DecompGrind: A Decomposition Framework for Robotic Grinding via Cutting-Surface Planning and Contact-Force Adaptation

Shunsuke Araki; Takumi Hachimine; Yuki Saito; Kouhei Ohnishi; Jun Morimoto; Takamitsu Matsubara

cs.ROMar 24, 2026

DecompGrind: A Decomposition Framework for Robotic Grinding via Cutting-Surface Planning and Contact-Force Adaptation

Authors: Shunsuke Araki, Takumi Hachimine, Yuki Saito, Kouhei Ohnishi, Jun Morimoto, Takamitsu Matsubara

Organizations: Division of Information Science, Graduate School of Science and Technology, Nara Institute of Science and Technology (NAIST), Nara, Japan. · Haptics Research Center, Keio University, Kanagawa, Japan. · Graduate School of Informatics, Kyoto University, Kyoto, Japan. · Brain Information Communication Research Laboratory Group (BICR), Advanced Telecommunications Research Institute International (ATR), Kyoto, Japan.

Abstract

Robotic grinding is widely used for shaping workpieces in manufacturing, but it remains difficult to automate this process efficiently. In particular, efficiently grinding workpieces of different shapes and material hardness is challenging because removal resistance varies with local contact conditions. Moreover, it is difficult to achieve accurate estimation of removal resistance and analytical modeling of shape transition, and learning-based approaches often require large amounts of training data to cover diverse processing conditions. To address these challenges, we decompose robotic grinding into two components: removal-shape planning and contact-force adaptation. Based on this formulation, we propose DecompGrind, a framework that combines Global Cutting-Surface Planning (GCSP) and Local Contact-Force Adaptation (LCFA). GCSP determines removal shapes through geometric analysis of the current and target shapes without learning, while LCFA learns a contact-force adaptation policy using bilateral control-based imitation learning during the grinding of each removal shape. This decomposition restricts learning to local contact-force adaptation, allowing the policy to be learned from a small number of demonstrations, while handling global shape transition geometrically. Experiments using a robotic grinding system and 3D-printed workpieces demonstrate efficient robotic grinding of workpieces having different shapes and material hardness while maintaining safe levels of contact force.

Figures & tables

Fig. 1: Overview of robotic grinding for object shaping with DecompGrind framework. This framework grinds the workpiece into the target shape while considering removal resistance. DecompGrind consists of Global Cutting-Surface Planning (GCSP) and Local Contact-Force Adaptation (LCFA). GCSP is used to plan the next removal shape from the current shape, and LCFA is implemented for grinding each decomposed removal shape with appropriate contact force.

Fig. 2: Cutting-surface model. The current shape $\mathbf{O}_{k}$ is divided into a next shape $\mathbf{O}_{k+1}^{\mathrm{Geom}}$ and a removal shape $\mathbf{r}_{k+1}^{\mathrm{Geom}}$ using the cutting surface as a geometric boundary. Here, the superscript $\mathrm{Geom}$ indicates a geometry-based representation.

Fig. 3: Framework of Bilateral Control-based Imitation Learning (BCIL). Demonstration: A demonstrator operates the leader robot while the follower robot executes the task via bilateral control, and the robot states are recorded as training data. Training: Using the collected training data, the policy is trained to predict the next leader state from the current follower state. The predicted next leader state is denoted by $\hat{\mathbf{z}}_{t+1}^{l}$ . Execution: The learned policy substitutes for the leader in bilateral control. The follower robot performs the task accordingly. The policy reproduces the demonstrator’s contact behavior to regulate contact force.

Fig. 4: Framework of DecompGrind. A cutting surface is planned from the current and target shapes. The robot is then set up at the initial position based on the planned cutting surface. Next, the learned LCFA is then used to grind the removal shape defined by GCSP. Grinding continues until the contact surface $\mathbf{c}_{t}^{\mathrm{con}}$ reaches the cutting surface $\mathbf{c}_{k}^{*}$ . Here, $\mathbf{c}_{t}^{\mathrm{con}}$ denotes the contact surface between the belt and the workpiece, $\mathbf{c}_{k}$ denotes a candidate cutting surface, and $\mathbf{c}_{k}^{*}$ denotes the selected optimal cutting surface. When $\|\mathbf{c}_{t}^{\mathrm{con}}-\mathbf{c}_{k}^{*}\|<\epsilon$ , the current shape is observed and the next cutting surface is planned. Here, $\epsilon>0$ is a small positive constant. The index $k$ denotes the update step of GCSP and $t$ denotes the time step of LCFA. The functions $\Gamma_{\mathrm{sur}}(\cdot)$ and $\Gamma_{\mathrm{sur}}^{-1}(\cdot)$ represent the conversion between the robot state and the surface representation.

Fig. 5: Configuration of grinding environment. The experimental setup consists of a leader-follower robot system connected via bilateral control. The workpiece is attached to the follower end-effector and observed with a 3D vision sensor. During the demonstration phase, the demonstrator manipulates the leader robot under contact-force feedback from the environment, and the follower robot performs grinding through bilateral control. In the execution phase, the follower robot performs grinding based on the policy learned via BCIL, without using the leader robot.

Panel of Table 6 — TABLE I: WorkPiece for Evaluation (WP-E): Initial and target shapes

Fig. 6: Analysis of contact surface distribution with GCSP in a simulated environment

Panel of Table 8 — TABLE II: WorkPiece for LCFA Training (WP-T): Different densities, diameters, and heights. WP-T1 corresponds to blue dashed line in Fig. 6 , and WP-T2 corresponds to red dashed line.

Panel of Table 9 — TABLE III: WorkPiece with Simple Shape (WP-S): Different density, diameter, and height

Fig. 7: Comparison of shape error for evaluated methods

Fig. 8: Execution Time and shape error for each method

Panel of Table 12 — TABLE IV: Final shapes after grinding for each method

	WP-S1		WP-S2		WP-S3		WP-S4		WP-S5
Method	Execution Time [s]	In-limit force ratio [%]	Execution Time [s]	In-limit force ratio [%]	Execution Time [s]	In-limit force ratio [%]	Execution Time [s]	In-limit force ratio [%]	Execution Time [s]	In-limit force ratio [%]
Demo-Speed-1	$5.3\pm 0.1$	$100$	$-$	$0$	$-$	$0$	$-$	$0$	$-$	$0$
Demo-Speed-2	$53.2\pm 0.2$	$100$	$40.5\pm 0.3$	$100$	$40.2\pm 0.6$	$100$	$27.6\pm 0.1$	$100$	$27.5\pm 0.1$	$100$
Proposed	$5.3\pm 0.1$	$100$	$\mathbf{6.0\pm 0.2}$	$100$	$\mathbf{9.2\pm 0.2}$	$100$	$\mathbf{10.9\pm 0.3}$	$100$	$27.8\pm 2.5$	$100$

TABLE V: Grinding time and in-limit force ratio for WP-S by each method

Fig. 9: Comparison of grinding forces for WP-S between proposed method and methods without LCFA. Force data for $t\geq 3~s$ are shown. (a) Tangential force error along the $\mathrm{Z}$ -axis relative to the target tangential force of 4 N. (b) Normal force along the $\mathrm{X}$ -axis.

	WP-E1		WP-E2		WP-E3
Method	Execution Time [s]	In-limit force ratio [%]	Execution Time [s]	In-limit force ratio [%]	Execution Time [s]	In-limit force ratio [%]
Demo-Speed-1	$405.7\pm 4.3$	$20$	$569.4\pm 7.0$	$16$	$\mathbf{178.7\pm 0.9}$	$100$
Demo-Speed-2	$243.8\pm 3.3$	$\mathbf{100}$	$212.6\pm 1.7$	$\mathbf{100}$	$239.6\pm 0.6$	$100$
Proposed	$\mathbf{182.6\pm 0.7}$	$\mathbf{100}$	$\mathbf{192.3\pm 3.1}$	$\mathbf{100}$	$\mathbf{180.0\pm 1.7}$	$100$

TABLE VI: Execution time and in-limit force ratio for WP by each method

	WP-E2
Method	Execution Time [s]	In-limit force ratio [%]
Training WP-E	$259.3\pm 10.3$	$60$
Training WP-T	$\mathbf{192.3\pm 3.1}$	$\mathbf{100}$

TABLE VII: Comparison of execution time and in-limit force ratio with different demonstration workpieces

Robotic Tool Use

Robotic tool use studies how robots select, grasp, and operate tools to perform tasks beyond their native capabilities.

See all 44 papers

Acoustic Feedback for Closed-Loop Force Control in Robotic Grinding
Feb 24, 2026Zongyuan Zhang, Christopher Lehnert, Will N. Browne +1Force-Controlled Robotic Manipulation Robotic Tool Use
School of Electrical Engineering and Robotics, Queensland University of Technology, 2 George St, Brisbane, 4000, Queensland, Australia · Australian Cobotics Centre, Queensland University of Technology, 2 George St, Brisbane, 4000, Queensland, Australia · Centre for Robotics, Queensland University of Technology, 2 George St, Brisbane, 4000, Queensland, Australia
Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials
Mar 11, 2026Cenk Cetin, Shreyas Pouli, Gabriella PizzutoContact-Rich Robotic Manipulation Force-Controlled Robotic Manipulation
School of Computer Science & Informatics, University of Liverpool, UK. · Department of Chemistry, University of Liverpool, UK.
RoboFin3D: A Sim-to-Real Platform for Robotic Surface Finishing
Sep 29, 2026Haowei Wen, Shangtao Li, Vaibhav Sanjay +3Robotics Simulation Sim-to-Real Transfer
Robotics Institute, School of Computer Science, Carnegie Mellon University

Feb 24, 2026J/K move · Enter open · S save

Acoustic Feedback for Closed-Loop Force Control in Robotic Grinding

Zongyuan Zhang, Christopher Lehnert, Will N. Browne, Jonathan M. Roberts

School of Electrical Engineering and Robotics, Queensland University of Technology, 2 George St, Brisbane, 4000, Queensland, Australia · Australian Cobotics Centre, Queensland University of Technology, 2 George St, Brisbane, 4000, Queensland, Australia · Centre for Robotics, Queensland University of Technology, 2 George St, Brisbane, 4000, Queensland, Australia

Force-Controlled Robotic Manipulation Robotic Tool Use

Open paper AskPDF arXiv

Acoustic feedback is a critical indicator for assessing the contact condition between the tool and the workpiece when humans perform grinding tasks with rotary tools. In contrast, robotic grinding systems typically rely on force sensing, with acoustic information largely ignored. This reliance on force sensors is costly and difficult to adapt to different grinding tools, whereas audio sensors (microphones) are low-cost and can be mounted on any medium that conducts grinding sound. This paper introduces a low-cost Acoustic Feedback Robotic Grinding System (AFRG) that captures audio signals with a contact microphone, estimates grinding force from the audio in real time, and enables closed-loop force control of the grinding process. Compared with conventional force-sensing approaches, AFRG achieves a 4-fold improvement in consistency across different grinding disc conditions. AFRG relies solely on a low-cost microphone, which is approximately 200-fold cheaper than conventional force sensors, as the sensing modality, providing an easily deployable, cost-effective robotic grinding solution.

Abstract

Robotic Tool Use

Explore similar work

Acoustic Feedback for Closed-Loop Force Control in Robotic Grinding

Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials

RoboFin3D: A Sim-to-Real Platform for Robotic Surface Finishing