Field And Space Robotics Laboratory
Robot Assisted Crucible Charging System


Shin-Etsu Handotai Co. Ltd.


Professor Steven Dubowsky
Dr. Long Sheng Yu, visiting scientist
Joseph Calzaretta, graduate research assistant
Tony Leier, graduate research assistant
Vivek "Superfly" Sujan, graduate research assistant
Kristie Yu, graduate research assistant
Melissa Tata, graduate research assistant

Problem Statement

The Task:

To charge a fused silica crucible with irregularly shaped polysilicon nuggets during CZ semiconductor wafer production. 

Key Constraints:


Characteristics of a charged crucible 

Technical Challenges

Overall system design

The following figure shows the overall design of the RACS laboratory demonstration system. The purpose of the system is to demonstrate the technology required to robotically fill crucibles in a factory setting. An Adept One Robot with a custom built gripper and controller is the primary component. The Adept One manipulator obtains nuggets from a feeding station within its workspace. The nugget is then passed over the nugget geometry acquisition (NGA) vision component where the bottom topology of the nugget is determined. While the nugget is being grasped, the surface geometry of the nuggets in the crucible is determined with the overhead vision system. A packing algorithm running on a control computer combines the nugget and crucible geometries to find a placement location. When a suitable placement location for the nugget is determined, the nugget is moved into the crucible and placed in its location. The Adept One robot, crucible, and vision systems share a common, stiff base.


System Parameters

In order for the SEH system to map an accurate and optimum location for a nugget within a field of nuggets, a vision system is chosen to help extrapolate the data needed for such an operation. The vision system uses 2 separate active laser triangulation systems to scan a 3D depth map of the nugget to be placed and the nugget field into which the nugget will be placed. These 2D vector field maps are then meshed together using a packing algorithm based on virtual trial and error, in order to find an optimum placement location.

In our system active laser triangulation is achieved by using the reflection of a laser line off the object onto a single CCD camera frame. The geometrical relationships of position and orientation is known between the camera and the laser light source. Using this information the illuminated points can be triangulated. For the Nugget Geometry Acquisition (NGA) phase, the laser line is scanned across the nugget by moving the nugget across the field of view by the manipulator/gripper at a constant velocity. For the crucible Surface Geometry Acquisition (SGA) phase, the laser line is scanned across the nugget field using a one dimensional galvanometer mirror scanner. Hence a full map of exposed surfaces is achieved. 

Nugget Geometry Acquisition (NGA)
Surface Geometry Aquisition (SGA)
  The project requires that a variety of irregularly shaped polysilicon nuggets be successfully grasped and accurately placed in the crucible. The range of sizes for nuggets that must be individually placed is shown in the following graph. The smallest nuggets that must be individually placed by the robot have a characteristic size of 2cm x 2cm x 1cm and weigh about 60 grams. The largest are 8cm x 8cm x 6cm and weigh 600 grams.
The end effector for the Adept 1 robot.

The end effector for the Adept One robot will pick up a high percentage of the nuggets and place them into the desired crucible location. The design consists of a closely packed triad of deep 20mm diameter suction cups mounted on a two degree of freedom gimbaled wrist. This wrist is attached to a JR3 six axis force/torque sensor that is connected to the Adept One. The suction cups are connected to a vacuum source. The vacuum can be released with a valve controlled by the master control computer. Nuggets smaller than 60 grams are placed within a bin that the manipulator can grasp and pour into the center of the crucible; these nuggets need not be precisely placed.


The major technical challenge of the control system is the sliding of the nugget into place while it contacts the crucible wall.. The following figure depicts the events of this wall-building process.


Wall-Building Process.

The manipulator brings the nugget from the scanning station to the crucible with a high speed slew motion. Then the nugget is brought into contact with the wall. In order to maintain a safe force level while moving the nugget along the wall, the system will switch out of position control to a different control algorithm Possible control algorithms include:

  1. Compliant Position Control: The manipulator will be very compliant in the direction normal to the wall surface, and stiffer in those directions parallel to the wall surface. Position control will be used. The manipulator is directed to move to a position beneath the surface of the wall. The springiness of the control scheme will cause the arm to exert a constant force against the wall (as it attempts to move the nugget into the crucible wall.)

  2. Hybrid Control: The manipulator will use standard position control in the directions parallel to the wall surface, and it will use force control in the direction normal to the wall surface. This algorithm involves more active monitoring of the wrist sensor. The following figure depicts a block diagram representation of a hybrid control system.

    Example of Hybrid Control System
  3. The blocks labelled F and P are selection matrices which effectively split up the control signal into force and position components. The force sensor allows for explicit force control which is more responsive than compliant position control, but raises stability issues.

  4. BaST-Assisted Control: This scheme can be used in conjunction with the two aforementioned control methods. The manipulator is mounted on a base force/torque sensor. The force felt at the base is a combination of gravitational, dynamic, and wrist contact forces. The gravitational forces can be calculated, and the wrist forces are measured. Thus the base sensor allows the system to estimate the dynamic forces on the manipulator, and subsequently the effective torques at each joint can be estimated. The control system can therefore control the torques at each joint, and thus greatly reduce unmodelled frictional effects. BaST control has been shown to improve the performance of slow, precise manipulator motions, and could prove useful to the RACS system.
The system will then detect the contact with the existing layer of nuggets and implement the planning/packing strategy.


Meshing of the nugget with the nugget field is accomplished using a packing algorithm. In this algorithm, the nugget to be placed and the pre-existing nugget landscape are represented by sets of vectors. The optimum nugget placement location is then located in a virtual environment. Once this location is determined, it is sent to the control system to be realized physically.
Vector Representation of Nugget Surface


Key Technical Issues for the Factory System Design:

The factory design must provide for easy crucible transport to and from the robot packing area and for complete operator safety and ergonomics. The design must accommodate robot access to all edges of a 36 inch crucible, to the small nugget holding device, and to the large nugget loading area. The operator must be able to sort the nuggets and transport them to the robot loading area without ever entering the robot's workspace while the robot is operation. There must be a precise placement of the crucible relative to the robot to coordinate the vision and controls systems.

Research is being performed to determine any additional requirements that the actual factory system will have beyond those of the laboratory system. The primary difference between the two systems is the crucible size, which is an 18 inch diameter for the lab and a 36 inch diameter for the factory. Also, there is a higher budget for the factory system, so better equipment could be utilized. The crucible could be packed using one very large robot, using two smaller robots, or using one smaller robot and a turntable. Another possibility is to use one robot to load a portion of several crucibles. For one large robot, commercially available SCARA, gantry, and inverted articulated arm robots with a workspace of at least 42 inches in the x and y directions are being investigated. The robot(s) chosen must have good reliability and speed capability and must be rated for a clean room environment.

Other areas currently being investigated are optimal nugget sorting weight, number of operators needed for each robot, and cycle times for packing the 36 inch crucible robotically.

Copyright 1995 by MIT.

All rights reserved.