Automating Deformable Gasket Assembly

TL;DR

The paper tackles automating gasket assembly, a long-horizon, high-precision manipulation of a 1D deformable object inserted into predefined channels. It compares a diffusion-policy learned from 250 human demonstrations with three procedural baselines, under a 100-trial physical evaluation, on straight, curved, and trapezoid channels. Results show that the Binary+ procedural strategy achieves the strongest alignment and insertion performance across challenging geometries, while diffusion-based learning yields mixed success with notable recovery challenges. The work provides a replicable experimental setup, public code and CAD resources, and highlights the practical viability of structured procedural strategies for tight-tolerance DLO insertion tasks in industrial contexts.

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. Code, CAD models, videos, and data can be found at https://berkeleyautomation.github.io/robot-gasket/

Figures (4)

• Figure 1: Gasket Assembly Example. This is an example of the Binary+ procedural algorithm rollout on the curved channel. Initially, the gasket is separate from the channel (1). The robot picks the gasket's midpoint (2), places it into the channel's midpoint (3), and presses down to insert (4). Then an endpoint of the gasket is picked, placed, and pressed into the endpoint of the channel (5,6,7); the same is done with the second gasket endpoint (8,9,10,11). We omit the remaining pick-place-press steps at the quarter and eighth points, as well as the second press that occurs at each point for reinforcing insertion. Finally, the gripper returns to the middle of the channel (12) and slides from the middle to one end (13). It returns to the middle and slides to the other endpoint (14,15). The end result is a successfully assembled gasket (16).
• Figure 2: Channels and Gaskets in Goal Positions. Each channel has 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 ($g_0, g_1$) and channel endpoints ($c_0, c_1$) lie nearly on top of each other, but in the trapezoid case, $c_0$ and $c_1$ also lie nearly on top of each other.
• Figure 3: Methods. 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 actuatio n 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.
• Figure 4: Evaluation Metric Examples. We provide examples for all four categories of the alignment and insertion evaluation metrics discussed in Section \ref{['subsec:metric']}. 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.