May 26, 2026

Human-to-Robot Skill Transfer using LfD: A Systematic Review

DOI
https://dx.doi.org/10.17504/protocols.io.261geqppog47/v1
  • Francisco Neves1,
  • Pedro Afonso Dias2,
  • Artur J. Cordeiro2,
  • Benedita Malheiro1
  • 1ISEP, Polytechnic of Porto;
  • 2INESC TEC, Institute for Systems and Computer Engineering, Technology and Science
Francisco Neves
Icon indicating open access to content
QR code linking to this content
DOI: https://dx.doi.org/10.17504/protocols.io.261geqppog47/v1
Protocol Citation: Francisco Neves, Pedro Afonso Dias, Artur J. Cordeiro, Benedita Malheiro 2026. Human-to-Robot Skill Transfer using LfD: A Systematic Review. protocols.io https://dx.doi.org/10.17504/protocols.io.261geqppog47/v1
License: This is an open access  protocol  distributed under the terms of the  Creative Commons Attribution License,  which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited
Protocol status: Working
We use this protocol and it's working
Created: May 26, 2026
Last Modified: May 26, 2026
Protocol  Integer ID: 317954
Keywords: robot skill transfer, robot skill transfer in the context, learning from demonstration, complex manipulation task, robotic manipulation, teleoperation, inverse reinforcement learning, systematic review industrial robotics, teaching modalities such as kinesthetic guiding, adaptable robotic system, foundation for adaptable robotic system, behavior cloning, including behavior cloning, learning, kinesthetic guiding, intuitive mechanism for human, core learning, demonstration, teaching modality
Abstract
Industrial robotics increasingly operates in complex dynamic environments that require high flexibility, making traditional programming methods obsolete. This research analyzes Learning from Demonstration as an intuitive mechanism for human-to-robot skill transfer in the context of robotic manipulation. A systematic literature review guided by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework examines the state of the art in Learning from Demonstration. The study evaluates core learning algorithms, including Behavior Cloning and Inverse Reinforcement Learning, together with teaching modalities such as kinesthetic guiding and teleoperation. The analysis also addresses fundamental theoretical challenges, including the correspondence problem and policy generalization. The findings establish a foundation for adaptable robotic systems that execute complex manipulation tasks.
Abstract
This protocol defines the methodological framework for a systematic literature review focused on Learning from Demonstration (LfD) methods and their applications in robotic manipulation. The study aims to analyze the evolution, methodological trends, teaching modalities, and theoretical challenges associated with human-to-robot skill transfer in dynamic industrial environments. Publications were retrieved from major scientific databases and analyzed through a structured review process guided by the PRISMA 2020 framework. The research examines core learning approaches, including Behavior Cloning and Inverse Reinforcement Learning, together with teaching modalities such as kinesthetic guiding and teleoperation. Additionally, the study investigates fundamental challenges including the correspondence problem and policy generalization, providing a comprehensive overview of the current state and future directions of adaptable robotic manipulation systems.
PICOC

Population

  • Stationary (fixed-base) robotic manipulators.

Intervention

  • LfD algorithms.

Comparison

  • Traditional hard-coded programming;
  • RL without prior demonstrations;
  • Hybrid LfD and RL architectures;
  • Cross-modality LfD comparisons.

Outcome

  • Learning/sample efficiency and overhead;
  • Task success and trajectory accuracy;
  • Generalization to novel parameters.

Context

  • Industrial applications of robot learning, focusing on LfD techniques for manipulation.
Research Questions
  1. Which LfD methods are applied to robotic manipulators?
  2. Which demonstration modalities and sensors are most common?
  3. Which evaluation metrics and benchmarks are used for comparability?
  4. What are the main technical limitations and open research directions?
Inclusion and Exclusion Criteria
Inclusion Criteria
  • 2020–2026
  • Language English full-text
  • Venue Peer-reviewed (IROS, ICRA, CVPR, CoRL)
  • Method LfD, IL, or PbD focus
  • Hardware Robotic manipulators
  • Validation Empirical manipulation tasks

Exclusion Criteria
  • Prior to 2020
  • Non-English
  • Non-peer-reviewed (books, workshops)
  • No demonstration-based learning
  • Non-manipulation platforms
  • Purely theoretical/ No experiments
Quality Assessment
  1. Is the study from a peer-reviewed, top-tier venue?
  2. Is the experimental setup clearly described?
  3. Are the robot manipulators well-specified?
  4. Is the LfD method clearly defined (algorithm, modality)?
  5. Are the demonstrations clearly described?
  6. Is a baseline comparison (e.g., RL, traditional) included?
  7. Are performance outcomes reported?
  8. Are efficiency outcomes (e.g., sample/time) reported?
  9. Are generalization or robustness outcomes reported?
  10. Is there a sufficient number of trials or demonstrations?
  11. Is the study reproducibility documented?
  12. Are technical limitations explicitly acknowledged?

Scoring criteria were defined as follows: 0 indicating no compliance, 0.5 indicating partial compliance, and 1 indicating full compliance.
Search Query and Databases
General Query

("learn* from demonstration" OR "learn* by demonstration" OR "imitation learning") AND (manipulat* OR "robot arm") AND PublicationYear >= 2020 AND PublicationYear <= 2026 Search in Title, Abstract, and Keywords

Web of Science

TI=(("learn* from demonstration" OR "learn* by demonstration" OR "imitation learning") AND (manipulat* OR "robot arm"))

Scopus

TITLE ((learn* W/3 (demonstration OR imitation)) AND (manipulat* OR "robot arm")) AND PUBYEAR > 2019 AND PUBYEAR < 2026 TITLE (("imitation learning" OR "learning from demonstration") AND ("robot manipulation" OR "robotic arm")) AND PUBYEAR > 2019 AND PUBYEAR < 2026 TITLE (("imitation learning" OR "learning from demonstration") AND (policy OR "end-to-end") AND (manipulat* OR "robot arm")) AND PUBYEAR > 2019 AND PUBYEAR < 2026

IEEE Xplore

(("Document Title":learn* AND ("Document Title":demonstration OR "Document Title":imitation) AND ("Document Title":manipulat* OR "Document Title":"robot arm")))
Screening Process
- The PRISMA 2020 methodology was followed for systematic selection and using the framework Parsifal.
- Articles were filtered using keyword queries, then screened manually for topical relevance.
- Duplicates, non-relevant, and retracted records were excluded.
- The final selection is illustrated bellow.