Robotics and Autonomy. Control of Complex Systems (RMM)

Transcription

1 Robotics and Autonomy Richard M. Murray BE 107, 14 May 2015 Goals: Describe how behavior is implemented in robotic systems (vs biology) Discuss some of the ways that insights in robotics might impact biology and vice versa Reading: Sensing, Navigation and Reasoning Technologies for the DARPA Urban Challenge, J. W. Burdick, N. dutoit, A. Howard, C. Looman, J. Ma, R. M. Murray, T. Wongpiromsarn, DARPA Urban Challenge Final Report, A bio-plausible design for visual attitude stabilization, A. Censi, S. Han, S. B. Fuller, R. M Murray. Conf. on Decision and Control (CDC), 2009 Control of Complex Systems (RMM) Gwenyth Card (Caltech), Apr 06 Sawyer Fuller (Caltech), Nov 08 Alice 30 P4 2+ GHz 30G mem, 3 Gb/s net 8 LADAR, 10 cameras 200k+ lines of code 50 engineers, 1.5 years 600 miles of self-driving Dominant challenges: (design for) verification and robustness EXCAPE, 13 Jun 2013 Drosophila 300k Hz ~ sensors Most neurons are dedicated to sensor processing Architecture (?) Dominant challenges: decoding organization and architecture Neutrophil DNA: 3G base pairs 20k genes, ~3kb each Timescales: msecs (binding) to minutes (expr) to hours (div n) Architecture (?) Dominant challenges: (lack of) modularity, stochastic program g 2

2 Design of Modern (Networked) Control Systems! Decision-Making (mode, contingency and constraint management) Networking and Communications Cloud Resources Operators Other Subsystems State! Estimation (MHE, PF) Sensor Processing (KF) Online System Model (sys + env) design cycles Online! Optimization (MPC, RHC) Feedback! Control (PID) Layers of Abstraction! Physical Sensors Actuators Control = System Sensing Actuation External Environment Computation Networking (layering, synchronization, redundancy, scale) UTC IASE, 10 Nov Motivating Example: Alice ( ) Alice 300+ miles of fully autonomous driving 8 cameras, 8 LADAR, 2 RADAR 12 Core 2 Duo CPUs + Quad Core 3 Gb/s data network ~75 person team over 18 months (x 2) Software 25 programs with ~200 exec threads 237,467 lines of executable code IFAC World Congress

3 DGC07 System Architecture Logging/ Visualization Process Manager Simulation Health Manager Systems LADAR (6) Feature Classificat n Stereo/Road Finding Elevation Mapping Gimbaled Sensor Obstacle Detect/Track Mission Linux, TCP/IP,... World Map Traffic Obstacle Map Vehicles Follower Actuation Interface Vehicle State Estimator Vehicle Sensing Application of existing controls technology in Alice Receding horizon (optimization-based) control for path planning with obstacles; ~100 msec iteration rate Multi-layer sensor fusion: sensor bus allows different combinations of sensors to be used for perceptors + fusion at map level Low level (inner loop) controls: PID w/ anti-windup (but based on a feasible trajectory from RHC controller) Team Caltech, Apr 07 Navigation Properties Highly modular Rapidly adaptable Constantly viable Robust??? 5 Sensing System Sensing hardware 6 horizontal LADAR (overlapping) 1 pushbroom LADAR; 1 sweeping (PTU) 3 stereo pairs (color; ~10 Hz) 2 road finding cameras (B&W) 2 RADAR units (PTU mounted) 10 blade cpci high speed computing Team Caltech, Apr 07 6

4 Feeder Perceptors Mapper Team Caltech, Jan 08 7 Planning Hourglass Protocol stack based architecture s uses directives/responses to communicate Each layer is isolated from the ones above and below Had 4 different path planners under development, two different traffic planners. Engineering principle: layered protocols isolate interactions Define each layer to have a specific purpose; don t rely on knowledge of lower level details Important to pass information back and forth through the layers; a fairly in an actuator just generate a change in the path (and perhaps the mission) Higher layers (not shown) monitor health and can act as hormones (affecting multiple subsystems) Hybrid system control methodology Finite state automata control interactions between layers and mode switches (intersection, off road, etc) Formal methods for analysis of control protocol correctness (post race) Mission Traffic Follower Actuation Interface Vehicle - Eg: make sure that you never have a situation where two layers are in conflict Team Caltech, Apr 07 8

5 Exploit RNDF Have a skeleton for the roads => graph search Expand the spine by adding rails spaced at 0.5 m Add transitions for changing lanes, intersections Use dynamic graph (generated on the fly) to further augment alternatives Remove edges with obstacles; paint cost for lanes, etc Rail 7 Sample RNDF 1 1 Waypoint Lane Zone Stop Sign 2 Segment / Zone ID Checkpoint ID v1.0 N Parking Lot* Traffic Circle way Stop * Note: The southern 6 waypoints in the Parking Lot (Zone 14) are Checkpoints Team Caltech, Jan 08 9 Logic passing finished or obstacle disappeared ROAD REGION passing finished or obstacle disappeared DR,NP,S STO,P,S no collision-free path exists STO,NP,S DR,P,S backup finished or failed and the number of times Alice has switched to BACKUP is less than some threshold Alice has been stopped for long enough and there is an obstacle in the vicinity of Alice no collision-free path exists and there is only one lane STO,PR,S DR,P,S and the number of times Alice has switched to the DR,P,R state near the current position is less than some threshold no collision-free path exists and there is more than one lane STO,P,S DR,PR,S and the number of times Alice has switched to the DR,P,R state near the current position is less than some threshold DR,A collision-free path with DR,A is found STO,A backup finished or failed and the number of times Alice has switched to BACKUP exceeds some threshold BACKUP and the number of times Alice has switched to the DR,P,R state near the current position exceeds some threshold and there is more than one lane and the number of times Alice has switched to the DR,P,R state near the current position exceeds some threshold and there is only one lane DR,B STO,B and there is only one lane and there is more than one lane OFF-ROAD mode FAILED PAUSED Team Caltech, Jan 08 collision-free path with DR,P,R is found 10

6 Lessons Learned from Alice Online optimization solves nonlinear control problems Modern computation allows constrained optimization problems to be solved online Solutions exist for situations with more limited computation (multi-parametric optimization) Layered control architectures allow more efficient design Allows for separation of concerns between subsystems Provided a very modular design, capable of rapid (humancontrolled) adaptation Verification of control protocols is necessary, but hard Traditional methods of simulation and testing not sufficient Formal methods not easily applied to hand designed control protocols New tools for correct by construction design are needed Temporal logic(s) are powerful language for specifying desired behavior (combined with traditional measures) New tools are becoming available, but not yet ready for prime time Mission Traffic Follower Actuation Interface Vehicle UTC, 27 Oct Control of Complex Systems (RMM) BE 107, 14 Mar 2015 /BE 12

7 Bioplausible Command and Control Humbert, Dickinson & M CDC05 Different architecture than engineering Large collection of diverse sensors (many more than required) Very slow computation with lots of parallel pathways Stabilization of a goal image/optical flow Compute forces and torques based on spatial integration of optimal flow S. Fuller (Caltech) Integration kernel F encodes controller Sean Humbert (Caltech/U. Maryland) BE 107, 14 Mar 2015 /BE 13 Censi, Han, Fuller, Straw and M CDC09 Bootstrappable Design for Visual Pose Stabilization Approach: minimize image error J(q) = 1 2 (y g)2 = 1 2 S 2(y(s) g(s)) 2 ds Can write the resulting controller in terms of spatial gradients + products Lagged or delayed y also works S can be learned (up to a positive definite factor) by watching how the environment moves given f, τ Structure of the resulting controller Weights have a sparse structure: take positive and negative products of nearby sensors Controller with delayed y has similar structure to Reichardt correlator (rough measurement of optic flow) IDEAS, Jul 2014 /BE 14

8 Censi, Han, Fuller, Straw and M CDC09, IROS10 3D Pose Stabilization Experiments Goal: stabilize 3D position and orientation using bioplausible visual feedback IDEAS, Jul 2014 /BE 15 Conclusions and Discussion Autonomous vehicles Engineering architecture; exploit modern computat n Multi-layer, distributed, complex General architecture; many uses Bio-inspired vehicles Bioplausible design (slow computing) Small number of layers ; reactive behaviors Test bio-inspired principles BE 107, 14 Mar 2015 /BE 16