Academia.eduAcademia.edu

Figure 5 – uploaded by Lorenzo Panchetti

See full PDFdownloadDownload figure
Fig. 6. The cobot interaction is evaluated in terms of safety, easiness to teach, entertainment, goal achievement, and task completion. Three answers have been collected for results in the chart on the right, featuring two different users. A user participated in the activities twice, its second feedback is shown in dashed lines. THIS TABLE PRESENTS ERROR METRICS FOR THE FACTORY SCENARIO—30 RUNS ARE PERFORMED PER TASK.

Figure 6 The cobot interaction is evaluated in terms of safety, easiness to teach, entertainment, goal achievement, and task completion. Three answers have been collected for results in the chart on the right, featuring two different users. A user participated in the activities twice, its second feedback is shown in dashed lines. THIS TABLE PRESENTS ERROR METRICS FOR THE FACTORY SCENARIO—30 RUNS ARE PERFORMED PER TASK.

Related Figures (5)

Fig. 1. Flowchart detailing the method of TEAM. A set of demonstrations is recorded by a user and the motion is generalized through GMR. Then, the system is modelled as a set of damped spring models able to generalize the motion to different start and end joint angles. All parameters of the system are automatically optimized and no expert knowledge is needed.
Fig. 1. Flowchart detailing the method of TEAM. A set of demonstrations is recorded by a user and the motion is generalized through GMR. Then, the system is modelled as a set of damped spring models able to generalize the motion to different start and end joint angles. All parameters of the system are automatically optimized and no expert knowledge is needed.
Fig. 2. The ABB GoFa cobot is facing the test elevator door used for evaluation of TEAM in a maintenance scenario. sood metric of trajectory reproduction since it measures devi- ations from the set of demonstrations instead of adaptability. Since MSE penalizes the distance from a reference signal, MSE would assign a lower score to a successful trajectory adapting to different starting and end conditions, than to a trajectory close to the regression. In our work, we visually evaluate the reproductions and consider a reproduction failed if 1) the movement does not look like the demonstration, 2) the cobot collides with any element of the environment, 3) the task fails—e.g. the tool does not keep the door open or the brush is not in its support.
Fig. 2. The ABB GoFa cobot is facing the test elevator door used for evaluation of TEAM in a maintenance scenario. sood metric of trajectory reproduction since it measures devi- ations from the set of demonstrations instead of adaptability. Since MSE penalizes the distance from a reference signal, MSE would assign a lower score to a successful trajectory adapting to different starting and end conditions, than to a trajectory close to the regression. In our work, we visually evaluate the reproductions and consider a reproduction failed if 1) the movement does not look like the demonstration, 2) the cobot collides with any element of the environment, 3) the task fails—e.g. the tool does not keep the door open or the brush is not in its support.
THIS TABLE PRESENTS THE TRAINING TIME AND ERROR METRICS FOR THE MAINTENANCE TASK— 10 RUNS ARE PERFORMED PER TASK. FAILURES CORRESPOND TO THE NUMBER OF TIME THE ROBOT FAILED EITHER BY NOT REPRODUCING THE MOVEMENT OR BY NOT COMPLETING THE TASK. THE TABLE SHOWS THE JENSEN-SHANNON DIVERGENCE REPEATABILIT‘ OVER 50 DIFFERENT RUNS. THE NUMBER OF GMM COMPONENTS ee ee ere
THIS TABLE PRESENTS THE TRAINING TIME AND ERROR METRICS FOR THE MAINTENANCE TASK— 10 RUNS ARE PERFORMED PER TASK. FAILURES CORRESPOND TO THE NUMBER OF TIME THE ROBOT FAILED EITHER BY NOT REPRODUCING THE MOVEMENT OR BY NOT COMPLETING THE TASK. THE TABLE SHOWS THE JENSEN-SHANNON DIVERGENCE REPEATABILIT‘ OVER 50 DIFFERENT RUNS. THE NUMBER OF GMM COMPONENTS ee ee ere
Fig. 3. Bayesian optimization times for each task over 50 runs. Fig. 3. Bayesian optimization times for each task over 50 runs. grid search is between a factor of 50 to 100 slower than BO due to the larger number of evaluations computed. COMPARISON BETWEEN GRID SEARCH (GS) AND BAYESIAN OPTIMIZATION (BO). STATISTICS ARE COMPUTED OVER 50 RUNS
Fig. 3. Bayesian optimization times for each task over 50 runs. Fig. 3. Bayesian optimization times for each task over 50 runs. grid search is between a factor of 50 to 100 slower than BO due to the larger number of evaluations computed. COMPARISON BETWEEN GRID SEARCH (GS) AND BAYESIAN OPTIMIZATION (BO). STATISTICS ARE COMPUTED OVER 50 RUNS
Related topics:
Computer Science
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved