Notes on troubleshooting in the event of faulty performance:
- Make sure that there are no interfering factors in the episodes used (fingers in the image, strong reflections or variable, external light sources, etc.).
- If the tool and thus the TCP is not as centered as possible in the camera image, MIRAI's ability to learn may be impeded.
- For rotations, train MIRAI to use the shortest path (the smaller angle) if possible.
- Make sure that enough variations have been recorded for MIRAI (all angles, all possible approach positions)
- Have "Edge Cases" (target object only partially in the picture) and "Recovery Trials" (gripper already slightly past the target object) also been trained to support and stabilize MIRAI in all potential situations?
- Make sure that neither cables nor hoses create additional forces or moments on the tool.
For the "Motion skill" training variant:
- Note that when training an episode -that is, between starting and stopping the recording- no stopping of the motion is generated. If this is not completely possible, the total percentage of such episodes should not exceed 5% (e.g. 200 episodes of which no more than 10 episodes have a stop in the motion).
- "Recovery trials" increase precision, while edge cases make the skill robust by helping to improve overall stability with respect to possible variances.
In addition, it is possible that insufficient variances have been trained. Please also have a look at the following video.
Contact Customer Support if you need further assistance.