Flexible Manufacturing Line with Multiple Robotic Cells Heping Chen hc15@txstate.edu Texas State University 1
OUTLINE Introduction Autonomous Robot Teaching POMDP based learning algorithm Performance Optimization Single cell optimization Multi cell Optimization Discussion 2
Trim and Final Assembly Line 3
Problems in Production Line When one work station fails, the production line has to be stopped Low efficiency High cost For complex manufacturing processes, is it possible to optimize the system performance? Single work station optimization? Multi work station optimization? 4
Proposed Solution Station controller/ Computer Station controller/ Computer Station controller/ Computer Other devices Station controller/ Computer Station controller/ Computer 5
Problems Big part location errors Can happen after a new batch is launched Parts could come from different suppliers Typically manual teaching methods are used to adjust the parameters Production line has to be stopped Installing additional sensors generates new problems System upgrading cost Maintenance issues Sensor failure 6
Failure Correction Online Transfer Learning Online Transfer Learning Online Learning Online Learning Online Learning #1 FMS #2 FMS #n FMS Off-Line to On-Line Teaching MTS Off-Line to On-Line Teaching FMS:Fixed Manufacturing Station Kinect SICK MTS:Mobile Teaching Station 7
Autonomous Robot Teaching Reducing the operational cost Sensor accuracy problem Calibration is difficult Noise Target Hole Transmission Part Marker for Tool Detection Robot Tool 8
Autonomous Robot Teaching Autonomous Industrial Mobile Manipulator ABB IRB 140 Industrial Camera Hole Transmission Part Tool/Peg Adult Robot PC Controller Gigabit Ethernet Ethernet DALSA Industrial Camera IRC5 Robot Controller Workpiece Tool & Peg Child IRB140 9
POMDP Partial Observable Markov Decision Process State Transition is Uncertain Observation is uncertain State is partial observable Why POMDP? Using POMDP to estimate the underlying errors through executing actions and receiving observations 10
Belief State POMDP Method Real state is unknown How to make decision? Assign a belief b over state S Update the belief state in each step Defining Value Functions Approximating Value Functions Alpha Vector 11
OUTLINE Introduction Autonomous Robot Teaching MDP based learning algorithm POMDP based learning algorithm Performance Optimization Single work station optimization Multi work station optimization Discussion 12
Example: Single Stage Assembly 13
Assembly Process Parameters: search force, search radius, search speed, insertion force How to tune the Assembly Parameters? 14
Example: Multi-Stage Assembly 15
Problems Design of Experiment (DOE) Genetic Algorithm (GA) Genetic Algorithm (GA)+ANN Low Efficiency Low Accuracy Model Free Offline Manufacturing line has to be stopped Cannot deal with variations, including part location errors, part geometry errors and environmental errors. 16
Single Work Station Optimization First Time Through (FTT) rate How to balance FTT rate and cycle time? FTT rate is calculated statistically. Cycle time is recorded each assembly Modeling? Optimization methods? FTT prediction? Cycle time estimation? 17
Multi Work Station Optimization FTT rate of the whole system Cycle time of the production line 18
Performance Optimization Performance optimization Knowledge transfer What can we transfer? How to transfer? 1 n k 2 n k M n k 1 e k 1 x k 1 y k 2 e k 2 x k 2 y k M e k M x k M y k 1 2 M u k u k u k Stage 1 Stage 2 Stage M Multi-Stage Learning Methodology 19
Comments Robot teaching System optimization Interesting topics? Discussions 20