Experimental Evaluation of the Learning Cobot - Northwestern University
Experimental Evaluation of the Learning Cobot ES Boy1, E Burdet1,2, CL Teo1 and JE Colgate3 1: Dept of Mechanical Engineering, 2: Division of Bioengineering, National University of Singapore
[email protected] http://www.guppy.mpe.nus.edu.sg/~eburdet 3: LIMS, Dept of Mechanical Engineering, Northwestern University, USA
Abstract. Cobots can assist a human by mechanically constraining motion to software-defined guide paths. The Learning Cobot we introduced in [1] provides simple design tools for ergonomic guide paths. Walk-ThroughProgramming (WTP) lets the operator trace out a path in free mode and use a B-spline approximation of this path as a guide path for subsequent movements. Guide paths can also be designed by placing B-spline control points by mouse in a dedicated graphical user interface (GUI). This paper presents psychophysical experiments to evaluate these tools. User satisfaction was our evaluation criterion, as it encompasses analytical criteria like effort and motion jerkiness but is finer. The results demonstrate the efficacy of the Learning Cobot, and the necessity of using the WTP and GUI together. With the WTP the operator can design ergonomic guide paths well, i.e., with haptic feedback alone, but operators prefer also to have the GUI to visualize the path. Using the Learning Cobot, operators can gradually learn ergonomic guide paths and adapt to changes in the environment or in the task.
1
Introduction
Placing a window or a car door into its frame is a difficult operation, requiring simultaneous control of six degrees of freedom (DOF). Guide ways make it easy to move it into the frame using only translational push and pull. Cobots (Collaborative robots) developed Northwestern University [2-5] use a similar strategy. They realize software-defined mechanical guides along desired paths or surfaces. They are passive in that they do not generate motion, but only steer the wheels to direct it [2]. Forces perpendicular to the wheel headings are balanced by friction, constraining motion to the heading direction. To illustrate the cobot concept and its motion modes, we briefly describe the Scooter cobot used in the experiments reported here. The Scooter (Fig 1a) is a triangular vehicle moving on a plane, with a steerable wheel at each corner. In Free Mode (FM), the wheels turn like casters to align with the force exerted by the operator. The cobot behaves as if it had 3 DOF (i.e., motion in planar position and orientation). This force is measured by a force-torque sensor mounted on the handle. In Guided Mode (GM), each wheel is steered by a motor to follow a guide path coded in software.
40
E.S. Boy et al
force torque sensor
goal
obstacle obstacle i start
one of the three glidewheels for odometry
obstacl Fig. 1. (a): the Scooter cobot on which the Learning Cobot strategy was tested. (b): Subjects performing psychophysical experiments.
Cobot kinematics, design and control have been investigated extensively, and several planar and spatial cobots have been realized for the automotive industry [2-5]. This paper studies path-planning issues. It investigates how an ergonomic guide path adapted to a particular user and task can be designed and modified with changes in the environment. Optimal path planning has been investigated extensively in robotics [8-10], and [11] has proposed using an ergonomic cost function to derive guide paths for cobots. Since this approach requires knowledge of the environment, it is limited by the complexity and high cost of sensor processing and by sub-optimal sensor properties. Moreover, optimal guide paths vary from task to task and from person to person. It appears difficult to find a cost function suitable to every task, cobot and operator.
Experimental Evaluation of the Learning Cobot
41
Walk Through Programming bb
aa
control points
free mode
cubic B-spline
Dedicated Graphical User Interface c
control point placed using a mouse
Fig. 2. Guide path design tools. In Walk Through Programming (WTP), a path traced in free mode (a) is least-squared fitted using B-splines coded by discrete control points. (b) Alternatively, the control points are placed in a GUI using a mouse (c). The B-spline is used as guide path for subsequent movements.
The Learning Cobot introduced in [1] needs neither external sensors nor mathematical optimization to design a guide path. It relies on the well-developed sensing and inference capabilities of the human operator. The Learning Cobot uses two simple design tools. With Walk Through Programming (WTP), the operator traces a path in free mode; this is approximated by a B-spline path which can guide subsequent movement (Figs. 2a,b). Guide paths can also be defined (Fig. 2c) by placing B-spline control points in the dedicated Graphical User Interface (GUI). This paper reports experiments performed to evaluate the Learning Cobot. To evaluate the various path designing tools, we analyzed movements and user satisfaction in representative conditions. This was complemented by a questionnaire filled out by the subjects after the experiments. The results show the utility and complementarity of WTP and GUI.
42
E.S. Boy et al
2
Methods
The experiments described in sections 2.1 – 2.4 were approved by the Institutional Review Board of Northwestern University. The subjects were students without known motor disabilities, informed about the experiment, who gave their consent prior to participation. 2.1
Using WTP and the GUI without training
Five subjects (mean age 24, with standard deviation 2) performed the first experiment. They were first trained to drive the Scooter cobot in free mode in the environment of Fig. 7 until they could do the pin-into-hole task without colliding with obstacles. They did not experience motion guided by the cobot, the WTP or the GUI. Tests were performed in three representative task environments. In each, they performed six pin-into-hole trials in free mode. They were required to grade these movements according to their satisfaction i.e. grade 1 = fail, grade 2 = bad, grade 3 = okay, grade 4 = good and grade 5 = optimal. They were not informed that the paths might be used to guide subsequent movements. The next day, the same subjects were asked to move in guided mode along the paths that in free mode they had graded highest and lowest. They then grade these guided movements according to their satisfaction as mentioned above. Wilcoxon rank-sum tests [12] were used to compare satisfaction for movements in free and guided modes. 2.2
Training the WTP and GUI
Seven other subjects (mean age 24 with standard deviation 2) performed the second experiment. They were first trained to drive the Scooter cobot in three representative environments as described in Fig. 2 of [6]. They then learned to use the WTP and GUI to design guide paths in the environment of Fig. 7. To learn the WTP, the subjects were told to experience the resulting guidance extensively, and to modify a guide path until satisfied. On average, subjects realized 3 paths (least 2, most 5). When using the GUI, the subjects were told to note the regions they were unsatisfied with and to improve the path locally (Fig. 7) until satisfied. Each subject realized an average of 3 improvements (least 2, most 5), i.e., 4 trials.
Experimental Evaluation of the Learning Cobot
43
Fig. 3. Six paths traced in free mode were shown in random order to the subject to select a good guideway.
2.3
Path Definition
The next day, the subjects had to design guide paths in the two environments of Figs. 2 and 3. They had first to design six paths with the WTP, i.e., to trace them in free mode. They were told to grade each try on the scale of 1 – 5 similar to the grading used in section 2.1, i.e., to estimate how well the resulting path would guide motion. They then had to define a guide path with 16 control points using the GUI. To test how well guide paths can be evaluated visually, the paths traced using WTP were presented in random order (Fig. 3). The subjects graded these six paths. Additionally, five (of the seven) subjects were asked to move in GM along the best path defined with WTP and improve it using the GUI. On average, each improved the path twice (minimal 2, maximal 3), i.e., made three trials. 2.4
Guide paths Evaluation
Five guide paths were tested on the next day: • A: The highest-graded defined in WTP (free mode); • B: The lowest-graded defined in WTP; • C: The guide path defined in the GUI; • D: The path selected visually; • E: Path A changed in the GUI after tracing it in GM (Two subjects did not have E.) All subjects had to move the Scooter along these paths and grade its movements on scale 1 – 5 similar to that presented in section 2.1. In each environment, they were told to: i) make at least three series of movements along these five (for two subjects, four) guide paths, ii) refine the grading after each series, and iii) continue to try the paths until they had stable grading. The first series presented the paths in the order A,
44
E.S. Boy et al
B, C, D, E. In subsequent series the order could be modified at will, for example to compare two similar guide paths by examining them consecutively. Only final grades were analyzed, and more than one of the guide paths could obtain the highest grade. On average, the subjects did four movement series (most 4, least 6). Wilcoxon rank-sum tests compared performance with the various guide path design tools. For statistical treatment, grades were coded from 1 for “fail” to 5 for “optimal”. After the experiment, the subjects had to complete a questionnaire analyzing which features they considered important for good guide paths (Table 1), and how they liked the different design methods (Fig. 6). Analyses were performed on the answers, grading the ratings from 1 to 5 for the questions of Figs. 6 a,b,c, and assigning 1 to “yes” and -1 to “no” in the question of Fig. 6d. Three different quantitative criteria, defined in [6], were used to analyze the movements: i) Effort ε (τ ) to maneuver the Scooter during a whole movement; ii) The frequency content of torque σ (τ ) , reflecting movement smoothness; iii) The number of back-and-forth corrections, indicative of feedback control.
3.
Results
3.1
Good free paths may not be good guides
In the first experiment, the subjects had no prior knowledge of guided movements. They graded paths traced in FM without knowing that these would be used as guideways. As a result, the best paths traced in FM were not always selected as best path in GM (p>0.15). Actually, more than 40% of the paths preferred in GM were graded lowest in FM. Therefore, in the second experiment, the subjects were required to train in GM and experience how paths designed with the WTP and GUI guide motion, before starting to use and evaluate these design tools. 3.2
Evaluation of the WTP and GUI
Fig. 4 shows how the subjects graded the five different paths along which they tested motion guidance. We observe that: 1. The guide path corresponding to the highest-graded path traced in FM (‘WTP best’) was almost always preferred to the lowest-graded (‘WTP worst’) (p
[email protected] http://www.guppy.mpe.nus.edu.sg/~eburdet 3: LIMS, Dept of Mechanical Engineering, Northwestern University, USA
Abstract. Cobots can assist a human by mechanically constraining motion to software-defined guide paths. The Learning Cobot we introduced in [1] provides simple design tools for ergonomic guide paths. Walk-ThroughProgramming (WTP) lets the operator trace out a path in free mode and use a B-spline approximation of this path as a guide path for subsequent movements. Guide paths can also be designed by placing B-spline control points by mouse in a dedicated graphical user interface (GUI). This paper presents psychophysical experiments to evaluate these tools. User satisfaction was our evaluation criterion, as it encompasses analytical criteria like effort and motion jerkiness but is finer. The results demonstrate the efficacy of the Learning Cobot, and the necessity of using the WTP and GUI together. With the WTP the operator can design ergonomic guide paths well, i.e., with haptic feedback alone, but operators prefer also to have the GUI to visualize the path. Using the Learning Cobot, operators can gradually learn ergonomic guide paths and adapt to changes in the environment or in the task.
1
Introduction
Placing a window or a car door into its frame is a difficult operation, requiring simultaneous control of six degrees of freedom (DOF). Guide ways make it easy to move it into the frame using only translational push and pull. Cobots (Collaborative robots) developed Northwestern University [2-5] use a similar strategy. They realize software-defined mechanical guides along desired paths or surfaces. They are passive in that they do not generate motion, but only steer the wheels to direct it [2]. Forces perpendicular to the wheel headings are balanced by friction, constraining motion to the heading direction. To illustrate the cobot concept and its motion modes, we briefly describe the Scooter cobot used in the experiments reported here. The Scooter (Fig 1a) is a triangular vehicle moving on a plane, with a steerable wheel at each corner. In Free Mode (FM), the wheels turn like casters to align with the force exerted by the operator. The cobot behaves as if it had 3 DOF (i.e., motion in planar position and orientation). This force is measured by a force-torque sensor mounted on the handle. In Guided Mode (GM), each wheel is steered by a motor to follow a guide path coded in software.
40
E.S. Boy et al
force torque sensor
goal
obstacle obstacle i start
one of the three glidewheels for odometry
obstacl Fig. 1. (a): the Scooter cobot on which the Learning Cobot strategy was tested. (b): Subjects performing psychophysical experiments.
Cobot kinematics, design and control have been investigated extensively, and several planar and spatial cobots have been realized for the automotive industry [2-5]. This paper studies path-planning issues. It investigates how an ergonomic guide path adapted to a particular user and task can be designed and modified with changes in the environment. Optimal path planning has been investigated extensively in robotics [8-10], and [11] has proposed using an ergonomic cost function to derive guide paths for cobots. Since this approach requires knowledge of the environment, it is limited by the complexity and high cost of sensor processing and by sub-optimal sensor properties. Moreover, optimal guide paths vary from task to task and from person to person. It appears difficult to find a cost function suitable to every task, cobot and operator.
Experimental Evaluation of the Learning Cobot
41
Walk Through Programming bb
aa
control points
free mode
cubic B-spline
Dedicated Graphical User Interface c
control point placed using a mouse
Fig. 2. Guide path design tools. In Walk Through Programming (WTP), a path traced in free mode (a) is least-squared fitted using B-splines coded by discrete control points. (b) Alternatively, the control points are placed in a GUI using a mouse (c). The B-spline is used as guide path for subsequent movements.
The Learning Cobot introduced in [1] needs neither external sensors nor mathematical optimization to design a guide path. It relies on the well-developed sensing and inference capabilities of the human operator. The Learning Cobot uses two simple design tools. With Walk Through Programming (WTP), the operator traces a path in free mode; this is approximated by a B-spline path which can guide subsequent movement (Figs. 2a,b). Guide paths can also be defined (Fig. 2c) by placing B-spline control points in the dedicated Graphical User Interface (GUI). This paper reports experiments performed to evaluate the Learning Cobot. To evaluate the various path designing tools, we analyzed movements and user satisfaction in representative conditions. This was complemented by a questionnaire filled out by the subjects after the experiments. The results show the utility and complementarity of WTP and GUI.
42
E.S. Boy et al
2
Methods
The experiments described in sections 2.1 – 2.4 were approved by the Institutional Review Board of Northwestern University. The subjects were students without known motor disabilities, informed about the experiment, who gave their consent prior to participation. 2.1
Using WTP and the GUI without training
Five subjects (mean age 24, with standard deviation 2) performed the first experiment. They were first trained to drive the Scooter cobot in free mode in the environment of Fig. 7 until they could do the pin-into-hole task without colliding with obstacles. They did not experience motion guided by the cobot, the WTP or the GUI. Tests were performed in three representative task environments. In each, they performed six pin-into-hole trials in free mode. They were required to grade these movements according to their satisfaction i.e. grade 1 = fail, grade 2 = bad, grade 3 = okay, grade 4 = good and grade 5 = optimal. They were not informed that the paths might be used to guide subsequent movements. The next day, the same subjects were asked to move in guided mode along the paths that in free mode they had graded highest and lowest. They then grade these guided movements according to their satisfaction as mentioned above. Wilcoxon rank-sum tests [12] were used to compare satisfaction for movements in free and guided modes. 2.2
Training the WTP and GUI
Seven other subjects (mean age 24 with standard deviation 2) performed the second experiment. They were first trained to drive the Scooter cobot in three representative environments as described in Fig. 2 of [6]. They then learned to use the WTP and GUI to design guide paths in the environment of Fig. 7. To learn the WTP, the subjects were told to experience the resulting guidance extensively, and to modify a guide path until satisfied. On average, subjects realized 3 paths (least 2, most 5). When using the GUI, the subjects were told to note the regions they were unsatisfied with and to improve the path locally (Fig. 7) until satisfied. Each subject realized an average of 3 improvements (least 2, most 5), i.e., 4 trials.
Experimental Evaluation of the Learning Cobot
43
Fig. 3. Six paths traced in free mode were shown in random order to the subject to select a good guideway.
2.3
Path Definition
The next day, the subjects had to design guide paths in the two environments of Figs. 2 and 3. They had first to design six paths with the WTP, i.e., to trace them in free mode. They were told to grade each try on the scale of 1 – 5 similar to the grading used in section 2.1, i.e., to estimate how well the resulting path would guide motion. They then had to define a guide path with 16 control points using the GUI. To test how well guide paths can be evaluated visually, the paths traced using WTP were presented in random order (Fig. 3). The subjects graded these six paths. Additionally, five (of the seven) subjects were asked to move in GM along the best path defined with WTP and improve it using the GUI. On average, each improved the path twice (minimal 2, maximal 3), i.e., made three trials. 2.4
Guide paths Evaluation
Five guide paths were tested on the next day: • A: The highest-graded defined in WTP (free mode); • B: The lowest-graded defined in WTP; • C: The guide path defined in the GUI; • D: The path selected visually; • E: Path A changed in the GUI after tracing it in GM (Two subjects did not have E.) All subjects had to move the Scooter along these paths and grade its movements on scale 1 – 5 similar to that presented in section 2.1. In each environment, they were told to: i) make at least three series of movements along these five (for two subjects, four) guide paths, ii) refine the grading after each series, and iii) continue to try the paths until they had stable grading. The first series presented the paths in the order A,
44
E.S. Boy et al
B, C, D, E. In subsequent series the order could be modified at will, for example to compare two similar guide paths by examining them consecutively. Only final grades were analyzed, and more than one of the guide paths could obtain the highest grade. On average, the subjects did four movement series (most 4, least 6). Wilcoxon rank-sum tests compared performance with the various guide path design tools. For statistical treatment, grades were coded from 1 for “fail” to 5 for “optimal”. After the experiment, the subjects had to complete a questionnaire analyzing which features they considered important for good guide paths (Table 1), and how they liked the different design methods (Fig. 6). Analyses were performed on the answers, grading the ratings from 1 to 5 for the questions of Figs. 6 a,b,c, and assigning 1 to “yes” and -1 to “no” in the question of Fig. 6d. Three different quantitative criteria, defined in [6], were used to analyze the movements: i) Effort ε (τ ) to maneuver the Scooter during a whole movement; ii) The frequency content of torque σ (τ ) , reflecting movement smoothness; iii) The number of back-and-forth corrections, indicative of feedback control.
3.
Results
3.1
Good free paths may not be good guides
In the first experiment, the subjects had no prior knowledge of guided movements. They graded paths traced in FM without knowing that these would be used as guideways. As a result, the best paths traced in FM were not always selected as best path in GM (p>0.15). Actually, more than 40% of the paths preferred in GM were graded lowest in FM. Therefore, in the second experiment, the subjects were required to train in GM and experience how paths designed with the WTP and GUI guide motion, before starting to use and evaluate these design tools. 3.2
Evaluation of the WTP and GUI
Fig. 4 shows how the subjects graded the five different paths along which they tested motion guidance. We observe that: 1. The guide path corresponding to the highest-graded path traced in FM (‘WTP best’) was almost always preferred to the lowest-graded (‘WTP worst’) (p
