Vous êtes sur la page 1sur 23

Presented By Chaudhary Hemant Motiram (B.E. Mech) Email: hmc108@yahoo.co.in hmc2004@rediffmail.com Dipsankar Sinha (B.E. Mech) Email: rudycurz@rediffmail.



Machine Learning is a branch of Artificial Intelligence and is gaining importance. If Machine Learning is applied to robots then they can learn by themselves and thus help to reduce the tedious task of monitoring, programming and also will minimize the overhead charges. The paper takes into consideration the introduction of Machine Learning and its applications to robotics with a case study. The paper briefly explains the method of learning with a case study of system able of performing complex manipulation task. A manipulation system that can observe dynamic changes in the environment in real time. This consists of a grasping system that uses a high-speed visual and force feedback technique having a multi-fingered hand-arm with a hierarchical parallel processing system and a high-speed vision system SPE-256 and Grasping with high responsiveness and adaptive ness to dynamic changes in the environment is realized by Machine Learning.

Introduction to Machine Learning Necessity of Machine Learning Disadvantages of Present Systems Advantages of Machine Learning Applications of Machine Learning Learning Methods Requirements of Machine Learning Case Study: High Speed Grasping Using Visual And Force Feedback . Components of the System. DSP Subsystem Active Vision with SPE-256. 7-Axis Manipulator with Dextrous Multi-Fingered Hand Algorithm of High Speed Grasping Processing of the Components Processing for Active Vision Processing for Arm Processing for Hand 7. 9. 9. 9. 10. 10. 11. 11. 13. 13. 1. 2. 2. 3. 4.


Experimental Results Conclusion References

15. 18. 19.

Machine Learning is a new and rapidly developing area of technology with a great potential in a wide variety of applications. It deals with machines performing human like functions such as reasoning and interpretation. Machine Learning is a term that is usually applied to functions performed by machines that would normally require some properties of human intelligence. This typically involves either humanlike reasoning or senses. It is not just a software or hardware technology. It involves an entire process of data collection, computer logic and data processing .Development efforts in this area require advances not only in the fields of computer science but also in the basic understanding of the human senses and the brain. These machines do not operate like conventionally computer systems, which follow the instruction step by step. They use association, reasoning and decision making processes much like human brain would to solve problems. Some can learn from experience and communicate in natural language just as children grow from infancy to adulthood. This new generation systems differ from contemporary systems primarily in the ability to draw conclusions and make decisions based on the inference and deduction from base of knowledge. Because of their ability to handle data with some level of meaning and understanding, such machines can be used as intelligent problem solvers rather than merely as computation devices that follow strict instructions. Machine Learning can simulate many human capabilities, but it yet cannot create. This requires imagination and intuition, not just logic. Thus Machine learning is decision making ability by means of artificial intelligence to exhibit intelligent behavior without any help of a human being which involves Machines that think like humans and act like humans.


The standard method to perform any task by robot is to implement the robot with a hardwired control program designed by a human operator. The major disadvantages of these approaches are that because of the rigid nature of the control program the robot is not able to adapt to changes, for example in the environment (e.g. changes in surface colours or textures) or the robot itself (e.g. failure of individual sensors).Besides, the design of fixed control programs is costly and error-prone. There are numerous learning tasks that can be identified in the framework of robotics and robot control principally; Machine Learning can be applied to support the following actions during robot development and deployment: Initial knowledge acquisition and program generation, i.e., initial robot programming and world-model acquisition. Action and world knowledge refinement, i.e., acquisition of new knowledge in order to be able to solve new tasks as well as refinement and correction of existing knowledge helping the robot to deal with an uncertain and to adapt to a changing environment.


1) True robot position and perceived position drift apart considerably after journeys as short as 5 meters. 2) Because research on quantitative descriptions of mobile robot behavior is still in its infancy, mobile robotics to date is still an empirical discipline that uses existence proofs extensively. Robot systems to perform certain tasks are

implemented, but, for want of precise performance measures and behavioral descriptions, are not independently verified. 3) Due to several factors such as wheel slipping, the knowledge about the robot's position and orientation is inaccurate. 4) The sensor inputs are noisy. 5) The sensor characteristics (esp. when using ultrasonic sensors) might cause additional uncertainties. 6) Positions of objects that are detected in the environment are not known exactly.


Industrial and technical applications of mobile robots are continuously gaining in importance in particular under considerations of reliability (uninterrupted and reliable execution of monotonous tasks such as surveillance), accessibility (inspection of sites that are inaccessible to humans, e.g. tight spaces, hazardous environments or remote sites) or cost (transportation systems based on autonomous mobile robots can be cheaper than standard track-bound systems). Mobile robots are already widely used for surveillance, inspection and transportation tasks a further emerging market with enormous potential is that of mobile entertainment robots. 1) ADAPTS & RESPONDS TO CHANGING ENVIORNMENT. 2) DYNAMIC RESPONSE 3) ABILITY TO DETECT AND CORRECT ERRORS 4) PRECISE & SPECIAL FIXTURES ARE ELIMINATED 5) MOBILE 6) USERS THAT ARE INEXPERIENCED IN PROGRAMMING CAN USE THIS TECHNOLOGY.


1) 2) 3) 4) 5) 6) 7) 8) 9) 10) In CNC machines for prediction of deviation of quality and prediction of faults Driverless transportation systems. Transportation & Navigation Automated Inspection Robots Flexible Automation Military Missions Game Playing Augmented surgical manipulation tasks. A major aspect of automated inspection is to detect abnormalities automatically. They play an economic role also in the entertainment Industry (artificial pets being the best known example).



Just as a child is taught to learn in different ways Machine Learning also uses similar concepts for learning. 1) ADAPTIVE LEARNING. 2) SUPERVISED & UNSUPERVISED LEARNING. 3) REINFORCEMENT LEARNING. 4) EVOLUTION BASED METHODS.



1) The system should be able to adapt itself to changing conditions in the environment while performing a commanded task. 2) It must be able to acquire knowledge about its environment in form of a map or directly object-related information and to use this acquired knowledge efficiently. 3) The system should show continuous improvement of the system's capabilities. 4) System should have at least a limited ability to reason about the status and the usability of its own knowledge, in order to determine what kind of information is missing and what knowledge could possibly be acquired by experimentation, retrieved from a user demonstration, or directly be derived from user-given information. 5) Extraction of symbolic knowledge about the world from sensory data. 6) The applicability of learning technique with respect to a given task depends on the characteristics of both the task and the learning techniques. 7) ACTION RECOGNITION: One important issue in a system is the recognition of the human intent in order to, for example, provide appropriate assistance before the contact with the object has occurred. For example four different grasp types have been modeled for the purpose of recognizing the typical grasp for four different objects: a toy car, a pen, a glass and a cup shown in figure-2.1 Car is grasped with the Large Diameter Grasp. Glass is grasped with the Small Diameter Grasp. Cup is grasped with the Lateral Pinch Grasp. Pen is grasped with the Four-finger Thumb Grasp.

The possibility to recognize these grasps is investigated for giving a very simple measuring system.



Grasping motion is one of the most important processes for control of multifingered hands. To recognize objects and act upon them, humans typically use hand-eye coordination to identify the object, move the arm to the vicinity of it, preshape and align the hand and finally grasp and manipulate the object, similar concept is used in this system. To complete the grasping process multiple types of sensor information are needed such as visual information and force haptic information, and grasping motion should be


controlled by fusing this sensory information. Then the sensory information should be acquired at a rate as the cycle time of grasping control so that the grasping motion has responsiveness to dynamic changes in the environment. The initial design is motivated by the fact that complex tasks can be modeled as a sequence of elementary ones where each elementary action is represented by a set of tractable constraints originating from the e.g. robot kinematics, task representation, type of sensory input, etc. There are two important features to realize grasping in the real world, as follows: (3.1A) FLEXIBILITY UNDER MULTIPLE CONDITIONS A grasping system should have flexibility to complete various tasks under various conditions. The grasping process should be suitably changed according to the grasping condition, for example an object's position, an object's shape and an object's motion. To implement this, a hierarchical parallel processing architecture with several types of sensor is valid. Multiple types of sensory-motor fusion processing coexist in one system based on it. As a result flexibility in multiple grasping environments is realized. (3.1B) RESPONSIVENESS TO DYNAMIC CHANGES In the real world, the grasping environment changes dynamically and is possible that the object moves at high speed and sudden accidents happen during grasping. To overcome this, grasping control based on high speed sensory feedback is effective. Highspeed sensory feedback means to return feedback of external sensor information at a rate higher than the rate of control. Because the system can recognize an external environment in real time, responsiveness to dynamic changes in the grasping environment is realized.



In general the cycle time of 1ms is necessary to prevent mechanical resonance in robotic control. The cycle time of each sensory feedback should be 1ms to ensure stable grasping control. 4.1) DSP SUBSYSTEM The DSP subsystem is the main part which processes sensory feedback within 1ms. It has a hierarchical parallel architecture consisting of 7 DSPs connected to each other, and many I/O ports are installed for inputting various types of information in parallel. This system uses a floating-point DSP TMS320C40 which has high performance (275 MOPS) and 6 I/O ports (20 Mbytes/sec). By connecting severalC40 processors, a low bottle-neck hierarchical parallel architecture is realized. In the system the following I/O ports are


prepared; ADC (12 bit, 64 CH), DAC (12 bit, 24 CH), and Digital I/O (8 bit, 8 ports). These I/O ports are distributed on several DSPs to minimize the I/O bottleneck so that sensor signals are input in parallel. A parallel programming development environment has been prepared in which multi-process and multi-thread programming is easily realized. This function is useful to program parallel sensory feedback shown in figure-3.1 4.2) ACTIVE VISION WITH SPE-256. The active vision subsystem consists of a vision chip system called SPE-256 and a 2-axis actuator moved by DC servo motors. SPE-256 consists of a 16x16 array of processing elements (PE) and PIN photo-diodes (PD). The output of each PD is connected with a corresponding PE. Each PE is a 4-neighbor connected SIMD based processor which has a 24 bit register and a bit-serial arithmetic logic unit capable of AND, OR, and XOR operations etc. Because the visual processing is perfectly executed in parallel, high-speed visual feed-back is realized within 1ms. The SPE-256 is a scale-up model of an integrated vision chip and the next generation is currently being developed in which all elements of the vision chip architecture are integrated in one chip .The actuator part of the active vision subsystem has two degrees of freedom; pan and tilt. This is used to move the sensor platform and this is controlled by a DSP assigned for active vision control. 4.3) 7-AXIS MANIPULATOR WITH DEXTROUS MULTI-FINGERED HAND. The hand-arm subsystem is a 7-axis manipulator with a dextrous multi-fingered hand. The multi-fingered hand has 4 fingers and 14 joints. Its structure is similar to a human hand, in which a thumb finger is installed opposite to the other three fingers. Each joint is controlled by DC servo motors in a remote place using a control cable consisting of an outer casing and an inner wire. Each joint of the hand has a potentiometer for position control and a strain gauge for force control. The arm has 7 joints controlled by AC servo motors. An encoder is installed in each joint and a 6-axis force/torque sensor is installed at the wrist. Two DSPs are assigned to control the hand and the arm. Mapping using the joint space representation facilitates the similarity between hands poses. This is suitable for enveloping or power grasps. Mapping using the


Cartesian space is more suitable for representing the fingertip positions. This is a natural approach when, for example, precision grasps are considered.


The algorithm has 4 sensory feedbacks and one sensory information processing unit which are executed in parallel for following actions: 1) TRACKING CONTROL OF THE ACTIVE VISION 2) TRACKING CONTROL OF THE ARM 3) REACHING CONTROL OF THE ARM 4) GRASPING CONTROL OF THE HAND 5) VISUAL OBJECT RECOGNITION. 6) COLLISION AVOIDANCE These processes are distributed to 3 DSPs which are respectively assigned to the active vision, the arm, and the hand.


5.1) PROCESSING FOR ACTIVE VISION In the DSP for active vision, two processes are executed in parallel: 5.1A) Object Recognition by Visual Information: First, in SPE-256 a manipulated object is extracted in the image plane using an algorithm called Self Windowing which is realized by utilizing features of high-speed vision. Next some image features are calculated from the extracted image in the DSP as the center of the image and the angle of rotation of the image. Lastly, some 3-D geometrical parameters are calculated in the DSP as the 3-D position and the orientation


(role, pitch, and yow angles) of the object and the object size. Then the object shape is detected based on these parameters. In the present configuration we assume that the motion of an object is limited to a constraint plane given beforehand. Using parameters of both the plane and image features, 3-D parameters are calculated. 5.2B) Target Tracking Control of Active Vision In this system the objective of active vision control is to acquire reliable object information by keeping an object in sight. This is realized by using feedback control using the image feature which is a function of 1) The control input to the active vision. 2) The objective position on the image. 3) The joint angle vector of the Active Vision.




5.2) PROCESSING FOR ARM On the DSP for arm control, two sensory feedback controls are executed: 5.2A) Tracking control to object motion. The objective of tracking control is cancellation of the object motion by maintaining the relative position between the hand and the object. In the present configuration it is realized as a servo control so that the relative position error is kept to zero on the constraint plane shown in Figure-5.2A. 5.2B) Reaching control to the grasping position. The objective of reaching control is to reach the hand to the object by control of the relative position. In the present configuration this is controlled in the direction orthogonal to the constraint plane, as shown in Figure-5.2B. In this control the objective trajectory of the hand is given beforehand by integrating tracking motion and reaching motion which is a function of: 1) The control input to the arm. 2) The joint angle vector of the arm.


3) The position and the orientation of the object observed by vision. 4) The position and the orientation of the hand obtained by haptic sensors. 5) The objective trajectory for reaching. 6) Force feedback of the wrist force/torque sensor for compliance control. 5.3) PROCESSING FOR HAND In this system the objective of grasping control is to fix the object with the hand for manipulation. Using the compliance control method the hand control depends on the control input to the hand servo, the joint angle vector of the hand, the joint torque vector, the objective trajectory for grasping which is planned according to reaching motion. Furthermore, according to the object shape, pre-shaping motion is executed to set the appropriate hand shape for grasping. In the present configuration the grasping shape is changed by distinguishing a circle and a rectangle in the 2D image-plane, as shown in Figure-5.3A.






The experimental result is shown in Figure 6.1 & 6.2 as a continuous sequence of pictures. All sensory feedback is executed in parallel according to the object motion at high speed: tracking motion of the active vision, tracking and reaching motion of the arm, and grasping motion of the hand. In Figure-6.1 a close-up view of the same motion is shown. In this figure tracking is executed from 0:0ms to 0:5ms and both reaching and grasping motion start at 0:5ms and all motion is completed at 0:8ms. Then in Figure-6.2 a close-up view of the grasping motion of a spherical object is shown. It is shown that the shape of the hand is changed to a suitable shape for grasping of a sphere. In Figure-6.3 the trajectory of the hand is shown when grasping and releasing are alternately executed. In this figure, the Y axis position of the hand and the


object show the tracking motion, and the X axis position of the hand and objective trajectory for reaching motion show the reaching motion. This figure shows that both responsive tracking by visual feedback during the releasing phase and stable grasping by visual and force feedback during the grasping phase are realized. Collision Avoidance: To avoid occlusion problems, the hand-arm and the active vision are arranged opposite to each other and a manipulated object is placed in the middle position. And two planes, the beginning plane and the ending plane for reaching, is settled between the hand and the active vision, and reaching is planned so that the hand is transferred from the beginning plane to the ending plane. To make a trajectory which is difficult to estimate, a manipulated object is optionally moved by a human hand on the constraint plane. The collision is avoided by task switching from reaching subtask to object handling subtask as shown in figure-6.4.





Because these task switching are executed by real-time sensory feedback, they are executed at high-speed according to changes of an environment. Figure-6.5 shows on the Y axis the position of the hand and the object tracking motion, and the X axis position of the hand and objective trajectory for reaching motion and also handling and avoidance trajectory.

FIGURE-6.5: FEEDBACK RESPONSE OF THE POSITION FOR COLLISION AVOIDANCE. In these experiments, because the object is moved by a human hand, its trajectory is irregular and difficult to predict. Using the speed of the sensory feedback this problem is solved.As a result, grasping responsive to dynamic changes of object motion is realized. Various types of application for this system to realize responsive and flexible manipulation in the real world environment are been developed. Some of the possible applications of the experimental system can be simple pick and place operation involving separation of useful objects from scrap, performing functions such as welding, materiel handling in industries and chemical laboratories, robot arm for people who have no arms, completing everyday homework, complex assembly operations of automobiles, construction of walls, ground leveling, etc.


1) 2) Intelligent systems can learn & act accordingly. The main conclusion that can be drawn from an analysis of the application of Machine Learning to Robotics is that a successful employment of learning techniques on all levels of robot control is not possible without deeply revising the design criteria that are usually underlying the robot control system. In particular, it is necessary to identify both the tasks of the learning system and the tasks of the robot first and to design an architecture being able to host both the learning and the performance components afterwards. 3) Learning techniques on all levels of robot control is not possible without deeply revising the design criteria. Grasping is realized when decomposed and each subtask is executed in parallel. 4) It can be found that providing good examples and a good interface between the learning and the performance components is crucial for 5) success. subtasks

In short, besides the technological challenges of mobile robotics fundamental sensor-motor competences, robot navigation and application-oriented capabilities the scientific challenge is to move mobile robotics from a discipline of empirical practice towards a precise science.



INSPIRATON: Prof. Dr. Bernhard Scholkpf, Director of Max-Plank, Institute for Biologische Kybernetik, Germany. URL: bs@tuebingen.mpg.de

Journal of Machine Learning, August 2004, Last Three Pages, IIT Bombay Library, Recent section.

Computer Aided Manufacturing by John. H. Powers. Jr. McGraw Hill Int. Edition-1987, P-92 to 109, P- 164 to 216.

Franklin Watts Computer Library, Artificial Intelligence, Fred D Ignozio & Alen Wold, Mushroom Prod. London, P -1 to 15. Industrial Robotics, A Handbook of Automated Systems Design by Ken. Stonecipher- Hayden Book Company, Chapter-16, P-189 to 194, 166, 331 to 346. Machine & Human Learning by Yves Kodratoff & Alan Hutchinson, Advances in European Research-1989, P-313 to 328, 199 to 271.

High Speed Grasping using Visual & Force Feedback, University of Tokyo, Japan. URL: http://www.k2.t.u-tokyo.ac.jp/~sfoc/