Learning from Demonstration for Robotic Deburring and Polishing: A Systematic Mapping Study

1. Introduction

Deburring and polishing represent critical, high-precision surface finishing stages in the manufacturing of components across key industries, including aerospace, automotive, medical devices, and consumer electronics. These operations are characterized by complex physical interactions, demanding the simultaneous satisfaction of competing constraints: maintaining a controlled, continuous contact force normal to the workpiece surface, executing smooth spatial trajectories over freeform geometries, and dynamically adapting to material variations, tool wear, and evolving surface topologies. In traditional industrial practice, these requirements are met through the tacit expertise of highly skilled human craftsmen. These operators possess an intuitive understanding of optimal contact angles, material-specific feed rates, and corrective force responses—aspects that are highly challenging to formalize mathematically or encode using conventional control models. However, rising labor costs and a declining availability of skilled manual labor in industrialized economies have driven a pressing demand to automate these processes by transferring expert human skills to robotic systems in a reliable and flexible manner.

The automation of surface finishing using conventional robotic programming paradigms — such as CAD/CAM-based offline trajectory generation, teach-pendant programming, and model-based force control — has achieved only limited success in high-mix, low-volume production. These approaches exhibit several fundamental limitations. First, they require high-fidelity three-dimensional CAD models of the workpiece, which are frequently unavailable, inaccurate, or geometrically altered due to manufacturing tolerances in cast, forged, or individually worn components. Second, generating force-compliant trajectories for freeform surfaces requires highly specialized robotic engineering expertise, making the programming process prohibitively expensive and time-consuming for Small and Medium-sized Enterprises (SMEs). Third, conventional setups display poor adaptability; any change in part geometry or material properties requires the robotic path and control parameters to be extensively reprogrammed, causing substantial changeover downtime. These challenges are particularly pronounced in deburring, where burr size, location, and hardness are inherently stochastic, and in polishing, where uniform surface roughness demands real-time compliance far exceeding the capabilities of open-loop position controllers.

To overcome these limitations, the Learning from Demonstration (LfD) paradigm—also known as Programming by Demonstration (PbD) or imitation learning—has emerged as a compelling methodology. LfD enables robots to acquire complex manipulation skills directly from expert demonstrations rather than through explicit mathematical programming. In an LfD framework, a human operator guides the robot kinesthetically, via haptic teleoperation interfaces, or using instrumented tools, during which the system records motion trajectories, contact forces, and multimodal sensory data. Machine learning algorithms are then deployed to extract generalizable motion primitives or control policies from the demonstration data, allowing the robot to reproduce the task autonomously under varying environmental conditions. Over the past decade, LfD research has generated a rich family of algorithmic representations for surface finishing. These range from mathematically elegant Dynamic Movement Primitives (DMPs) and probabilistic representations, such as Gaussian Mixture Models (GMM) and Probabilistic Movement Primitives (ProMPs), to recent deep generative architectures, including Diffusion Policies and Neural Ordinary Differential Equations (Neural ODEs). The maturation of collaborative robots (cobots) equipped with joint torque sensors, such as the Franka Emika Panda, has further accelerated the practical feasibility of LfD-based finishing on factory floors.

Although the broader domain of robot learning from demonstration has been surveyed in several comprehensive reviews, these works address LfD in general manipulation contexts and do not provide a focused, systematic analysis of its application to contact-rich surface finishing. The domains of robotic deburring and polishing present unique physical and mechatronic challenges that are absent from general LfD tasks: the necessity of coupling spatial trajectory imitation with contact force profile learning, the management of tool-workpiece interaction dynamics under freeform shapes, the integration of multimodal sensory feedback, and the strict constraints imposed by industrial deployability. To the best of the authors’ knowledge, no systematic mapping study has yet been conducted that exclusively addresses LfD-based methodologies in robotic deburring and polishing, synthesizes the algorithmic and sensory configurations employed, characterizes the performance evaluation metrics used across studies, and identifies the open research challenges that hinder transition from laboratory settings to industrial practice. This absence of a dedicated structured synthesis represents a significant gap in the literature, limiting the ability of both researchers and practitioners to identify the current state of the art, understand methodological trade-offs, and prioritize future research directions in this rapidly evolving field.

To address this gap, the present work conducts a systematic mapping study of the literature on LfD-based robotic deburring and polishing, covering peer-reviewed publications from 2016 to 2026. A systematic mapping study is a form of secondary research that aims to provide a broad overview of a research area through the classification and thematic aggregation of primary studies, and is particularly well-suited to emerging fields where the body of evidence is growing but remains insufficiently synthesized. The study protocol was pre-registered on the Open Science Framework (OSF) prior to the literature search to ensure methodological transparency and minimize reporting bias (https://doi.org/10.17605/OSF.IO/5Y6BR).

The principal contributions of this work are as follows:

1. Structured Taxonomy: A structured taxonomy of LfD architectures applied in robotic deburring and polishing, categorizing 24 primary studies into five methodological clusters: DMP-based methods, probabilistic and statistical models, deep learning and generative AI architectures, autonomous dynamical systems, and direct adaptive control approaches.

2. Mechatronic Analysis: A systematic analysis of sensory and control configurations, documenting how robotic platforms, force/torque sensors, vision systems, and haptic interfaces are integrated to support LfD-based surface finishing.

3. Evaluation Metrics Synthesis: A comprehensive characterization of evaluation metrics employed in the field, spanning kinematic trajectory accuracy, dynamic force tracking performance, and physical surface quality measures.

4. Industrial Challenges Roadmap: An evidence-based synthesis of industrial deployment challenges, identifying the key technical barriers that currently impede the adoption of LfD-based surface finishing in production environments.

To guide this systematic mapping study, the following four Research Questions (RQs) were defined:

• RQ1 – Methodological Landscape: Which Learning from Demonstration (LfD) architectures, trajectory learning algorithms, and mathematical representations are most commonly used in robotic deburring and polishing tasks?
• RQ2 – System and Sensory Configuration: How are robotic systems, control architectures, and sensory modalities (e.g., force sensing, vision, multimodal perception) designed and integrated to support LfD-based surface finishing operations?
• RQ3 – Evaluation and Performance Metrics: What performance evaluation criteria and validation metrics are used to assess the effectiveness of LfD-based robotic deburring and polishing methods?
• RQ4 – Industrial Deployment Challenges: What are the key technical challenges, limitations, and research gaps that hinder the transition of LfD-based robotic surface finishing methods from laboratory settings to industrial applications?

The remainder of this paper is organized as follows. Section 2 describes the methodology of the systematic mapping study, including the protocol and registration, eligibility criteria, search strategy, quality assessment framework, and screening and selection process in accordance with PRISMA 2020 [1] guidelines [1]. Section 3 presents the results of the mapping, beginning with the PRISMA flow diagram, study characteristics, bibliometric analysis, and quality assessment (Section 3.1–3.4), followed by responses to the research questions: the methodological landscape (RQ1, Section 3.5), system and sensory configuration (RQ2, Section 3.6), evaluation metrics (RQ3, Section 3.7), and industrial deployment challenges (RQ4, Section 3.8). Section 4 provides an integrated discussion of the findings. Section 5 acknowledges the limitations of this study, Section 6 outlines priority directions for future research, and Section 7 concludes the paper.

2. Methodology

2.1. Protocol and Registration

This systematic mapping study was conducted in strict accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement. To ensure methodological transparency, minimize reporting bias, and guarantee reproducibility, a detailed research protocol was drafted and registered prior to the commencement of the literature search and screening process. The protocol is publicly accessible on the Open Science Framework (OSF) registry under the following persistent identifier: https://doi.org/10.17605/OSF.IO/5Y6BR. No amendments or deviations were made to the registered protocol during the conduct of this study.

2.2. Eligibility Criteria

To identify relevant studies that address the defined research questions (RQ1–RQ4), a set of formal inclusion and exclusion criteria was established. The scope of this mapping study was restricted to primary studies focusing on the intersection of Learning from Demonstration and robotic surface finishing processes.

Studies were deemed eligible for inclusion if they met all of the following conditions:

• Thematic Relevance: The study must explicitly investigate robotic surface finishing operations, specifically deburring or polishing.
• Methodological Approach: The proposed robotic control or programming framework must incorporate an LfD-based method (e.g., kinesthetic teaching, teleoperated demonstration, or haptic imitation learning) where motion or force policies are extracted from human demonstrations.
• Technical Depth: The study must provide experimental validation or detailed theoretical formulations regarding trajectory generation, force regulation, control architectures, or sensor configurations.
• Publication Type and Language: Only peer-reviewed academic journal articles or conference proceedings published in the English language were included.

Studies were excluded if they exhibited any of the following characteristics:

• Out-of-Scope Operations: Studies addressing generic robotic manipulation tasks (e.g., pick-and-place, assembly, peg-in-hole insertion, or trajectory tracking in free space) without a specific surface finishing context.
• Non-Demonstration Control: Research utilizing traditional robotic programming (e.g., offline CAD/CAM programming, manual teach-pendant jogging) or model-based adaptive controllers that do not learn or adapt based on human demonstration data.
• Insufficient Quality or Length: Extended abstracts, editorial prefaces, technical reports, white papers, book reviews, or unpublished master’s/doctoral theses.
• Accessibility Barriers: Articles for which the full-text version was unavailable or lacked sufficient data to extract parameters necessary for answering the research questions.

A summary of the eligibility criteria is presented in Table 1.

2.3. Search Strategy and Information Sources

A comprehensive, systematic search of the literature was executed to retrieve all peer-reviewed studies published between January 1, 2016, and June 24, 2026. This decade-long timeframe was selected to capture the most recent and significant advancements in LfD architectures, collaborative robotic hardware, and deep learning algorithms.

The search was conducted across four major electronic databases:

1. IEEE Xplore: Targeted for its high concentration of robotics, automation, and control engineering publications.

2. Scopus: Utilized for broad, multidisciplinary coverage of engineering and physical sciences literature.

3. Web of Science (Core Collection): Employed to capture high-impact journals and citation indexes.

4. Google Scholar: Used as a supplementary resource to capture early-access articles, conference papers, and preprints, thereby minimizing potential publication bias.

To ensure a high recall rate, the search queries were constructed by combining keywords from two main thematic blocks using Boolean operators:

• Block A (Learning Paradigm): `"Learning from Demonstration"`, `"LfD"`, `"Imitation Learning"`, `"Programming by Demonstration"`, `"DMP"`, `"Dynamic Movement Primitives"`, `"ProMP"`, `"Probabilistic Movement Primitives"`, `"Skill learning"`, `"Task learning"`.
• Block B (Application Context): `"Deburring"`, `"Polishing"`, `"Surface finishing"`, `"Surface-finishing"`, `"Finishing"`, `"Robotic finish*"`.

The search strings were tailored to the syntax requirements of each database. For the IEEE Xplore query, the date restriction was applied manually via the search interface filters (2016–2026) because the query editor does not support date syntax directly inside the search string. The queries were executed on June 24, 2026. The specific search strings applied to each database are summarized in Table 2.

2.4. Screening and Selection Process

Upon completing the database searches, all identified records were imported into reference management software, and duplicate entries were systematically removed. The study selection process strictly followed the PRISMA 2020 [1] guidelines [1], executing three sequential screening phases:

1. Identification: The initial database searches yielded a total of 288 records (Scopus: 57; Web of Science: 43; IEEE Xplore: 29; Google Scholar: 159). Following the elimination of 65 duplicate entries, 223 unique records were compiled for screening.

2. Screening (Title and Abstract Review): A title and abstract review was conducted on the 223 unique records against the predefined inclusion and exclusion criteria. This phase resulted in the exclusion of 197 records that lacked relevance to LfD-based robotic deburring or polishing, leaving 26 records for full-text eligibility assessment.

3. Eligibility (Full-Text Review): The full texts of the 26 remaining studies were retrieved and analyzed. During this phase, 2 studies were excluded due to out-of-scope applications: one study focused exclusively on grinding-based material removal (Kim et al., 2022) and another study addressed sanding without a polishing or deburring component (Eiband et al., 2025).

Following this rigorous process, exactly 24 primary studies were deemed eligible and included in the final corpus for systematic mapping and thematic synthesis. The screening and selection process was conducted by a single primary researcher, with borderline or ambiguous cases reviewed multiple times to verify strict adherence to the thematic scope. A standardized data extraction form was utilized to ensure reliability.

2.5. Methodological Quality Assessment

To evaluate the methodological rigor of the included primary studies, a domain-specific quality assessment (QA) framework was developed. Standard clinical risk-of-bias tools are not directly applicable to experimental robotics research. Therefore, a rubric was constructed based on established methodological quality criteria for software and systems engineering studies, adapted to the specific experimental characteristics of LfD-based robotic surface finishing.

The framework comprises six key criteria (QA1–QA6), each evaluated on a three-point scale: 2 (fully met), 1 (partially met), and 0 (not met or not reported):

• QA1 — Experimental Validation: Does the study validate the proposed LfD framework on a physical robotic manipulator performing a surface finishing task?
• QA2 — Reproducibility and Statistical Reporting: Are the experiments repeated over multiple trials, and are quantitative statistical measures (e.g., mean and standard deviation) reported?
• QA3 — Baseline Comparison: Is the proposed method compared against a baseline (e.g., standard model-based control, human expert performance, or alternative LfD algorithms)?
• QA4 — Generalizability: Is the learned skill tested on varying workpiece geometries, orientations, or materials?
• QA5 — Quantitative Performance Metrics: Does the study report quantitative evaluation metrics for both kinematic trajectory accuracy and dynamic force tracking?
• QA6 — Industrial Relevance: Is the task validated using an industrial workpiece, or are real-world industrial deployment constraints (e.g., tool wear, cycle time, worker safety) discussed?

Based on the cumulative score (ranging from 0 to 12), the studies were classified into three quality tiers:

• High Quality (H): Cumulative score of 9 to 12.
• Moderate Quality (M): Cumulative score of 5 to 8.
• Low Quality (L): Cumulative score of 0 to 4.

The full results of the methodological quality assessment are reported in Table 8 (Section 3.4).

2.6. Data Extraction and Synthesis

A standardized data extraction template was designed to systematically capture the technical characteristics of each included study. The template captured data across five primary dimensions:

1. Bibliographic Metadata: Authors, publication year, publication venue, and the geographic location of the corresponding author.

2. LfD Algorithmic Architecture: Trajectory representation model, learning algorithms, and mathematical formulations (e.g., DMPs, GMM, GMR, Diffusion Policy, Neural ODEs).

3. Sensory Modalities: Sensors integrated during the human demonstration phase and the robot execution phase (e.g., 6-DOF force/torque sensors, joint torque sensors, RGB/RGB-D cameras, haptic interfaces).

4. Hardware and Control Platform: The specific robotic manipulator model, degrees of freedom (DOF), and low-level force/position control strategies (e.g., impedance control, admittance control, variable stiffness control).

5. Application Environment: The specific surface finishing task (deburring, polishing, cleaning, rust removal) and the shape/material of the test workpiece.

Data extraction was completed by the primary researcher in a continuous session to ensure coding consistency. For studies where specific hardware parameters were not reported or were irrelevant (e.g., studies validating skills using custom instrumented tools unattached to a robotic arm), the corresponding fields were coded with a `-` symbol to maintain data integrity. The extracted data were then synthesized qualitatively to answer the research questions, and quantitative summaries (frequencies and percentages) were calculated for the algorithmic and sensory distributions. For the purpose of this systematic mapping study, the primary outcomes of interest are defined as the qualitative LfD algorithmic architectures, mechatronic sensory configurations, low-level control systems, and the corresponding quantitative evaluation metrics (e.g., force and trajectory RMSE, surface roughness $Ra$) reported across the corpus.

2.7. Effect Measures and Synthesis Rationale

Due to the qualitative and taxonomic nature of this systematic mapping study, standard statistical effect measures (such as risk ratios, relative risks, or mean differences) commonly employed in clinical systematic reviews are not applicable. Instead, the synthesis is structured around descriptive and frequency-based thematic aggregation of algorithmic, mechatronic, and control features. Quantitative baseline performance metrics (e.g., trajectory tracking root-mean-square errors and surface roughness measures) are summarized qualitatively to establish baseline benchmarks without performing statistical meta-analyses.

3. Results

3.1. Study Selection and PRISMA Flow Diagram

The initial systematic search across the four electronic databases yielded 288 records. After importing these records into reference management software, 65 duplicate entries were identified and removed, leaving 223 unique records for the screening phase.

During the title and abstract screening, 197 records were excluded as they did not meet the predefined eligibility criteria (primarily due to lack of LfD methodology or relevance to robotic surface finishing). The remaining 26 studies were retrieved for full-text eligibility assessment.

Following a detailed full-text review, 2 studies were excluded: one focused exclusively on grinding-based material removal (Kim et al., 2022) and the other addressed sanding without a polishing or deburring component (Eiband et al., 2025). This process resulted in a final corpus of 24 primary studies included in the systematic mapping and thematic synthesis. The study selection process is illustrated in Figure 1.

3.2. Quantitative Study Characteristics

A summary of the extracted data from the 24 included primary studies is presented in Table 3. The corpus represents a decade of research (2016–2026) and documents a diverse range of LfD algorithms, sensory modalities, robotic systems, and surface finishing applications.

3.3. Bibliometric Analysis

3.3.1. Temporal Distribution

The temporal distribution of the 24 selected studies between 2016 and 2026 reveals a significant growth in research activity. As summarized in Table 4, the annual publication volume remained low but stable from 2017 to 2022, with a sharp increase beginning in 2023. The years 2024 and 2025 represent the peak period of academic output, accounting for 58.3% of the entire corpus. This trend reflects the growing interest in utilizing learning-based paradigms to automate contact-rich manufacturing tasks.

3.3.2. Geographical Distribution

The geographical analysis, determined by the affiliation of the corresponding author, indicates a highly centralized research landscape. As shown in Table 5, East Asian institutions—specifically from China—account for the vast majority of the research, contributing 15 papers (62.5%). Europe forms the next largest research cluster, led by Austria (8.3%) and supported by Germany, Slovenia, and Portugal. Turkey represents a significant individual contributor with 12.5% of the publications.

3.3.3. Quantitative Methodological and Sensory Distributions

To provide a quantitative overview of the technological configurations, the 24 included studies were classified based on their primary LfD algorithmic approach (RQ1) and sensory modalities (RQ2).

Dynamic Movement Primitives (DMPs) and their force-coupled extensions represent the most prevalent mathematical framework, utilized in 9 studies (37.5%). Probabilistic models (GMM-GMR, ProMPs, FMPs, GPs) follow closely with 8 studies (33.3%). Modern deep learning and generative AI architectures represent an emerging cluster with 4 studies (16.7%), as summarized in Table 6.

Regarding sensory configurations, force and torque sensing remains the dominant modality. As detailed in Table 7, 17 studies (70.8%) rely exclusively on force/torque feedback (either via external F/T sensors or joint torque measurements) to regulate tool-workpiece contact. Vision-only systems account for 3 studies (12.5%), while multimodal configurations combining force, vision, and haptic feedback represent 4 studies (16.7%).

3.4. Methodological Quality Assessment Results

The methodological quality of the 24 included studies was evaluated using the six-criterion rubric (QA1–QA6) defined in Section 2.5. The results are presented in Table 7.

Overall, the corpus demonstrates a high-to-moderate level of methodological rigor: 10 studies (41.7%) were rated High quality (score 9–12) and 14 studies (58.3%) were rated Moderate quality (score 5–8). No study was rated Low quality, reflecting the peer-reviewed selection process.

The most consistently satisfied criterion was QA1 (Experimental Validation), which was met by 100% of the studies. The most common limitation was QA2 (Reproducibility and Statistical Reporting), reflecting a general lack of statistical replication (e.g., reporting mean ± SD over multiple trials) across the experimental setups, with only 2 studies (8.3%) fully meeting this criterion.

Table 8. Methodological Quality Assessment of Included Studies (n = 24).

Source	QA1	QA2	QA3	QA4	QA5	QA6	Total	Rating
Min et al. [2]	2	1	1	1	2	2	9	High
Li et al. [3]	2	2	2	1	2	1	10	High
Si et al. [4]	2	1	1	1	1	1	7	Moderate
Wang et al. [5]	2	1	1	1	2	1	8	Moderate
Wang et al. [6]	2	1	1	1	2	1	8	Moderate
Acikgoz et al. [7]	2	1	1	1	1	2	8	Moderate
Zhai et al. [8]	2	1	1	1	2	1	8	Moderate
Fischer et al. [9]	2	2	2	2	1	1	10	High
Kulak et al. [10]	2	1	1	1	1	1	7	Moderate
Zhang et al. [11]	2	1	1	1	2	1	8	Moderate
Wu et al. [12]	2	1	1	1	2	2	9	High
Wu et al. [13]	2	1	1	2	2	1	9	High
Haninger et al. [14]	2	1	1	1	2	1	8	Moderate
Möhl et al. [15]	2	1	2	2	1	2	10	High
Parvizi et al. [16]	2	1	1	1	1	2	8	Moderate
Wang et al. [17]	2	1	2	2	2	1	10	High
Xu et al. [18]	2	2	2	1	2	1	10	High
Shen et al. [19]	2	1	1	2	2	2	10	High
Hamdan et al. [20]	2	1	1	1	2	1	8	Moderate
Liao et al. [21]	2	1	1	1	2	1	8	Moderate
Li et al. [22]	2	1	1	2	1	1	8	Moderate
Ke et al. [23]	2	1	2	2	1	2	10	High
Nemec et al. [24]	2	1	1	1	2	1	8	Moderate
Duarte et al. [25]	2	1	1	1	1	1	7	Moderate
Fully Met (score=2)	24	2	6	6	14	6	-	-
% Fully Met	100%	8.3%	25.0%	25.0%	58.3%	25.0%	-	-

Table 8. Methodological Quality Assessment of Included Studies (n = 24).

Source	QA1	QA2	QA3	QA4	QA5	QA6	Total	Rating
Min et al. [2]	2	1	1	1	2	2	9	High
Li et al. [3]	2	2	2	1	2	1	10	High
Si et al. [4]	2	1	1	1	1	1	7	Moderate
Wang et al. [5]	2	1	1	1	2	1	8	Moderate
Wang et al. [6]	2	1	1	1	2	1	8	Moderate
Acikgoz et al. [7]	2	1	1	1	1	2	8	Moderate
Zhai et al. [8]	2	1	1	1	2	1	8	Moderate
Fischer et al. [9]	2	2	2	2	1	1	10	High
Kulak et al. [10]	2	1	1	1	1	1	7	Moderate
Zhang et al. [11]	2	1	1	1	2	1	8	Moderate
Wu et al. [12]	2	1	1	1	2	2	9	High
Wu et al. [13]	2	1	1	2	2	1	9	High
Haninger et al. [14]	2	1	1	1	2	1	8	Moderate
Möhl et al. [15]	2	1	2	2	1	2	10	High
Parvizi et al. [16]	2	1	1	1	1	2	8	Moderate
Wang et al. [17]	2	1	2	2	2	1	10	High
Xu et al. [18]	2	2	2	1	2	1	10	High
Shen et al. [19]	2	1	1	2	2	2	10	High
Hamdan et al. [20]	2	1	1	1	2	1	8	Moderate
Liao et al. [21]	2	1	1	1	2	1	8	Moderate
Li et al. [22]	2	1	1	2	1	1	8	Moderate
Ke et al. [23]	2	1	2	2	1	2	10	High
Nemec et al. [24]	2	1	1	1	2	1	8	Moderate
Duarte et al. [25]	2	1	1	1	1	1	7	Moderate
Fully Met (score=2)	24	2	6	6	14	6	-	-
% Fully Met	100%	8.3%	25.0%	25.0%	58.3%	25.0%	-	-

3.5. RQ1: Methodological Landscape

Based on the systematic mapping of the literature, LfD architectures in robotic deburring and polishing are categorized into five primary methodological clusters. These approaches range from mathematical trajectory parameterization to advanced deep generative artificial intelligence models:

3.5.1. Dynamic Movement Primitives (DMP) and Variants

DMPs represent the most prevalent framework (37.5% of studies), utilizing a system of second-order differential equations (spring-damper systems) augmented with a non-linear forcing term to encode and reproduce demonstrated trajectories. The transformation system is mathematically formulated as:

$$\tau \dot{v}=K_p(g-x)-D_{p}v+(g-x_0)f\left(s\right)$$

(1)

$$\tau \dot{x}=v$$

(2)

where x is the position, v is the velocity, g is the goal position, $x_0$ is the starting position, $K_p$ and $D_p$ are the stiffness and damping gains, and $\tau$ is a time scaling parameter. The forcing term $f\left(s\right)$ is parameterized using Gaussian basis functions ${\psi }_i\left(s\right)$ as:

$$f\left(s\right)={\displaystyle \frac{{\sum }_i{w}_i{\psi }_i\left(s\right)}{{\sum }_i{\psi }_i\left(s\right)}}s$$

(3)

where $w_i$ are the weights learned from demonstration, and s is the phase variable governed by the canonical system $\tau \dot{s}=-{\alpha }_{s}s$ (with decay rate ${\alpha }_s$). While standard DMPs guarantee spatial and temporal scaling, they are kinematically restricted and cannot adapt to contact force variations. To address this, studies have introduced several domain-specific modifications:

• Force-Controlled Dynamic Coupling DMPs (FDC-DMP): Introduced by Shen et al. [19], this framework adds a coupling term driven by interaction forces directly into the DMP acceleration equation, enabling the robot to dynamically modify its path (e.g., avoiding obstacles during bus body polishing) without altering the global target.
• B-Spline DMPs (BDMPs): Wang et al. [17] proposed replacing the standard Gaussian basis functions in DMPs with B-splines. This modification significantly improves trajectory modeling accuracy with a smaller number of basis functions. The forcing term is parameterized as $f\left(s\right)={\sum }_j{B}_j\left(s\right)w_j$, where $B_j\left(s\right)$ represents the B-spline basis functions. When optimized using Policy Improvement with Path Integrals (${\mathrm{PI}}^2$) reinforcement learning, BDMPs demonstrate high generalization capability for polishing trajectories and force profiles under unseen workpiece positions.
• Riemannian DMPs: Liao et al. [21] extended DMPs to Riemannian manifolds (e.g., Cartesian space and 2-D sphere manifolds) to simultaneously model human motion, 3-D endpoint stiffness, and contact forces from a one-shot demonstration, solved via Quadratic Programming (QP). Trajectories on the manifold $\mathcal{M}$ are generated by mapping the states to the tangent space $T_x\mathcal{M}$, maintaining geometrical properties of robot orientations.
• Neural Network-Augmented DMPs: Wang et al. [5] integrated a Phase-Modulated Diagonal Recurrent Neural Network (PMDRNN) with DMPs to adaptively predict trajectory offsets based on real-time force tracking deviations, mitigating environmental uncertainties.

3.5.2. Probabilistic and Statistical Models

Probabilistic models (33.3% of studies) represent demonstrations as joint probability distributions, capturing task variations, correlations, and rhythmic patterns across multiple demonstrations:

• Gaussian Mixture Models and Regression (GMM-GMR): Used to model the joint distribution of time, space, and force parameters. Wu et al. [13] utilized GMMs to encode human polishing dynamics, combining GMR with a variable impedance controller to regulate contact compliance. Zhai et al. [8] integrated GMM-GMR with a vector-valued Gaussian Process (GP) to online-modulate robotic trajectories when subjected to human external forces.
• Probabilistic Movement Primitives (ProMPs): Unlike DMPs, ProMPs capture the statistical variance of demonstrations. ProMPs represent a trajectory as a linear combination of basis functions: $y\left(t\right)=\mathsf{\Phi }\left(t\right)w+ϵ$, where $w\sim \mathcal{N}({\mu }_w,{\Sigma }_w)$ captures the statistical variance across multiple demonstrations. Wang et al. [6] developed Arc-Length ProMPs (AL-ProMP), mapping the probability distribution of contact forces to spatial coordinates (arc-length $s_l$) rather than time t, formulating the trajectory as $y\left(s_l\right)=\mathsf{\Phi }\left(s_l\right)w+ϵ$. This formulation prevents trajectory distortions during non-linear speed scaling.
• Fourier Movement Primitives (FMP): Grounded in signal processing, Kulak et al. [10] proposed FMPs using Fourier series as basis functions: $y\left(t\right)=a_0+{\sum }_{k=1}^K(a_{k}cos\left(k\omega t\right)+b_{k}sin\left(k\omega t\right))$. FMPs approximate periodic, multi-frequency signals (e.g., circular polishing patterns) from unaligned demonstrations without requiring phase alignment or frequency extraction.

3.5.3. Deep Learning and Generative AI Architectures

Representing 16.7% of the studies, these approaches leverage deep neural networks to directly map high-dimensional visual or proprioceptive observations to continuous control actions:

• Diffusion Policies: Ke et al. [23] and Li et al. [3] utilized diffusion models to generate continuous, expert-like motion-force trajectories. In the DP-RRL framework [3], the Diffusion Policy generates a trajectory distribution by iteratively denoising a random sequence $x^K,x^{K-1},\dots ,x^0$ using a noise predictor ${ϵ}_{\theta }(x^k,k,O)$ conditioned on observation O. The residual RL agent then predicts a displacement $\Delta F_t$ to correct the reference force based on contact dynamics.
• Neural Ordinary Differential Equations (Hyper-NODEs): Xu et al. [18] developed a Hyper-NODE architecture to generate continuous position and orientation (quaternion) trajectories. The system dynamics are modeled as:

  $${\displaystyle \frac{dh\left(t\right)}{dt}}=f(h\left(t\right),t;\theta )$$

  (4)

  where $h\left(t\right)$ represents the continuous hidden state, and $\theta$ are the weights generated by a hypernetwork. Paired with Control Barrier Functions (CBF), the system guarantees obstacle avoidance in local non-polishing areas (LNP-areas) while estimating admittance control parameters.
• Neural Network Morphing: Möhl et al. [15] designed a keypoint-driven non-linear morphing network to transfer demonstrated trajectories between 3D point clouds of geometrically similar objects without CAD models.

3.5.4. Autonomous Dynamical Systems (DS)

Grounded in control theory, DS-based approaches (8.3% of studies) model robot motion as time-invariant differential equations to ensure global asymptotic stability and immediate reactivity to physical perturbations:

• Stable Limit Cycles: Duarte et al. [25] represented periodic human polishing movements (e.g., circular or elliptical motions) using a time-invariant DS with a stable limit cycle attractor. The system is formulated as a second-order nonlinear dynamical system of the form $\dot{x}=f\left(x\right)$, where the trajectories are forced to converge asymptotically to a closed orbit $\mathcal{C}$. This formulation guarantees that the robot converges back to the demonstrated polishing pattern even after being physically displaced.
• DS-based Imitation Learning: Si et al. [4] proposed a dynamically stable energy field framework to provide virtual haptic guidance during teleoperated human demonstrations, reducing operator physical workload.

3.5.5. Direct Adaptive Control and Parameter Estimation

Representing 4.2% of the studies [2], this approach focuses on decoupling demonstrations into pure "motion skills" (discrete pose sequences) and "force skills" (desired normal forces) to achieve accurate skill transfer through a computed-torque impedance control law on complex surfaces (e.g., violin bodies).

3.6. RQ2: System and Sensory Configuration

To support LfD-based robotic surface finishing, mechatronic architectures must be carefully designed to capture high-fidelity human demonstrations, perceive unstructured environments, and execute contact-rich tasks safely:

3.6.1. Sensory Modalities and Multimodal Perception

• Force and Torque Sensing: As the primary driver of closed-loop execution, force feedback is integrated in 87.5% of the studies (either as the sole sensor or in multimodal setups). While end-effector 6-DOF F/T sensors are standard, Hamdan et al. [20] proposed a dual-force sensor configuration. One sensor measures the human operator’s guiding force ($F_h$), while the second measures the tool-workpiece interaction force ($F_{int}$). By calculating the environmental reaction force:

  $$F_{env}=F_h-F_{int}$$

  (5)

  the system isolates the environmental dynamics (stiffness, friction) from human guidance inputs.
• Vision and Spatial Perception: To handle geometrically complex surfaces, depth sensors (e.g., overhead or wrist-mounted RGB-D cameras) are integrated. These sensors capture raw 3D point clouds, which are processed via PointNet++ or keypoint-based neural networks to reconstruct surface meshes [23] or guide trajectory morphing [15].
• Kinematic and Biometric Tracking: Teleoperation and kinesthetic demonstration interfaces utilize haptic devices (e.g., Geomagic Touch) or wearable inertial measurement units (IMUs). More advanced setups incorporate surface electromyography (sEMG) sensors on the human arm to capture synergistic muscle activity, translating muscle co-contraction directly into robot joint stiffness parameters.
• Instrumented Tools: To facilitate platform-independent demonstrations, Fischer et al. [9] designed custom instrumented tools and mechanical alignment plates to record high-quality contact forces and orientations directly on the workpiece.

3.6.2. Control Architectures

Due to the rigid nature of physical contact during surface finishing, pure position control is unsafe. Consequently, studies rely on indirect force control strategies:

• Variable Impedance and Admittance Control: These strategies model the robot-workpiece interface as a mass-spring-damper system. The low-level dynamic behavior is governed by the admittance control law:

  $$M_d(\ddot{x}-{\ddot{x}}_d)+D_d(\dot{x}-{\dot{x}}_d)+K_d(x-x_d)=F_{ext}-F_{ref}$$

  (6)

  where $M_d$, $D_d$, and $K_d$ are the desired mass, damping, and stiffness matrices, $x_d$ is the reference trajectory, x is the actual position, $F_{ext}$ is the external interaction force, and $F_{ref}$ is the target reference force. Variable impedance controllers (Wu et al., 2023, 2025) modulate stiffness $K_d$ and damping $D_d$ online. For instance, stiffness is lowered when transitioning onto hard, brittle materials (e.g., iron) to prevent impact chatter, and increased on soft materials (e.g., wood) to ensure uniform material removal.
• Iterative Learning Control (ILC): To compensate for repetitive tracking errors, ILC is integrated with impedance control. Zhang et al. [11] combined GMM trajectory models with Dynamic Time Warping ILC (DTW-ILC), iteratively updating the robot’s reference path based on stiffness estimation to achieve fast force convergence over multiple polishing passes.

3.6.3. Robotic Systems and Hardware Integration

The Franka Emika Panda (7-DOF) and Universal Robots (UR5, UR10e) cobots are highly preferred due to their backdrivability, joint torque sensing, and active gravity compensation, which are crucial for kinesthetic teaching. For high-precision micro-machining (e.g., deburring), where cobots lack sufficient stiffness, specialized setups are used, such as a 6-DOF hexapod paired with high-speed piezoelectric actuators controlled via teleoperated haptic interfaces [7]. Control loops typically separate high-frequency force control (running at $1000\mathrm{Hz}$) from low-frequency visual perception and policy planning (running at $10-30\mathrm{Hz}$) to maintain real-time stability.

3.7. RQ3: Evaluation and Performance Metrics

The effectiveness of LfD-based robotic finishing is evaluated using a multi-tiered framework, spanning kinematic accuracy, dynamic force tracking, and physical surface quality:

3.7.1. Kinematic and Trajectory Accuracy Metrics

• Dynamic Time Warping (DTW) Distance: Quantifies the spatiotemporal similarity between the demonstrated human trajectory and the robot’s executed path, especially when feed rates vary.
• Root Mean Square Error (RMSE): Calculates the spatial deviation (in millimeters) between the executed end-effector path $x_i$ and the demonstrated trajectory ${\widehat{x}}_i$ over N samples:

  $${\mathrm{RMSE}}_p=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{\parallel x_i-{\widehat{x}}_i\parallel }^2}$$

  (7)
• Pearson Correlation Coefficient (r): Measures the shape similarity of the trajectories:

  $$r={\displaystyle \frac{{\sum }_{i=1}^N(x_i-\overline{x})({\widehat{x}}_i-\overline{\widehat{x}})}{\sqrt{{\sum }_{i=1}^N{(x_i-\overline{x})}^2}\sqrt{{\sum }_{i=1}^N{({\widehat{x}}_i-\overline{\widehat{x}})}^2}}}$$

  (8)

  with values closer to 1.0 indicating high imitation fidelity.
• Relative Smoothness ($r_{c_p}$ for position, $r_{c_q}$ for orientation): Evaluates the jerk of the generated trajectory to ensure smooth robotic motion.

3.7.2. Force Tracking and Dynamic Interaction Metrics

• Force Root Mean Square Error (Force RMSE): The primary metric to evaluate force tracking performance, measuring the deviation between the executed contact force $F_i$ and the demonstrated reference force profile ${\widehat{F}}_i$:

  $${\mathrm{RMSE}}_F=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{(F_i-{\widehat{F}}_i)}^2}$$

  (9)
• Maximum Impact Force ($F_{max}$): Evaluates system compliance and safety during initial tool contact or material transitions.
• Mean Force Deviation ($\Delta F$): Quantifies the stability of the normal force during continuous polishing.

3.7.3. Surface Quality and Process-Specific Metrics

• Surface Roughness (Ra, Rq, Rz): Measured using contact profilometers or white-light interferometers. A reduction in average roughness ($Ra$) verifies successful surface smoothing.
• Material Removal Rate (MRR): Replicating the expert’s material removal strategy is evaluated based on Preston’s equation:

  $$\Delta z=k_{p}Pv\Delta t$$

  (10)

  where $\Delta z$ is the thickness of the removed material, $k_p$ is Preston’s coefficient (depending on tool and material properties), P is the contact pressure (directly proportional to normal force $F_n$), and v is the relative tool speed. Studies like Kim et al. (2023) and Min et al. [2] estimate these parameters online to adapt forces dynamically on curved surfaces.
• Remaining Stain Ratio (RSR): In cleaning applications, image segmentation is used to calculate the percentage of stains remaining on the surface post-execution.

3.8. RQ4: Industrial Deployment Challenges

Despite laboratory advancements, several critical technical barriers hinder the transition of LfD-based finishing to industrial assembly lines:

3.8.1. The Sim-to-Real Gap and Complex Contact Dynamics

Deep Reinforcement Learning (DRL) and generative models (e.g., Diffusion Policies) require thousands of training episodes. However, physics engines struggle to simulate contact dynamics (friction hysteresis, tool deformation, workpiece stiffness, and material removal profiles) with high fidelity. Consequently, policies trained in simulation exhibit degraded performance when deployed on physical hardware.

3.8.2. Demonstration Quality and Hardware Constraints

Kinesthetic teaching is prone to human error and kinematic limitations:

• Kinematic Interference: The physical weight and joint limits of the robot arm restrict the operator’s natural movement, leading to distorted demonstrations.
• Sensor Noise: High-speed spindle rotation and pneumatic tool vibrations generate significant mechanical noise, degrading force/torque sensor readings during the teaching phase.
• Cognitive Overload: Controlling the robot’s spatial path, tool orientation, and contact force simultaneously in real time places a high cognitive demand on the human expert.

3.8.3. Generalization to Complex Geometries and LNP-Areas

• Local Non-Polishing (LNP) Areas: Industrial components often feature functional geometry (e.g., threaded holes, slots, and ribs) that must remain untouched. Autonomously detecting and avoiding these LNP-areas while maintaining a constant normal force on the surrounding freeform surface remains an open control problem.
• Geometric Generalization: Trajectory generalization models (e.g., DMPs) often distort orientations when scaling trajectories to highly curved, non-planar workpieces, risking collision or uneven polishing.

3.8.4. Multimodal Perception and Computational Complexity

Fusing high-dimensional point clouds, haptic data, and force feedback in real time requires substantial computational resources. Traditional LfD models like Gaussian Processes or GMMs suffer from poor scalability, making high-frequency closed-loop control ($>500\mathrm{Hz}$) difficult to achieve on standard industrial controllers.

3.8.5. Need for Robust Human-in-the-Loop (HITL) Systems

One-shot offline learning is vulnerable to environmental changes. Industrial deployment requires active online correction mechanisms (HITL), allowing human operators to intuitively intervene, adjust control parameters (e.g., stiffness or feed rate), and correct localized trajectory segments in real time without restarting the programming process.

4. Discussion

4.1. Methodological and Academic Perspectives: Temporal Evolution (2016–2026)

The systematic mapping of the literature reveals a significant temporal evolution in the methodological landscape of robot learning from demonstration for surface finishing tasks. In the early phase of the reviewed period (2016–2020), research was heavily dominated by Dynamic Movement Primitives (DMPs) and classical statistical representations such as Gaussian Mixture Models with Gaussian Mixture Regression (GMM-GMR). These approaches focused primarily on the mathematical parameterization of human hand trajectories, relying on second-order spring-damper equations to guarantee spatial scaling and temporal robustness. However, these early frameworks faced critical limitations in contact-rich tasks due to their inability to dynamically adapt contact force profiles under unmodeled surface variations.

From 2021 onwards, a distinct paradigm shift is observed toward adaptive mechatronic integration and hybrid control. Researchers increasingly coupled movement primitives with variable impedance and admittance control, transitioning LfD from a purely kinematic imitation tool into a dynamic, multimodal learning framework. The latest period (2024–2026) is characterized by the emergence of deep generative AI models, such as Diffusion Policies and Neural Ordinary Differential Equations (Neural ODEs). These frameworks process high-dimensional spatial point clouds directly to generate obstacle-aware continuous trajectories in real time, eliminating the need for rigid pre-computed workpiece geometry models. This progression highlights a transition from simple trajectory replication to complex, sensory-driven reactive behaviors capable of generalizing to entirely new surface topologies.

Beyond the algorithmic representations, the systematic mapping reveals notable quantitative benchmark trends across the primary studies. In terms of force tracking accuracy, Dynamic Movement Primitives (DMPs) coupled with adaptive variable impedance or admittance controllers consistently report force tracking root-mean-square errors (RMSE) between $0.5\mathrm{N}$ and $1.5\mathrm{N}$ under laboratory conditions (e.g., Wang et al., 2023a; Liao et al., 2024). Regarding surface quality, polishing frameworks report substantial improvements, with average surface roughness ($Ra$) typically reduced from initial post-machining values of $1.5–3.0\mu \mathrm{m}$ down to finished tolerances of $0.1–0.3\mu \mathrm{m}$ (e.g., Min et al., 2024; Zhang et al., 2024). For spatial trajectory reproduction, deep learning and generative AI models (such as Diffusion Policies and Hyper-NODEs) demonstrate high spatial fidelity, achieving Pearson correlation coefficients (r) greater than 0.95 and endpoint trajectory RMSE values under $2.0\mathrm{mm}$ relative to the expert human demonstrations (e.g., Li et al., 2025; Ke et al., 2025; Xu et al., 2025). These values provide a quantitative baseline for evaluating future LfD-based robotic finishing frameworks.

Table 9. Quantitative Performance Baseline Benchmarks in LfD-Based Robotic Surface Finishing.

Metric Category	Metric	Performance Baseline Range	Primary Algorithmic Family	Representative Studies
Force Tracking Accuracy	Force RMSE (Root Mean Square Error)	$0.5\mathrm{N}-1.5\mathrm{N}$	DMP + Adaptive Variable Impedance Control	Wang et al. [5], Liao et al. [21]
Surface Quality	Average Surface Roughness ($Ra$)	$0.1–0.3\mu \mathrm{m}$ (Initial: $1.5–3.0\mu \mathrm{m}$)	GMM-GMR & ILC-Based Polishing	Min et al. [2], Zhang et al. [11]
Spatial Trajectory Accuracy	Trajectory RMSE & Pearson Correlation (r)	RMSE < $2.0\mathrm{mm}$, $r>0.95$	Deep Generative AI (Diffusion Policy / Hyper-NODE)	Li et al. [3], Ke et al. [23], Xu et al. [18]

Table 9. Quantitative Performance Baseline Benchmarks in LfD-Based Robotic Surface Finishing.

Metric Category	Metric	Performance Baseline Range	Primary Algorithmic Family	Representative Studies
Force Tracking Accuracy	Force RMSE (Root Mean Square Error)	$0.5\mathrm{N}-1.5\mathrm{N}$	DMP + Adaptive Variable Impedance Control	Wang et al. [5], Liao et al. [21]
Surface Quality	Average Surface Roughness ($Ra$)	$0.1–0.3\mu \mathrm{m}$ (Initial: $1.5–3.0\mu \mathrm{m}$)	GMM-GMR & ILC-Based Polishing	Min et al. [2], Zhang et al. [11]
Spatial Trajectory Accuracy	Trajectory RMSE & Pearson Correlation (r)	RMSE < $2.0\mathrm{mm}$, $r>0.95$	Deep Generative AI (Diffusion Policy / Hyper-NODE)	Li et al. [3], Ke et al. [23], Xu et al. [18]

4.2. Task-Specific Synthesis: Polishing vs. Deburring

A key finding of this systematic mapping is the severe methodological imbalance between polishing and deburring applications. Out of the 24 included primary studies, polishing and cleaning operations represent the vast majority, while only 3 studies (Acikgoz et al., 2017; Parvizi et al., 2017; Li et al., 2020) focus primarily on robotic deburring. This asymmetry is driven by the fundamentally different physical dynamics of the two processes. Polishing is characterized by continuous, low-frequency contact forces over smooth surfaces, which align well with the spatial smoothness assumptions of primitives like DMPs or Gaussian Processes. In contrast, deburring involves highly non-linear, high-frequency impact forces encountered when the tool contacts rigid, variable-sized burrs. These transient contact dynamics introduce severe chattering, tool-wear uncertainties, and risk of mechanical failure, which are exceptionally difficult to capture via standard imitation learning.

Consequently, the 3 deburring studies in the corpus rely on specialized teleoperated haptic interfaces or specified DMPs (sDMPs) optimized via heuristic algorithms (e.g., Particle Swarm Optimization) to learn safety-critical force reflexes. The underrepresentation of deburring represents a critical research gap: while robotic polishing is approaching industrial readiness, robust LfD-based deburring remains largely restricted to simplified laboratory settings. Future research must address this gap by developing high-frequency adaptive force policies capable of stabilizing rigid tool-workpiece interactions during burr contact.

4.3. Economic and Operational Implications for SMEs

For Small and Medium-sized Enterprises (SMEs) operating in high-mix, low-volume (small batch) production environments, traditional robotic automation is often economically infeasible due to the prohibitive setup and programming costs of CAD/CAM systems. LfD transforms this financial landscape by providing three distinct cost-saving and operational advantages. First, LfD eliminates the need for highly paid, specialized robotic programming engineers. By using intuitive kinesthetic teaching or teleoperation, existing skilled shop-floor operators can ’teach’ the robot optimal finishing strategies through direct physical demonstration, utilizing their tacit craftsmanship without writing code. Second, LfD drastically reduces setup and changeover times. Traditional CAD/CAM programming requires watertight 3D models and offline path planning, a process that can take hours or days for complex parts. An LfD system can adapt to a new part geometry in minutes through a single expert demonstration or keypoint-driven neural morphing, minimizing changeover downtime. Third, LfD lowers hardware barriers by allowing the use of cost-effective sensors (e.g., consumer-grade RGB-D cameras) and avoiding the need for expensive, force-accurate simulators. By consolidating these mechatronic, personnel, and temporal savings, LfD represents a financially viable pathway for SMEs to automate complex finishing operations under highly variable production requirements.

4.4. Comparative Synthesis of the Included Studies

To systematically map the LfD literature in robotic surface finishing, Table 10 synthesizes the core commonalities (similarities) and key distinctions (differences and unique contributions) of all 24 primary studies included in the corpus.

4.5. Identification of Research Gaps in the LfD Literature

The synthesis of the 24 primary studies reveals four critical research gaps that currently limit the adoption of LfD-based robotic finishing in industrial environments:

1. The Tactile Complexity and Mechanical Hazards of Deburring:

There is a severe imbalance in task representation, with 21 studies focusing on polishing/cleaning and only 3 addressing deburring. Polishing involves low-frequency contact normal to smooth surfaces, which aligns well with standard spatial primitives. In contrast, deburring involves high-frequency, non-linear impact forces when contacting rigid, stochastic burrs. These dynamics introduce chattering, high tool wear, and risk of mechanical failure, which standard LfD frameworks fail to model or stabilize.

2. Generalization to Freeform Geometries without CAD Templates:

Most current LfD frameworks assume flat or simple curved workspaces. Generalizing a demonstrated trajectory to highly curved, freeform 3D objects frequently leads to spatial trajectory distortion and orientation misalignment (quaternion errors). While keypoint morphing [15] and Mesh-DMPs [23] have emerged to address this, they require high computational power and struggle to maintain constant contact forces on highly non-planar surfaces.

3. The Sim-to-Real Gap in Physical Interaction Physics:

Modern deep learning models (e.g., Diffusion Policies, Residual RL) offer excellent trajectory generation capabilities but require thousands of training episodes. Running these on physical hardware is impractical due to safety risks and tool wear, necessitating training in simulation. However, physics engines cannot model contact physics (friction hysteresis, tool compliance, material removal profiles) with high fidelity. Consequently, policies trained in simulation degrade when deployed on physical factory floors.

4. Lack of Intuitive online Human-in-the-Loop (HITL) Correction:

The majority of LfD frameworks rely on offline, one-shot demonstrations. If the demonstration is noisy or if the workshop environment drifts (e.g., due to tool wear), the robot cannot adapt. There is a lack of real-time, online interactive correction systems that allow human operators to physically intervene, correct localized trajectory segments, and adapt impedance parameters on the fly.

4.6. How the Current Study Addresses These Research Gaps

The present systematic mapping study contributes directly to addressing these literature gaps by establishing a structured, evidence-based foundation for future LfD research:

• Addressing the Task Imbalance: By exposing the severe deficit in deburring research (only 12.5% of the corpus), this study provides a clear technical analysis of why deburring is exceptionally challenging (high-frequency transient contact, mechanical chattering) and catalogued the haptic and PSO-based control strategies used by the few successful deburring studies. This guides future researchers toward the control and haptic configurations necessary to tackle deburring.
• Providing a Mechatronic and Sensory Roadmap: This study maps the exact mechatronic configurations required to support LfD (such as the dual-force sensor configuration of Hamdan et al. to isolate human forces from environmental reaction forces). By documenting the frequency of force-only (70.8%) vs. multimodal (16.7%) setups, it provides a design guideline for building hardware platforms capable of perceiving both geometric and tactile environments.
• Synthesizing Quantitative Performance Benchmarks: To assist researchers in bridging the sim-to-real gap, this work extracted and synthesized concrete, quantitative baseline performance metrics reported in physical experiments (force tracking RMSE of $0.5\mathrm{N}-1.5\mathrm{N}$, trajectory RMSE under $2.0\mathrm{mm}$, and finished surface roughness $Ra$ of $0.1\mu \mathrm{m}-0.3\mu \mathrm{m}$). These benchmarks provide a standard against which simulated policies can be verified and validated.
• Structuring Algorithmic Trade-offs: By categorizing the LfD algorithms into five methodological families (Table 6) and comparing their characteristics (Table 10), this study maps which algorithms are best suited for specific challenges. For instance, it highlights that while DMPs excel at spatial scaling, probabilistic models (like GMM-GMR) are better suited for online trajectory deformation during human intervention (HITL), and generative AI is best for visual point-cloud mapping, allowing practitioners to select the optimal control architecture.

5. Limitations of this Study

The findings and sectoral implications of this systematic mapping study should be interpreted in light of several methodological limitations:

• Database Coverage and Search Strategy: The literature search was restricted to Scopus, Web of Science, IEEE Xplore, and Google Scholar. While these are the primary repositories for engineering and robotics research, some publications indexed in regional or specialized databases may have been omitted. Additionally, the keyword-based search queries centered around terms like "Learning from Demonstration" and "DMP" might have missed relevant studies that utilize alternative terminologies, such as "skill transfer" or "human-robot co-manipulation," despite sharing the same underlying architecture.
• Language Bias: In accordance with the screening protocol, only articles published in the English language were included. Given that 62.5% of the included literature originates from China, and that countries like Japan and Germany possess strong academic and industrial backgrounds in robotic manufacturing, excluding non-English publications likely introduced a language bias. Highly innovative papers published in Chinese, Japanese, or German may have been overlooked.
• Exclusion of Grey Literature: To ensure scientific rigor, only peer-reviewed journal articles and conference proceedings were included, while patents, technical white papers, and corporate reports were excluded. Because real-world industrial implementations of LfD are frequently protected as commercial trade secrets, the exclusion of grey literature may have limited our ability to map the exact degree of current commercial adoption.
• Selection Bias and Single Screener Limitation: Since the screening, eligibility selection, and data extraction processes were conducted by a single primary researcher rather than by two independent reviewers as recommended by the PRISMA 2020 [1] guidelines [1], there is an inherent risk of selection bias. Although borderline or ambiguous cases were re-evaluated multiple times to ensure coding consistency and adherence to the eligibility protocol, the lack of a second independent auditor represents a methodological limitation.

6. Future Directions

Based on the synthesized evidence, future research in LfD-based robotic surface finishing should prioritize three key technological integrations:

6.1. Multimodal Perception and High-Dimensional Datasets

Future LfD frameworks must transition from relying on simple 1-D force profiles to fusing rich, high-dimensional multimodal sensory data. Fusing markerless RGB-D tracking, 3D point clouds, and dual-force haptic feedback will enhance environmental context awareness. Furthermore, integrating acoustic emission signals directly into the control loop will allow robots to monitor material-specific machining states (e.g., surface delamination or tool chatter) in real time. To support these frameworks, the research community must develop large-scale, open-access demonstration datasets covering diverse workpiece geometries, and design computationally efficient algorithms capable of processing high-dimensional inputs without the scalability bottlenecks associated with classical Gaussian Processes.

6.2. Bridging the Sim-to-Real Gap and Safe Exploration

While Deep Reinforcement Learning (DRL) and Diffusion Policies offer powerful tools for policy generation, their reliance on simulation presents a major bottleneck due to the difficulty of simulating contact physics. Future research must focus on developing lightweight online adaptation mechanisms (lifelong learning) that allow robots to refine their skills directly on the workshop floor under drifting environmental conditions. Furthermore, deploying these systems in unstructured environments necessitates the integration of safe-exploration strategies, such as Control Barrier Functions (CBFs), to autonomously detect and recover from extreme-force failure cases during contact-rich physical interactions.

6.3. Interactive Human-in-the-Loop (HITL) Systems

To handle sensor uncertainties and human demonstration variations, future architectures should transition from offline one-shot training to interactive online learning. Implementing HITL correction interfaces will enable human operators to physically intervene during execution, correct localized trajectory segments, and tune variable impedance parameters (such as stiffness or damping) in real time. This interactive optimization loop will improve generalization across distinct object categories without requiring full task re-demonstrations, significantly enhancing the flexibility of rajectory morphing.t

7. Conclusions

This systematic mapping study provides a comprehensive, PRISMA 2020 [1]-compliant synthesis of Learning from Demonstration (LfD) methodologies applied to robotic deburring and polishing between 2016 and 2026. By analyzing 24 primary peer-reviewed studies, this work established a structured taxonomy categorizing LfD approaches into five methodological clusters, with Dynamic Movement Primitives (DMPs) and probabilistic models representing the dominant frameworks, and deep generative policies (e.g., Diffusion Policies, Neural ODEs) emerging as the state of the art.

The synthesis of mechatronic configurations highlights that force/torque sensing remains the foundational sensory modality (70.8%), integrated primarily with variable impedance and admittance control architectures to guarantee physical interaction compliance. Quantitatively, the state of the art demonstrates high performance in laboratory environments, achieving force tracking errors (RMSE) between $0.5\mathrm{N}$ and $1.5\mathrm{N}$, trajectory reproduction errors under $2.0\mathrm{mm}$ ($r>0.95$), and surface roughness ($Ra$) reductions to $0.1–0.3\mu \mathrm{m}$.

However, several technical barriers must be resolved before LfD can be widely adopted on factory floors. These include the sim-to-real gap in contact physics, the high-frequency impact dynamics of deburring, the autonomous avoidance of local non-polishing areas (LNP-areas), and the lack of intuitive online human-in-the-loop correction interfaces. Addressing these gaps will facilitate the transition of LfD from a laboratory concept to a flexible, cost-effective automation tool, particularly for SMEs operating in high-mix, low-volume manufacturing environments.