Automating Deformable Gasket Assembly

Simeon Adebola*1, Tara Sadjadpour*1, Karim El-Refai*1, Will Panitch1, Zehan Ma1, Roy Lin1, Tianshuang Qiu1 Shreya Ganti1, Charlotte Le1, Jaimyn Drake1, Ken Goldberg1
1The AUTOLab at UC Berkeley
*Equal Contribution

The Gasket/Channel Detection box shows gasket segmentation (above) and channel segmentation (below). The Template Matching box shows the three templates for the curved, straight and trapezoid channel. The Straight/Curved Actuation box shows selection and actuation strategies for the straight and curved channels: (a) is Unidirectional insertion, (b) is Binary search insertion, and (c) is Binary+ insertion. The colors on the channels represent the locations the robot attempts to place and press the gasket into while the numbers represent the order they are placed and pressed. Endpoints are green, midpoints are pink, half-points are blue and the quartile-points are cyan. The arrows indicate the direction(s) of the slide(s). For the trapezoid channel, we treat each segment of the trapezoid as an instance of the straight channel. In the unidirectional approach (d) we process each segment in a counterclockwise manner, starting at the blue segment. For hybrid and binary (e), we evaluate the blue segment, then the cyan segments, and finally the red segment. The learned policy proceeds directly from the initial state to actuation (f). The Final State box shows the final assembled gasket.

Abstract

In Gasket Assembly, a deformable gasket must be aligned and pressed into a narrow channel. This task is common for sealing surfaces in the manufacturing of automobiles, appliances, electronics, and other products. Gasket Assembly is a long-horizon, high-precision task and the gasket must align with the channel and be fully pressed in to achieve a secure fit. To compare approaches, we present 4 methods for Gasket Assembly: one policy from deep imitation learning and three procedural algorithms. We evaluate these methods with 100 physical trials. Results suggest that the Binary+ algorithm succeeds in 10/10 on the straight channel whereas the learned policy based on 250 human teleoperated demonstrations succeeds in 8/10 trials and is significantly slower.

Video

Channel Types

We present three channel types: open straight, open curved and closed trapezoid. See the link at page top for CAD files.

Each channel is shown here with a gasket fully inserted. The straight channel (A) and the curved channel (B) are both open-ended channels whereas the trapezoid channel (C) is closed. This means that for all channels, the gasket endpoints (g0, g1) and channel endpoints (c0, c1) lie nearly on top of each other, but in the trapezoid case, c0 and c1 also lie nearly on top of each other




Physical Trials & Results

We carry out 100 physical trials: 10 for the learned diffusion policy on the straight channel, 10 for each procedural algorithm on each channel: straight, curved and trapezoid.

Alignment Results and Insertion Results for all four approaches: learned diffusion mode, unidirectional, binary search and Binary+.

Learned Diffusion Policy (Success)

The learned diffusion policy successfully carrying out the task on the straight channel. Trial time was for a maximum of 10 minutes

Learned Diffusion Policy (Failure)

The learned diffusion policy failing to complete the task on the straight channel. Trial time was for a maximum of 10 minutes

Unidirectional Algorithm (Success)

The unidirectional algorithm successfully carrying out the task on the straight channel. Trial time was for a maximum of 3 minutes 30 seconds.

Unidirectional Algorithm (Failure)

The unidirectional algorithm failing to complete the task on the straight channel. Trial time was for a maximum of 3 minutes 30 seconds.

Binary Search Algorithm (Success)

The binary search algorithm successfully completes the task on the curved channel. Trial time was for a maximum of 3 minutes 30 seconds.

Appendix

  1. Human Demonstrations

    Human demonstrations are collected as follows:

    1. The channel is fixed in place horizontally across the workspace, separating the workspace into a lower and upper section.
    2. The gasket is randomly dropped in either the lower or upper section so that it does not overlap itself and does not touch the channel.
    3. The midpoint of the gasket is grasped and placed on top of the midpoint of the channel. The gripper then presses the gasket down into the channel.
    4. One endpoint of the channel is chosen arbitrarily. The gasket is placed on top of the selected endpoint of the channel and pressed down into the channel.
    5. The remaining endpoint of the gasket is then placed on top of the other endpoint of the channel and pressed to insert it into the channel.
    6. The gripper is moved to the quartile points (the order of the quartile points the gripper goes to is chosen arbitrarily) and pressed down on the gasket such that at those points the gasket is inserted into the channel.
    7. The gripper goes to the 'eighth' points (again the order of the points the gripper goes to is chosen arbitrarily) and presses down on the gasket such that at those points the gasket is inserted into the channel.
    8. The gripper goes to the midpoint of the gasket, moves down slowly to the channel surface such that the gripper touches the channel surface, and moves horizontally with no vertical movement towards one of the endpoints of the channel (chosen arbitrarily). The gripper returns to the midpoint of the channel and repeats this motion towards the other endpoint of the channel. This 8-step procedure is repeated for each human demonstration.

  2. Experimental Evaluation Metrics Breakdown

    After the robot execution has terminated, a human judge visually rates performance into one of four alignments categories, as follows:

    1. 0% - 25%: A major alignment failure, in which the robot has successfully aligned less than 25% of the gasket with the channel.
    2. 25% - 50%: A partial alignment failure, in which between 25% and 50% of the gasket has been successfully aligned.
    3. 50% - 75%: A partial alignment success, in which between 50% and 75% of the gasket has been properly aligned.
    4. 75% - 100%: A full alignment success, in which the robot has properly aligned at least 75% of the gasket length with the channel.

    Similarly, a human judge visually rates performance into one of four insertion categories, as follows:

    1. 0% - 25%: A major insertion failure, in which less than 25% of the gasket is inserted into the channel.
    2. 25% - 50%: A partial insertion failure, in which between 25% and 50% of the gasket is inserted.
    3. 50% - 75%: A partial insertion success, in which between 50% and 75% of the gasket is inserted.
    4. 75% - 100%: A full insertion success, in which at least 75% of the gasket length is inserted.
  3.  

    Figure 4 shows qualitative results from the trials of the three analytical algorithms in increasing order of success.


    Evaluation Metrics

    Fig. 4: Evaluation Metric Examples. We provide examples for all four categories of the alignment and insertion evaluation metrics discussed in Section V-D. We show the final gasket and channel states after the robot attempts gasket assembly. For alignment we only consider the view from the overhead camera to determine alignment between the gasket and channel. To determine the snug fit of the insertion, we consult both the overhead view (top row) and the front view (bottom row), because (f), for example, shows how a gasket that is aligned with the channel can have poor insertion.

Citation

If you use this work or our dataset or find it helpful, please consider citing: (bibtex)

  @inproceedings{gasketassembly2024,
    author    = {Adebola, Simeon* and Sadjadpour, Tara* and El-Refai, Karim* and Panitch, Will and Ma, Zehan 
                and Lin, Roy and Qiu, Tianshuang and Ganti, Shreya and Le, Charlotte and Drake, Jaimyn and Goldberg, Ken},
    title     = {Automating Deformable Gasket Assembly},
    journal   = {CASE},
    year      = {2024},
}