Divide et Impera: Decoding Impedance Strategies for Robotic Peg-in-Hole Assembly

TL;DR

This work shows that robotic peg-in-hole assembly under impedance control does not rely on a single optimal parameter set; instead, many viable impedance configurations exist. Using the Elementary Dynamic Actions framework, the authors decompose control into submovements, oscillations, and impedance, and analyze a large impedance-parameter space with PCA and K-means to reveal task-specific and generalized assembly strategies. A neural network predictor provides practical guidance by estimating feasible impedance parameters, improving tuning efficiency and accessibility for less-experienced programmers. The study demonstrates across four peg types that impedance solutions inhabit a low-dimensional subspace with distinct strategy families, and it provides public code and CAD data to enable replication and extension in industrial robotic assembly.

Abstract

This paper investigates robotic peg-in-hole assembly using the Elementary Dynamic Actions (EDA) framework, which models contact-rich tasks through a combination of submovements, oscillations, and mechanical impedance. Rather than focusing on a single optimal parameter set, we analyze the distribution and structure of multiple successful impedance solutions, revealing patterns that guide impedance selection in contactrich robotic manipulation. Experiments with a real robot and four different peg types demonstrate the presence of task-specific and generalized assembly strategies, identified through K-means Clustering. Principal Component Analysis (PCA) is used to represent these findings, highlighting patterns in successful impedance selections. Additionally, a neural-network-based success predictor accurately estimates feasible impedance parameters, reducing the need for extensive trial-and-error tuning. By providing publicly available code, CAD files, and a trained model, this work enhances the accessibility of impedance control and offers a structured approach to programming robotic assembly tasks, particularly for less-experienced users.

Figures (13)

• Figure 1: Exemplary peg-in-hole assembly using a KUKA LBR robot. All control parameters were defined with respect to the spatially fixed coordinate frame highlighted in orange.
• Figure 2: (A) The three primitives of Elementary Dynamic Actions (EDA). Submovements and oscillations are kinematic primitives used to generate discrete and periodic movements respectively Nah_2023_EDA_kin, and mechanical impedance is a dynamic primitive used to handle physical interaction hogan2017physicalhogan2018impedancehogan2022contactlachner_energy_2021lachner_shaping_2022. (B) The three primitives are combined using a Norton equivalent network model hogan2014general.
• Figure 3: Experimental setup used during the trials. The workpiece with the hole was secured to prevent any movement throughout the four 40-hour long experimental trials.
• Figure 4: Distribution of success rates (blue bars) for each impedance parameter for the four different pegs (triangle, square, hexagon, and cylinder). For each histogram, the red solid lines represent the best linear fit, while the red dashed lines show the confidence intervals. An asterisk (*) was added for those linear fits that presented a significant slope.
• Figure 5: Left Column: The successful (green) and unsuccessful (red) trials of the cylinder, hex, square, and triangle pegs are projected into the solution space obtained using PCA. Note that often the failures fall outside the PCA-reduced solution space. 2D projections of these graphs can be found in \ref{['fig:SuccessvsFailure_combined_projections']}. Right Column: The coefficients of each of the three Principal Components are shown.