal Cord
Muscles
http://www.nature.com/nrn/journal/v5/
CPS‐ CPS‐Synoptic Scheme Human
Vehicle Model Environment E i t Model
Vehicle
Environment
``We know that a lot of the brain has an internal neural simulator”… We know that a lot of the brain has an internal neural simulator “to anticipate or predict the future for a given a input” Eric Kandel (Charlie Rose interview, 2008)
CPS‐ CPS‐Synoptic Scheme Human
Vehicle Model Environment E i t Model
Vehicle
Environment
CPS with Driver Assistance System (DAS) Human
Vehicle Model Environment E i t Model
Vehicle DAS
Environment
Vast Majority of DAS systems Human
Vehicle Model Environment E i t Model
Vehicle DAS
Environment
AntiAnti-lock Braking Systems
Longitudinal Force
Maximum Braking g
Maximum Acceleration Steer Angle
Long. Slip g p
Longitudinal Force
Lateral Force
CounterCounter-Steering and Over-Steering Over-
Vast Majority of DAS systems Human
Vehicle Model Environment E i t Model
Vehicle DAS
Environment
Vehicle CPS Main Issues
• • • • Complexity/Compositionality… y y “Shy” Autonomy No Guarantees/Heuristic Tunings u a o o e cep o / og o g o ed Human MotionPerception/Cognition Ignored
CPS Main Issues
• • • • Complexity/Compositionality y y “Shy” Autonomy No Guarantees/Heuristic Tunings u a o o /pe cep o /cog o g o ed Human Motion/perception/cognition ignored
Research Goals
Provably Safe and Adaptive Autonomy
Quantifying Uncertainties (Human and Environment)
Quantifying Uncertainties: Tire/Road Interaction Tire/Road Interaction
CYBER™TYRE
Quantifying Uncertainties: Tire/Road Interaction Tire/Road Interaction
Tangential and Lateral Sidewall Tangential and Lateral Sidewall Tange ent
Lateral
Quantifying Uncertainties: Driver Models
Objective: Exploit real‐time driver state (joint angles) detection to extract compatible set of trajectories
Methodology: 1‐Stereo cameras for 3D reconstruction of the scene 2‐ Articulated tracking, use body sensor network for validation
Real‐ Real‐Time 3D Reconstruction and the Middlebury Stereovision Dataset Middlebury Stereovision Dataset
Real‐ Real‐Time 3d Reconstruction Driver
Provably Safe Adaptive Autonomy
Vehicle Actuators Vehicle and Tire Sensor Data
Adaptive and Predictive p Autonomy
Driver Intent
www.mpc.berkeley.edu
Set‐ Set‐Based Control Design
• Compute N‐steps controllable set Compute N steps controllable set
…
• Given compute such that
…
Set‐ Set‐Based Control Design
• Compute N‐steps controllable set Compute N steps controllable set
…
Benefit in Autonomy Concept
At time j • Driver intent • Among all possible actuations choose the one that solves
Adaptive and Predictive Autonomy p y
www.mpc.berkeley.edu
Main Limitation: Real‐Time Computation Main Limitation: Real‐ ``Predictive Control”: Borrelli, Bemporad Morari
• PWA Model
www.mpc.berkeley.edu www mpc berkeley edu
• One‐step robust controllable set
‐For linear (A,B) system: ‐For linear systems, result is polytope For linear systems result is polytope ‐For PWA systems, result is union of polytopes
Process Industry 4Ghz, 4Ghz 1 Terabyte
Automotive 50Mhz, Mbytes 50Mh 2 Mb t
Safe Control Design Through Simplified Models Simplified Models
“True” System Simplified System
……
XN −2 ZN−2
XN −1 ZN −1
: k‐step controllable set for “True” System when control law p y designed for simplified system is applied.
2D Example
States: Inputs: Disturbance: Assume constant: A
y ,
M
f ,Vx
PWA approximation of Pacejka tire model PWA approximation of Pacejka tire model
2D Example – Robust Set Computation 2D Example –
0.6 0.4 0.2
0 x2 -0.2 -0.4 -0.6 -0.8 -2
0.6 0.4 0.2 0 x2 -0.2 -0.4 -0.6 -0.8 270 272 274 276 x1 278
-1.5
-1
-0.5
0 x1
0.5
1
1.5
2
280
282
From 2‐D to 12‐ From 2‐D to 12‐D Example Experimental results @ 72 Kph on Ice Experimental results @ 72 Kph on Ice Experimental results @ 72
Conclusions/Outlook
• Developing concept and methods for Provably Safe Adaptive Autonomy – Quantifying uncertainties in • Vehicle/Road Interaction • Human/Car Interaction – Real‐time Computation of Controllable Sets p with different level of granularity
The Team
UC Berkeley
• • • • • • • • • Gurkan Erdogan Ramanarayan Vasudevan Ricardo Cervera Ricardo Cervera Navarro Sanghyun Hong Theresa Lin Ye Zhuang Yiqi Gao Ruzena Bajscy Karl Hedrick
Ford Research Labs (Dearborn,USA) • Jahan Asgari, Eric Tseng, Davor Hrovat Pirelli Research Labs (Milano,Italy) ( , y) • Federico Mancosu, Giorgio Audisio