Not all motion skills will work perfectly in the first round of testing. For example, you may notice that robot movements are not reaching the desired target positions or may overly depend on a particular lighting situation or object appearance. If you identify any unexpected behaviors or the skill doesn't seem to be responding as intended, there are steps you can take to improve the performance.
The most common reason for under-performing skills is an insufficient number of training episodes. When working with a new skill, it is important to show Mirai all possibilities within the field of view so that it can learn and understand the targets and what you want it to focus on. This includes but is not limited to:
- Showing edges of the workpiece
- Showing outer edges of the field of view (keep in mind that part of the workpiece must still be in view)
- Showing rotated views that include any possible axis movements
- Showing perspectives higher above the workpiece (Z height)
- Showing corrective positions (recovery episodes)
Quick Links:
Best Practices
Hand Placement
Place your hand just below the force/torque sensor along the robot arm. An episode will fail if your hand falls into the field of view of the cameras - be very mindful of this while rotating!
Holding the wrist near the F/T sensor ensures that even tiny motions from the tip of the tool/gripper are properly captured in the training. Reaching from higher will be more difficult to steer and may result in more jerky or unpredictable pathing.
Pick a Consistent Strategy and Stick to It
For best results, avoid showing the AI two different paths to reach the same target. When you show the system more than one path, it can't decide which way is best and ultimately will try to blend these two paths together creating a "middle path" and fail. We recommend moving along the shortest path from the starting position to target position, whenever possible.
Move With a Consistent Speed
Mirai will pick up any inflections in speed along a path or during a motion. For example, when moving the tool, it accidentally moved too quickly in the beginning and then changed to a slower rate towards the end of the episode - the AI will copy this motion identically. If your desired result is to maintain a smooth and consistent rate of speed, this will need to be properly demonstrated in your training episodes.
Be ready to 'Stop recording'
When setting up your workstation, tablet placement will help your trainings run much more efficiently and lead to fewer failed episodes. We recommend placing the tablet within reach so that you can quickly tap "stop recording" as soon as your episode is complete. This will be especially important when training motion skills. Any delay at the end of the episode will be mimicked when performing the skill later. For this reason, make sure to press the 'Stop recording' button the moment the task is done; do not wait.
Motion Skill Method
Begin by placing the TCP (tool center point) at a start position of your choosing (this will need to be the same for each recording). Enter the recording loop, calibrate the force/torque sensor, and proceed to the ‘Start recording’ screen. Immediately, guide the robot to perform the desired motion (Any delay/lag will be imitated in the skill). The trainer will need to demonstrate about 200 repetitions of the given task in all possible variations of position and rotation. For a typical motion application, this takes about 2-5 hours, depending on the complexity of the task.
-
Press ‘start recording’.
-
Immediately, guide the robot to perform the desired motion
-
Move in a smooth, confident line from start to finish without pauses or detours.
-
Immediately press ‘stop recording’ after reaching the end of the desired motion.
For more in-depth instruction on Recording motion skills (Click here)
Recovery Episodes:
Some skills may encounter behaviors where the robot has moved to a position outside of the desired path. This could look like overshooting a target position, losing course unexpectedly, or if Mirai is unable to locate the target, and as a results steers in unexpected directions. “Recovery episodes,” In other words, show the robot how to correct a mistake.
To incorporate recovery skills:
- Begin by placing the robot in a starting position slightly outside the ideal path. For Example: In a task that involves touching a target object, start with the robot in a position beyond the target to simulate an overshooting error.
- Record the same as you normal would by guiding the robot back to the desired target/path/appearance and clearly demonstrate how to find it's way back. Depending on the complexity of your task, we recommend about a third of the recordings for a given skill be trained as recovery episodes.
The key to successful recovery training is to start these 'recovery recordings' with the robot already positioned in a point outside of the ideal path -- if the robot starts from the an expected start position, then the system will only learn what to do in perfect conditions.
Important: We do not recommend using recovery episodes to account for user error or to correct bad recordings. For example: If the camera is accidentally bumped during a recording, or if any adjustments/changes have been made to the camera angle, room lighting, or image focus, we recommend to delete these recordings and start over. The use of failed recordings will require additional unnecessary training time or a complete redo of the skill training from scratch.
Other Considerations:
Golden Rules for Motion Skills
Every motion and action that is required to solve the task has to be seen in each recorded episode. Try training without actually recording and check the camera image. Key points to note are:
- Your hands should not appear in the camera image at any point
- The end effector and the object should always be visible
Include all variance that you require for the skill to be robust against:
- Variations in the starting and end points of trajectories or the position of target objects in the task
- Changing background or moving objects in the background. (This is only relevant if the background is visible in the recorded episodes.)
- Differences in the color and/or shapes of the objects/working parts.
- Changes in the lighting conditions (e.g. - variations in daylight). Direct exposure to sunlight can negatively impact the stability of the skill and should be avoided if possible.
- Imprecisions and variations in gripping positions at the TCP
-
Ensure the robot arm is at the start position configured in skill setup by selecting the ‘send robot to reference position’ dialog
-
Make sure your hands are not visible in the camera's field of view
-
Be mindful not to push against the workpiece, the robot, or connected tools
-
Check that you not touching the robot when calibrating the F/T sensor using the on-screen button
-
Demonstrate variation by repositioning objects
-
Check the image preview to confirm the robot has a clear field of view (Once set, the following recordings must maintain the exact same field of view.
Substantial improvements are usually achieved within 50-200 additional recordings depending on skill type. Once episodes are finalized, make sure to upload your work to the cloud by pressing the ‘Start Cloud Training’ button. The status will change to ‘Cloud training in progress’ until the new skill is processed. Successful uploads will be shown under the "skill version" tab of each skill with a timestamp of the upload as well as version history.
Below is a visual aid that outlines the complete training process:
VIDEO TUTORIALS:
Training Tip: Keep the Target in the Image
Comments
0 comments
Article is closed for comments.