Precision-Driven Hybrid Control for 3D Microassembly Dan O Popa - - PowerPoint PPT Presentation
Precision-Driven Hybrid Control for 3D Microassembly Dan O Popa Aditya N Das Dan O. Popa, Aditya N. Das Texas Microfactory Texas Microfactory Automation & Robotics Research Institute University of Texas at Arlington, Texas, USA
Presentation at ICRA 2009, Kobe, Japan, May 17, 2009 p y
DFW
NEXT GENERATION ROBOTICS Humanoids | Precision Microrobotics | Distributed Robotics SMART MICROMACHINES Smart Materials | Smart Devices | Smart Networks MICROMANUFACTURING Packaging | Assembly | Testing | Reliability
μνrobotics
Packaging
Precision Design
Robotics Modularity Packaging
Sensors, Actuators, Pow er Scavengers Materials
y
I nteractivity Structural Health Managem ent Biom im etic Microrobotic Sw arm s Hum anoids
bl h l d l
i i d i l l
Serial Manual Teleoperated May use multiscale, multirobot platforms with microgrippers Probes on manual stages platforms with microgrippers. Vary in throughout and yield. Microgripper arrays Automated Deterministic Stochastic Flip Chip Bonding/Stacking Self assembly Parallel Distributed Array
Deterministic assembly:
Stochastic assembly:
This presentation proposes a precision-adjusted path-planning and hybrid control strategy to improve the This presentation proposes a precision adjusted path planning and hybrid control strategy to improve the yield and speed of serial, automated microassembly.
Deterministic Stochastic
Deterministic Stochastic Top-down Serial Assembly of small components is difficult :
interdependency S li f h i Bottom-up Parallel Self assembly
for equipment
10-6 10-7 10-8 10-9 10-3 10-4 10-5 10-2 in meters
vibrations Effect of gravitational forces Effect of surface forces Ki i /D i l F C l
D.O. Popa, H. Stephanou, “Micro and meso scale robotic assembly”, in SME Journal of Manufacturing Processes, Vol. 6, No.1, 2004, 52‐71.
Kinematic/Dynamic control Force Control Assembly time Predictability (yield)
Automated , Serial or Parallel Hybrid Microassembly is a viable pathway for mass production of microsystems with modular component dimensions above ?0 microns. ff Automated Deterministic Microassembly is difficult and requires careful consideration of several important tradeoffs: i. Serial microassembly is slow but can have high yield and construct complex 3D shapes for part dimensions above ?0 microns. ii. Open-loop control with calibration or models can be used, but:
i. If it relies on high repeatability it usually leads to both expensive and bulky hardware. ii If it relies on compliance then it must be engineered into ii. If it relies on compliance then it must be engineered into parts and end-effectors.
iii. Closed-loop feedback control can provide higher precision but:
i. Usage in quasi-static operation decreases throughput (“ ) (“look and move”). ii. Usage in dynamic operation leads to vibrations which are
bandwidth. iii. Workspace is limited, this applies to sensors , actuators and end-effectors sharing the workspace.
Examples of 3D microassembly from Texas Microfactory
Key factors driving the assembly yield at small scales:
Precision requirements (how accurate the robots are) Tolerance requirements (how accurate the parts are) Throughput requirements (how fast we want to assemble) Interaction forces between parts (how parts interact) Sensory vs. sensorless ability (what kind of sensors do we have available) Part design (are the parts “assemblable”?).
The goal of our work is to formulate a framework with quantitative metrics and decision rules for The goal of our work is to formulate a framework with quantitative metrics and decision rules for microassembly cells: ‐ Design rules – how many robots, heteroceptive sensors, their specs, etc. ‐ Hybrid Control rules – when to switch between controllers Path Planning rules which paths to follow to increase precision ‐ Path Planning rules – which paths to follow to increase precision
Our approach i. High yield assembly guarantees from “new” precision metrics ii. Precision adjusted hybrid control for faster throughput through a “complexity index” j y g p g p y iii. “Precise” path planning algorithms to reduce sensor cost and ensure higher precision iv. Complex 3D simulation of microassembly in a virtual world with high realism for planning and validation before being ported to the assembly system.
Mi / f t i i t f t f d t f b i t f t t t ith
dimensions described in micrometers or nanometers.
the fundamental practices of precision engineering . However, these practices are rooted in pre‐20‐th century technology.
standards are set by NIST and ANSI. They have set precision standards for automated machines, including robots robots.
mechanical structure of the system making them very bulky.
Intelligent Robotics offers the possibility to reduce the size of micro/nano manufacturing systems, but precision concepts and standards in robotics are rarely followed.
We propose a new precision framework for top down robotic assembly systems with guaranteed high yields, reasonable high speeds, and reasonable low cost.
conventional definitions of accuracy, repeatability, resolution.
Resolution: The smallest output increment that a machine can perform. Repeatability: The ability of a machine to return to the same state over many cycling attempts.
good accuracy Poor
Accuracy: The maximum expected difference between the actual and the ideal (desired) output for a given input In Metrology, Resolution and Accuracy are mean values, while
repeatability
gy, y , Repeatability is a statistical distribution. Outputs vary, and examples could be, motion (position), measurement (distance, angle, temperature, intensity, etc), parts produced (tolerance of parts).
poor accuracy good repeatability
For Precision Robotics, these metrics should :
Functions.
with similar metrics at other points of measurement.
good accuracy good repeatability
Be redefined to closely match the type of controller, sensor and actuator system used.
Direct: Measurement of motion at the tool
‐ Pros: few sources of errors, related to the tool position measurement instrument only. ‐ Cons: more complex measurement sensors, servo loop controller K2 more complex
Indirect: Measurement of motion at the joints
‐ Cons: lots of sources of errors: measurement instrument error (for instance joint encoders), kinematic parameters, other unknown factors (friction, backlash, temperature variation). Inverse model construction needed. ‐ Pros: Calibration loop controller K1 simple, model needed only once. Both direct and indirect measurements are needed for high precision machines. Part/tool position measurement feedback
‐
feedback Desired end‐effector position Xd
Inverse i i d l
+
Ex=Xd‐X U F
End‐effector position X
Joint position feedback Q
‐
Kinematic Model
Desired Joint position Qd
+
Eq=Qd‐Q
Gripper frame part frames f3 f4
f2 f3 f4 f1
Global frame Fiducial frame
[Popa02] D. Popa, B.H. Kang, J. Sin, “Reconfigurable Microassembly System for Photonics Applications,” in Proc. IEEE Conf. on Robotics [Popa02] D. Popa, B.H. Kang, J. Sin, Reconfigurable Microassembly System for Photonics Applications, in Proc. IEEE Conf. on Robotics and Automation, Washington, D.C., 2002. [Popa04] D. Popa, H. Stephanou, “Micro and Meso Scale Robotic Assembly”, in SME Journal of Manufacturing Processes, vol. 6 No. 1, 2004, 52‐71. [Popa06] D. Popa, R. Murthy, et.al., “M3‐Modular Multi‐Scale Assembly System for MEMS Packaging”, in proc. Of IEEE/RSJ Int’l Conference on Intelligent Robots and Systems (IROS ’06), Beijing, China, October 2006. [Popa09] D.O. Popa, R. Murthy, A. N. Das, “M3‐ Deterministic, Multiscale, Multirobot Platform for Microsystems Packaging: Design and Quasi Static Precision Evaluation ” in IEEE Transactions on Automation Science and Engineering (T ASE) April Packaging: Design and Quasi‐Static Precision Evaluation, in IEEE Transactions on Automation Science and Engineering (T‐ASE), April 2009.
1. Accuracy: The robotic system is commanded to place the end‐effector at a designated position in 3D‐space which may involve translation and/or rotation in R6. A fixed or mobile sensor (for instance, an optical microscope) is used to determine the error between actual and desired position of the end‐effector. The error distribution with respect to the sensor gives the measure position of the end effector. The error distribution with respect to the sensor gives the measure for accuracy, σacc. (Note: conventionaly, accuracy is a mean value, not a distribution). After repeated motions (regardless of control method)
⎞ ⎛
⎟ ⎞ ⎜ ⎛ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ Δ + + ≤
= ∞ → N N j j N Sk SPk S k accuracy
q K N q q
1 2 2 2 2 _
1 ) ( 1 lim min σ σ σ
Sensor frame localization uncertainties are σSk, and its measurement uncertainty of the k‐th DOF of part P is given by σSPk, ΔKj is the variation in position measurement j of the k‐th DOF
⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ + + ≤
= ∞ → j j N k accuracy
q K N Sk q SPk q K
1 _
) ( 1 lim ) ( )) ( ( ) ( μ μ
p g y
SPk, j
p j about its mean. Go one step further: use open loop control with an nominal inverse model to drive the system when making these measurements, σo=σaccuracy E.g. define accuracy as the robot intrinsic g ,
g f y precision as measured by heteroceptive sensors, but without using this data in any way.
2. Calibrated Model: Calibration Error is the uncertainty value of the forward kinematic map K(q) averaged over the set of all features part σKk(q). This uncertainty depends on the number of measurements N, the number of features Q, and the sensor uncertainty for all measurements, e.g.
~
SPki
Clearly,
~
) ( )) ( ( ) ( lim ) (
k l b
Sk q SPk q k Q N K q K μ μ + ≥ =
Z Y Xp Zp
Yc
Zc
2 2 2 , 2 _ , _
) ( lim ) ( ), ( )) ( ( ) , , , ( lim ) (
Sk SPk Kk Q N k n calibratio Q N k n calibratio
q q Sk q SPk q k Q N K q K σ σ σ σ μ μ + ≥ = + ≥ =
∞ → ∞ →
Yo Xc
between two predefined but arbitrary points in 3D‐space (corresponding to two respective joint p y p p ( p g p j coordinate vectors) one of which is measured through a sensor. The error distribution with respect to the sensor gives the measure for repeatbility, σrep. (Note: this corresponds to the traditional definition of repeatablility or precision) p p y p
1 2 2 2 _
= ∞ →
N j j N SPk S k repeat
2 2 2 t lib ti
repeat n calibratio c
manipulator can sense and execute in Cartesian space, 3σres transformed into a distribution with a mean 3σres and uncertainty σres given by: (Note: this is typically one or the other, usually the minimum increment that the robot can
_ SPk k k res
execute). During servo operations:
1 1
S
− −
s s
2 _ 2 2 state steady repeat s
Consider an assembly with tolerance Δk12 along DOF k, between a part 1 manipulated from position A, and part 2 fixed on a substrate at position B.
If the use of fixtures at taught locations along DOF k, will cause the assembly operation to succeed 99%+ of the time, will incur a speed cost C1=f1(AB), where f1 is a monotonically increasing function of the distance between A and B, and a sensor/feedback cost C2=0 (no heteroceptive sensors or feedback)
σ 3
12 >
Δ
2
( p )
the use of servoing along DOF k will cause the assembly operation to succeed 99%+ of the time, will incur a speed cost C1=f1(AB), and a sensor/feedback cost C2=f2(AB).
the use of calibration along DOF k with sufficient calibration points N,
s
k σ 3
12 >
Δ
c
k σ 3
12 >
Δ
g p , will cause the assembly operation to succeed 99%+ of the time, will incur a speed cost C1= f1(NAB) , and a sensor/feedback cost C2= f3(AB). The monotonically increasing functions fi are assembly system designer’s choice.
c
σ
12
y g
i
y y g Assembly planning can be done to decide how to switch between the three modes of
weights. g
Precision metrics were used in conjunction with assembly workcell design parameters to
Increase assembly speed due to precision adjusted hybrid control (e.g. switching between different operation modes depending on precision requirements).
Hardware incarnations at Texas Microfactory - three progressively smaller workcells:
with multiple end effectors: microgrippers zoom
Laser solder reflow (delivery optics) – 3DOF Zoom‐camera system – 2DOF Gripper Manipulator 4DOF Hot plate for die attach
with multiple end‐effectors: microgrippers, zoom microscope, laser for solder reflow.
dimensions of O(10‐1~10‐2m), handles parts of size
4DOF Tool tray with quickchange end‐effectors
dimensions of O(10 10 m), handles parts of size O(10‐2~10‐4m), and achieves accuracies in the scale of O(10‐4~10‐6 m).
Parts tray Fine manipulator 3DOF
Packaged MOEMS device using the M³ Schematic and Control System Diagram of M³ using the M Schematic and Control System Diagram of M
Assembly and packaging of heterogeneous MOEMS device.
Top Chip MOEMS Die Sn/Au Preform Carrier Optical Fiber
Manipulator accuracy < tolerance required (fiber to package, In preform insertion) Manipulator repeatability < tolerance required (die
Manipulator resolution < tolerance required (fiber to die) attach, top chip attach)
ΔX ΔY ΔZ Δθ Δϕ ΔΨ
D.O. Popa, R. Murthy, A. N. Das, “M3‐
FIBER TO PACKAGE 300 300 186 1.73 1.73
(µm) (µm)
(YAW) deg
ϕ
(PITCH) deg (ROLL) deg Deterministic, Multiscale, Multirobot Platform for Microsystems Packaging: Design and Quasi‐Static Precision Evaluation,” in IEEE Transactions on Automation Science and Engineering (T‐ ASE) April 2009
FIBER TO TRENCH 4 4 25 0.2
workspace with a total of 19 DOF.
Three μ³ end‐effectors share a common 8 cm³‐ size workspace
assembly with high yield guarantees.
O(10‐2~10‐3m), handles MEMS components of O(10‐
3~10‐5 m) in size, achieves O(10‐6~10‐7 m) accuracy.
p
10 m) in size, achieves O(10 10 m) accuracy.
Silicon MEMS microgrippers and microsnap fastners Assembled MOEMS: scanning mirrors, micro‐ball lenses, microinterferometer 1mmx1mm Cu coupon assembled onto a MEMS die using the μ³.
A Das P Zhang W H Lee D Popa H Stephanou “µ³: Multiscale Deterministic Micro‐Nano Assembly System for Construction of On‐
Wafer Microrobots,” in Proceedings of ICRA 2007.
2 2 2
2 3 2 2 2 1 k k k
Repeated Zyvex Connector Assemblies with High Yield Microspectrometer Assembly y g y
R Murhy A N Das D O Popa “High Yield Assembly of Compliant MEMS Snap Fasteners ” in Proc of
ASME Int’l Conf. on Micro and Nano Systems, August 2008.
_
ri ABi
5
a EIs =
x
F
1
a = α
B A
y
F
4 3 2
, , a t a w a l
s s s
= = =
1 1 θ
y
_
a
_
a
j ABx
j ABx
f ABz r
Microassembly of MEMS,” in proc. of ANS Conference Sharing Solutions for Emergencies and Hazardous Environments, Salt Lake City, Utah, February 2006.
yield j ABy
,
Skidmore, “Tolerance analysis of placement distributions in tethered micro‐ , y p electro‐mechanical systems components,” in Proceedings of IEEE International Conference on Robotics and Automation, May, 2004.
Reference Points Test Points
in microns X-Position Y-Position Z-Position
Array of Zyvex connectors
Using probing calibration: much better precision: microns Mean 1.6267 2.0144 2.2457 Std Dev. 1.014 1.9067 1.5869
Snapshots of Calibration Experiment
for Construction of On‐Wafer Microrobots,” in Proceedings of ICRA 2007.
Snapshots of Calibration Experiment
Advantages:
for visible and NIR range.
fasteners fabricated by DRIE
Michelson interferometer based Fourier Transform spectrometer by ARRI, University of Texas at Arlington.
approximately 30grams total weight
A.N. Das, J. Sin, D. O. Popa, H.E. Stephanou, “Design and Manufacturing of a Fourier Transform Microspectrometer,” in Proc. of IEEE Int’l Conf. on Nanotechnology, August 2008.
Leg Actuator 900µm Joint
Constructed Via 2 ½D MEMS Assembly,” in Proc. of 2008 IEEE Int’l Conf. on Intelligent Robots and Systems (IROS ’08), Nice, France, Sept. 22‐26, 2008.
XY stage Cable drive TCP Microrobot volume:2mm x 3mm x 1mm W k V l 50 50 75 Work Volume: 50μm x 50μm x 75μm Actuation: Electrothermal Transmission: XY – direct drive ϕψ ‐ cable driven
p , , p , y
From a few robots+controllers to many µrobot via assembly and die bonding. The workspace of a N³ robot is of the order O(10‐4m), has manipulators consisting of components with dimensions O(10‐3~10‐4m), and handles nanoscale components ranging between O(10‐5~10‐7 m)
in size. If we are to linearly extrapolate the accuracy requirements for such a robot from the larger scales, we should require it to have O(10‐7~10‐9 m) accuracy.
D O Popa R Murthy A N Das “M³ μ³ and N³: Top‐down Deterministic Macro to Nano Robotic Factories with Yield and Speed
M , μ , and N : Top down, Deterministic Macro to Nano Robotic Factories with Yield and Speed Adjusted Precision Metrics,” in Proc. of 2008 Int’l Workshop on Microfactories (IWMF ’08), Evanston, Illinois, Oct. 6‐8, 2008.
2> σ2 2+ σ3 2 ?
2 ) sgn( 1 ) (
2 1 2 3 2 2 3
σ σ σ σ − + + = Ω
2
2 ) sgn( 1 ) (
2 1 2 3 2 2 3
σ σ σ σ − + + = Ω
Complexity Index The assembly tasks can be accomplished by a control structure that will be selected among the following cases:
sensing only.
, g p p p g y
dependency mapping. Rules for controller selection:
3 Select controller 3 only when Ω(σ ) 0
A : Assembly process Ti : General pick and place task
“Complexity Index”
T1 Tai P x1
Ti : General pick and place task
sTi : Special tasks
Tai : Pick, Tbi : Move, Tci : Place
Hybrid Controller Block Diagram
A T2
sTi
Tbi Tci x2 xi * ∑
i s i
P : “Motion uncertainty” S : “Complexity Index”
i
Tn xn σ1 ┤ Ω
i i ci bi ai i
n T i d Total
= i T i d Total
bi
1
Determination of complexity index
3D rendering of 3D rendering of virtual microparts
n*(δD/v) 10 ~ 20 sec. n*[(δD/v) + τs]+ m*τc 300 ~ 600 sec.
Cumulative uncertainty in part positioning after travelling over
Open loop
positioning after travelling over 16750microns having 10 sensor fields 3D rendering of assembly configuration of microspectrometer; (a) virtual μ³ robotic assembly setup (b) close-up view of the micro-part (mirrors, grippers, lens holders) and device dies (assembly sockets) and (c) diagram of assembled 2½D MEMS d ff h h lf i l h
Closed loop
parts and off-the-shelf optical components on the spectrometer substrate
2 sgn 1 ) (
2 3 2 2 2 1
σ σ σ σ + − − = Ω
precision
Complexity Index (Ω):
For task 2: Ω = 0 as; σ1
2 > (σ2 2 + σ3 2)
Robot precision parameters: Resolution = 0.01µm
Ω = 1 Task is difficult Ω = 0 Task is simple
Identification of complexity
Repeatability = 0.07µm Accuracy = 0.4µm A calibration model based open loop controller is used to assemble part2
Process Description σ1 (given) Σ2 (measured) (average value) σ3 (measured) (average value) Complexity Index Ω Controller Identification of complexity
controller is used to assemble part2.
) Task 1 Ball lens holder onto fixed socket 3µm 0.5µm 2.7µm 1 Open loop nominal and closed loop visual servoing Task 2 MEMS mirror
4.5µm 0.5µm 1.8µm Open loop: calibration based socket 4 5µ 5µ µ Task 3 Ball lens holder onto fixed socket 3µm 0.6µm 3.2µm 1 Open loop nominal and closed loop visual servoing Task 4 MEMS mirror t i Open loop nominal d l d l i l
socket 1.5µm 0.3µm 3.8µm 1 and closed loop visual servoing
Assembly sequence 1 (Sensor count = 1) Precision sought (linear in x & y and angular in θ) Average Time Taken in seconds Yield Factor (out of 20 iterations) Final Assembly Outcome (<60% Failure) Part 1 assembly 6μm 0 5 degree 90 3 out of 20 Failed (15%) Part 1 assembly Lens holder 1 onto fixed socket Distance=[11mm, 4 breaks] 6μm, 0.5 degree 90 3 out of 20 Failed (15%) 3μm, 0.5 degree 130 16 out of 20 Succeed (80%) 1μm, 0.25 degree 220 20 out of 20 Succeed (100%) Part 2 assembly Micro mirror 1 onto fixed socket Distance=[24mm, 8 breaks] 6μm, 0.5 degree 105 1 out of 20 Failed (5%) 3μm, 0.5 degree 140 18 out of 20 Succeed (90%) 1μm, 0.25 degree 240 20 out of 20 Succeed (100%) Part 3 assembly Lens holder 2 onto fixed socket Distance=[27mm, 6 breaks] 6μm, 0.5 degree 110 0 out of 20 Failed (0%) 3μm, 0.5 degree 145 11 out of 20 Failed (55%) 1μm, 0.25 degree 245 20 out of 20 Succeed (100%) Part 4 assembly Micro mirror 2 onto movable socket Distance=[33mm, 8 breaks] 6μm, 0.5 degree 115 0 out of 20 Failed (0%) 3μm, 0.5 degree 160 8 out of 20 Failed (40%) 1μm, 0.25 degree 310 19 out of 20 Succeed (95%)
† Assuming image process time to be 100msec per iteration and robot velocity of 0.85mm/sec.
Assembly Process & Resulting Parameters For i Pure Open Loop (Nominal d l b d) Pure Open Loop (Calibration b d) Closed Loop (Nominal and i l i ) Hybrid Controller (Calibration based Open loop + Nominal d l d i l Microspectrometer Assembly (four parts per die) model based) based) visual servoing) model and visual servoing based closed loop)
Overall Yield 20% 30% 99.9% 95% Overall Yield 20% 30% 99.9% 95% Average Cycle Time (Calibration & Manipulation) 6 to 10 minutes 11 to 15 minutes 50 to 80 minutes 20 to 35 minutes Sensor Count 1 2 2
60% faster than closed loop 5 times the yield of open loop Nearly 100% yield Assuming image process time to be 100msec per iteration and robot velocity of 0.85mm/sec. The variation in time is due to fact that each iteration uses different path sequences. Similarly the overall yield is calculated by averaging the success rate
In this assembly scenario two each from two types of microparts are picked up from two different dies and placed
sockets on a third die. The assembly is carried out in a pre‐ specified sequence. Calibration time: 12 minutes Total assembly time: 75 minutes (hybrid control) control)
Process yield: 100% Reconfiguration time: 5 minutes
Automated, deterministic serial microassembly provides a pathway to build high yield and cost effective complex heterogeneous microsystems. A major drawback of closed loop serial microassembly is low speed and space constraints in A major drawback of closed loop serial microassembly is low speed, and space constraints in sensing. Precision measures for robotic system are redefined to associate resolution with visual servoing, repeatability with calibrated operation, and accuracy with open loop operation. g p y p y p p p The High Yield Assembly Condition (HYAC) guarantees nearly 100% assembly yield of compliant
By implementing a hybrid controller which selects suitable assembly control based on the By implementing a hybrid controller which selects suitable assembly control based on the precision requirements of a specific task in the sequence, additional objectives, such as low cycle time can be accomplished. We presented a virtual 3D microassembly environment having necessary and sufficient visual, kinematic dynamic realism in order to test multiple assembly schemes prior to assembly. We demonstrated assemblies of complex microsystems such as the microspectrometer showing that the hybrid controller is more efficient than pure open loop, or pure closed loop control.