Optimizing Force Signals from Human Demonstrations of In-Contact Motions

arXiv:2507.15608 · cs.RO · Submitted 2025-07-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Optimizing Force Signals from Human Demonstrations of In-Contact Motions".

Paper summary: Dev: Looking at the paper, "Optimizing Force Signals from Human Demonstrations of In-Contact Motions," it’s clear that while kinesthetic guiding is intuitive for robot programming, the raw force data it produces isn't always usable directly.

Rosa: The authors are tackling this by comparing different filtering techniques and proposing a way to detect those unwanted contact peaks in the signal >

Taro: And they found that optimizing these signals can increase the quality of an individual motion by up to twenty percent depending on their error criterion >

Dev: The implication for me is that we can't just take whatever force data comes from a demonstration and feed it into our control system without some kind of pre-processing step >

Rosa: Right. It suggests that we need a way to clean up the input so it corresponds better to the user's actual goal, which is what this paper is all about >

Taro: So, for someone just listening, it means when you're teaching a robot how to interact with something by hand guiding, you might need to be mindful of signal noise and choose your filtering method carefully based on the motion type >

Dev: And they did flag something important: they noted that additional information about the motion itself needs to be considered or user assistance is required when comparing those critical parameters >

Rosa: So, in short, this paper provides a framework for optimizing force signals from human demonstrations to get them closer to the human intention >

Conclusion: Rosa: So we've looked at how this paper tries to clean up those noisy force signals from human demonstrations using two steps: first finding and cutting out those unwanted contact peaks, and then applying different smoothing methods >

Dev: Yeah, it’s about taking that raw data from a robot being guided by a person and trying to make sure the robot's resulting action actually matches what the human intended >

Rosa: The authors are asking how much we can actually improve that signal quality, and they found they could boost the motion quality by up to twenty percent depending on how you measure it >

Dev: That percentage sounds nice, but you have to look at the trade-offs because they point out that positional difficulty matters more than the force profile difficulty when you’re choosing a filter >

Rosa: They give examples, like for straight line movements, they need a certain cutoff frequency of zero point one five hertz for the Sinc in time filter >

Dev: And then for just pointing at a spot, that same cutoff frequency has to be six point five hertz, which shows how much the context of the motion changes what you need >

Rosa: It seems like they’re saying you can't just use one universal filter setting because the underlying motion itself dictates what works best >

Dev: Exactly. They end by saying that you either need more information about the specific motion or you need some kind of user assistance when picking those critical parameters >

Rosa: It makes me wonder how this filtering approach holds up when we're trying to use machine learning to teach robots new skills from demonstrations >

Johannes Hartwig, Fabian Viessmann, Dominik Henrich

University of Bayreuth

cs.RO

Submitted: 2025-07-21

Updated: 2025-08-14

Comments: This is a preprint of a chapter the following work (and accepted for publication): Annals of Scientific Society for Assembly, Handling and Industrial Robotics 2024. The final authenticated version is will be linked here: http://dx.doi.org/[tba]

DOI: 10.1007/978-3-031-91463-8_7

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 76/100

The gist: Optimizing force signals from human demonstrations of in-contact motions explores methods to refine noisy input signals from kinesthetic guiding to better match human intention <ref:2507.15608#pg2>.

Key concepts

Contact Peaks
These are specific moments in a force signal where the rate of change (derivative) flips sign, indicating a distinct contact event. The paper identifies these peaks using techniques like relative thresholds or z-scores to isolate them from general noise in the measured force data.
Sinc-in-time Filter
This is one smoothing technique where the effectiveness depends on a cutoff frequency (fc). A lower fc value means more aggressive smoothing, while a higher fc allows for less smoothing. The optimal fc varies significantly depending on whether the demonstration involves linear or point motions.
Savitzky-Golay Filter
This filter uses a window width (h) as its main parameter to smooth the force signal. A wider window results in greater smoothing of the data, effectively reducing noise. The paper suggests this technique is useful for different types of demonstrations based on their positional difficulty.

Terminology

Summary

Optimizing force signals from human demonstrations of in-contact motions explores methods to refine noisy input signals from kinesthetic guiding to better match human intention <ref:2507. p. 0

Where the paper says it

However, imprecise and noisy input signals from human demonstrations pose problems when reproducing motions directly or using the signal as input for machine learning methods. This paper explores optimizing force signals to correspond better to the human intention of the demonstrated signal.

15608#pg2>. This paper examines the question of to what extent measured force signals can be optimized so that they correspond better to the human intention of the demonstrated signal <ref:2507. p. 1
Where the paper says it

We will explore these methods, applying their fundamental approach to our inaccurate signal. Therefore, this paper examines the question of to what extent measured force signals can be optimized so that they correspond better to the human intention of the demonstrated signal.

15608#pg3>.

Introduction and Problem Statement

The need for flexible and agile automation using cobots is increasing, making robot programming more accessible to small and medium-sized enterprises <ref:2507.15608#pg2>. Utilizing the programming by demonstration (PbD) paradigm with kinesthetic guiding allows humans to program robot trajectories without prior knowledge of robot programming <ref:2507. p. 0

Where the paper says it

Therefore, a future-oriented goal is to improve the usability of robot programming and the interaction between humans and robots [4]. Utilizing the programming by demonstration (PbD) paradigm, specifically with kinesthetic guiding, can enable humans to program robot trajectories without prior robot programming knowledge [5].

15608#pg2>. However, specific tasks like polishing or grinding complex shapes require continuous contact, necessitating a method to specify contact forces over time and space <ref:2507. p. 1
Where the paper says it

2 Johannes Hartwig, Fabian Viessmann, Dominik Henrich Nevertheless, specific tasks can only be adequately addressed partially through trajectory programming, such as when a robot task consists of polishing [6], engraving [7], or grinding complex shapes [8]. Here, continuous contact between the tool and the object becomes imperative, and a method to specify the contact forces to be applied over time and space is needed.

15608#pg3>. The force profiles generated by humans during demonstrations are often noisy or inaccurate due to factors such as unfamiliar guidance and sensor noise <ref:2507. p. 1
Where the paper says it

However, the force profiles generated by humans during the demonstration are often noisy or inaccurate (see Fig. A prominent approach here is to learn from multiple demonstrations, often combining methods such as dynamic time warping (DTW), Gaussian mixture models (GMMs), or dynamical systems.

15608#pg4>.

Methodology for Signal Optimization

The approach involves a two-step process: first, identifying and removing unwanted contact peaks from the force signal, and second, applying different smoothing algorithms to generate an optimized signal <ref:2507.15608#pg5>. Contact peaks are formally defined as a sequence where the derivative changes sign during the contact peak's time interval <ref:2507.15608#pg6>. The identification of these peaks uses conditions such as a relative threshold to set the difference in magnitude between forces and a z-score to identify outliers <ref:2507. p. 4

Where the paper says it

When dealing with these unwanted short peaks, it must generally hold that the force ku at the end of the peak u corresponds approximately to the force kw at the beginning of the peak w. For this purpose, we use a relative threshold to set the difference in magnitude between the forces ku and kw to the total peak height |ku − kv|.

15608#pg7>.

Filtering Techniques Compared

Three filtering techniques are considered for signal smoothing: Sinc-in-time filter, Savitzky-Golay filter, and RANSAC line approximation <ref:2507.15608#pg7>. The Sinc-in-time filter uses a cutoff frequency fc as its critical parameter, where a lower cutoff frequency results in greater smoothing <ref:2507. p. 5

Where the paper says it

Therefore, this is an optimal low-pass filter, and the critical parameter of this algorithm is the cutoff frequency fc. The lower the cutoff frequency is, the greater the smoothing, and vice versa.

15608#pg7>. The Savitzky-Golay filter's critical parameter is the window width h; a wider window leads to greater smoothing <ref:2507.15608#pg7>. For RANSAC line approximation, the critical parameter is the maximum segment length l; a greater length l results in more smoothing <ref:2507. p. 5
Where the paper says it

This length limits the time interval in which the respective model applies. Therefore, the critical parameter of this smoothing algorithm is the maximum segment length l; The greater the length l is, the greater the smoothing, and vice versa.

15608#pg7>.

Evaluation and Findings

Evaluation across 10 users demonstrated that the quality for an individual motion could be increased by up to 20 % concerning the error criterion <ref:2507.15608#pg2>. The difficulty of the positional motion should be weighted more heavily than the difficulty of force profile when considering smoothing <ref:2507. p. 7

Where the paper says it

Furthermore, the minimum error for linear and circular demonstrations is still much higher than for point demonstrations. From this, we conclude that the difficulty of the positional motion should be weighted more heavily than the difficulty of force profile.

15608#pg10>. For instance, for linear demonstrations, the minimum error for Sinc-in-time is fc = 0. p. 7
Where the paper says it

This leads to the conclusion that the smoothing must be increased if the human moves the robot in addition to exerting force. Furthermore, the minimum error for linear and circular demonstrations is still much higher than for point demonstrations.

15 Hz, while for point demonstrations it is fc = 6. p. 7
Where the paper says it

Here, whether these peaks belong in the false positive class is unclear. We observed that it is challenging to exert a force while the user has to move the robot.

5 Hz <ref:2507. p. 0
Where the paper says it

†These authors contributed equally. arXiv:2507.

15608#pg10>. The paper concludes that additional information about the motion must be considered or user assistance is required when comparing critical parameters <ref:2507. p. 8
Where the paper says it

Thus, additional information about the motion must be considered, or user assistance is required. 4 Conclusion This paper compares different methods to optimize force signals from kinesthetic guiding.

15608#pg10>.

Conclusion and Future Work

The study proposes a method to remove contact peaks and filtering techniques to make the smoothed signal correspond more closely to human intention <ref:2507.15608#pg5>. High degrees of filtering are required, especially for motions where the human demonstrator exerts force and moves the robot simultaneously. p. 8

Where the paper says it

Two different approaches improve the error measure and might be helpful for varying use cases. Although significant improvement is achieved, a high degree of filtering is required, particularly for motions in which the human demonstrator exerts a force and moves the robot simultaneously.

Future work should address modeling critical parameters dependent on positional motion and investigate the advantages of such filtering when learning robot motions utilizing machine learning methods. p. 8
Where the paper says it

Therefore, future work should address the modeling of critical parameters dependent on the positional motion. In addition, it should be investigated to what extent such filtering offers advantages when learning robot motions utilizing machine learning methods.

REFERENCES

  1. Dietz, T, et al. p. 9

    Where the paper says it

    10 Johannes Hartwig, Fabian Viessmann, Dominik Henrich References 1. Dietz, T, et al.

    (2012). Programming system for efficient use of industrial robots for deburring in SME environments. p. 9
    Where the paper says it

    Dietz, T, et al. Programming system for efficient use of industrial robots for deburring in SME environments.

    In 7th German Conference on Robotics. p. 9
    Where the paper says it

    Programming system for efficient use of industrial robots for deburring in SME environments. In 7th German Conference on Robotics.

    VDE.<ref:2507.15608#pg2>

  2. Riedelbauch, D, Hartwig, J, Henrich, D (2019). p. 9

    Where the paper says it

    In 7th German Conference on Robotics. Riedelbauch, D, Hartwig, J, Henrich, D (2019).

    Enabling End-Users to Deploy Flexible Human-Robot Teams to Factories of the Future. p. 9
    Where the paper says it

    Riedelbauch, D, Hartwig, J, Henrich, D (2019). Enabling End-Users to Deploy Flexible Human-Robot Teams to Factories of the Future.

    In IROS 2019 Workshop Factory of the Future. p. 9
    Where the paper says it

    Enabling End-Users to Deploy Flexible Human-Robot Teams to Factories of the Future. In IROS 2019 Workshop Factory of the Future.

    IEEE.<ref:2507.15608#pg2>

  3. Kildal J, et al. p. 9

    Where the paper says it

    IEEE Robotics & Automation Magazine, 26(1). Kildal J, et al.

    (2018). Potential users’ key concerns and expectations for the adoption of cobots. p. 9
    Where the paper says it

    Kildal J, et al. Potential users’ key concerns and expectations for the adoption of cobots.

    Procedia CIRP 72. p. 9
    Where the paper says it

    Potential users’ key concerns and expectations for the adoption of cobots. Procedia CIRP 72.

    <ref:2507.15608#pg5>

  4. Villani V, et al. p. 9

    Where the paper says it

    Procedia CIRP 72. Villani V, et al.

    (2018). Survey on human–robot collaboration in industrial settings: Safety, intuitive interfaces and applications. p. 9
    Where the paper says it

    Villani V, et al. Survey on human–robot collaboration in industrial settings: Safety, intuitive interfaces and applications.

    Mechatronics 55. p. 9
    Where the paper says it

    Survey on human–robot collaboration in industrial settings: Safety, intuitive interfaces and applications. Mechatronics 55.

    <ref:2507.15608#pg2>

  5. Amanhoud, W, et al. p. 9

    Where the paper says it

    Mechatronics 55. Amanhoud, W, et al.

    (2019). A dynamical system approach to motion and force generation in contact tasks. p. 9
    Where the paper says it

    Amanhoud, W, et al. A dynamical system approach to motion and force generation in contact tasks.

    In Robotics: Science and Systems. p. 0
    Where the paper says it

    Optimizing Force Signals from Human Demonstrations of In-Contact Motions Johannes Hartwig † , Fabian Viessmann † , and Dominik Henrich Chair for Applied Computer Science III (Robotics and Embedded Systems), University of Bayreuth, D-95440 Bayreuth, Germany {johannes. hartwig|fabian.

    <ref:2507.15608#pg3>

  6. Koropouli, V, et al. p. 9

    Where the paper says it

    In Robotics: Science and Systems. Koropouli, V, et al.

    (2011). Learning interaction control policies by demonstration. p. 9
    Where the paper says it

    Koropouli, V, et al. Learning interaction control policies by demonstration.

    In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems. p. 9
    Where the paper says it

    Learning interaction control policies by demonstration. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems.

    <ref:2507.15608#pg4>

  7. Jinno, M, et al. p. 9

    Where the paper says it

    In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems. Jinno, M, et al.

    (1995). Development of a force controlled robot for grinding, chamfering and polishing. p. 9
    Where the paper says it

    Jinno, M, et al. Development of a force controlled robot for grinding, chamfering and polishing.

    In Proceedings of 1995 IEEE International Conference on Robotics and Automation. p. 9
    Where the paper says it

    Development of a force controlled robot for grinding, chamfering and polishing. In Proceedings of 1995 IEEE International Conference on Robotics and Automation.

    IEEE.<ref:2507.15608#pg5>

  8. Bargmann, D, et al. p. 9

    Where the paper says it

    In Proceedings of 1995 IEEE International Conference on Robotics and Automation. Bargmann, D, et al.

    (2021). Unobstructed programming-by-demonstration for force-based assembly utilizing external force-torque sensors. p. 9
    Where the paper says it

    Bargmann, D, et al. Unobstructed programming-by-demonstration for force-based assembly utilizing external force-torque sensors.

    In 2021 IEEE International Conference on Systems, Man, and Cybernetics. p. 9
    Where the paper says it

    Unobstructed programming-by-demonstration for force-based assembly utilizing external force-torque sensors. In 2021 IEEE International Conference on Systems, Man, and Cybernetics.

    IEEE.<ref:2507.15608#pg10>

  9. Yang, C, et al. p. 9

    Where the paper says it

    In 2021 IEEE International Conference on Systems, Man, and Cybernetics. Yang, C, et al.

    (2018). A learning framework of adaptive manipulative skills from human to robot. p. 9
    Where the paper says it

    Yang, C, et al. A learning framework of adaptive manipulative skills from human to robot.

    IEEE Transactions on Industrial Informatics, 15(2). p. 9
    Where the paper says it

    A learning framework of adaptive manipulative skills from human to robot. IEEE Transactions on Industrial Informatics, 15(2).

    <ref:2507.15608#pg10>

  10. Montebelli, et al. p. 9

    Where the paper says it

    IEEE Transactions on Industrial Informatics, 15(2). Montebelli, et al.

    (2015). On handing down our tools to robots: Single-phase kinesthetic teaching for dynamic in-contact tasks. p. 9
    Where the paper says it

    Montebelli, et al. On handing down our tools to robots: Single-phase kinesthetic teaching for dynamic in-contact tasks.

    In 2015 IEEE International Conference on Robotics and Automation. p. 9
    Where the paper says it

    On handing down our tools to robots: Single-phase kinesthetic teaching for dynamic in-contact tasks. In 2015 IEEE International Conference on Robotics and Automation.

    IEEE.<ref:2507.15608#pg10>

  11. Steinmetz, F, et al. p. 9

    Where the paper says it

    In 2015 IEEE International Conference on Robotics and Automation. Steinmetz, F, et al.

    (2015). Simultaneous kinesthetic teaching of positional and force requirements for sequential in-contact tasks. p. 9
    Where the paper says it

    Steinmetz, F, et al. Simultaneous kinesthetic teaching of positional and force requirements for sequential in-contact tasks.

    In 2015 IEEE-RAS 15th International Conference on Humanoid Robots. p. 9
    Where the paper says it

    Simultaneous kinesthetic teaching of positional and force requirements for sequential in-contact tasks. In 2015 IEEE-RAS 15th International Conference on Humanoid Robots.

    IEEE.<ref:2507.15608#pg10>

  12. Conkey, A, Hermans, T (2019). p. 9

    Where the paper says it

    In 2015 IEEE-RAS 15th International Conference on Humanoid Robots. Conkey, A, Hermans, T (2019).

    Learning task constraints from demonstration for hybrid force/position control. p. 9
    Where the paper says it

    Conkey, A, Hermans, T (2019). Learning task constraints from demonstration for hybrid force/position control.

    In 2019 IEEE-RAS 19th International Conference on Humanoid Robots. p. 9
    Where the paper says it

    Learning task constraints from demonstration for hybrid force/position control. In 2019 IEEE-RAS 19th International Conference on Humanoid Robots.

    IEEE.<ref:2507.15608#pg10>

  13. Suomalainen, M, et al. p. 9

    Where the paper says it

    In 2019 IEEE-RAS 19th International Conference on Humanoid Robots. Suomalainen, M, et al.

    (2022). A survey of robot manipulation in contact. p. 9
    Where the paper says it

    Suomalainen, M, et al. A survey of robot manipulation in contact.

    Robotics and Autonomous Systems, 156. p. 9
    Where the paper says it

    A survey of robot manipulation in contact. Robotics and Autonomous Systems, 156.

    <ref:2507.15608#pg10>

  14. Hartwig, J, et al. p. 0

    Where the paper says it

    Optimizing Force Signals from Human Demonstrations of In-Contact Motions Johannes Hartwig † , Fabian Viessmann † , and Dominik Henrich Chair for Applied Computer Science III (Robotics and Embedded Systems), University of Bayreuth, D-95440 Bayreuth, Germany {johannes. hartwig|fabian.

    p. 0
    Where the paper says it

    Optimizing Force Signals from Human Demonstrations of In-Contact Motions Johannes Hartwig † , Fabian Viessmann † , and Dominik Henrich Chair for Applied Computer Science III (Robotics and Embedded Systems), University of Bayreuth, D-95440 Bayreuth, Germany {johannes. hartwig|fabian.

    (2023). Visualization of Forces and Torques for RobotProgramming of In-Contact Tasks. p. 9
    Where the paper says it

    Hartwig, J, et al. Visualization of Forces and Torques for RobotProgramming of In-Contact Tasks.

    Annals of Scientific Society for Assembly, Handling and Industrial Robotics 2023 (to appear). p. 9
    Where the paper says it

    Visualization of Forces and Torques for RobotProgramming of In-Contact Tasks. Annals of Scientific Society for Assembly, Handling and Industrial Robotics 2023 (to appear).

    <ref:2507.15608#pg2>

  15. Hartwig, J, et al. (2023). Input and Editing of Force Profiles of In-Contact Robot Motions via a Touch Graphical User Interface. p. 9

    Where the paper says it

    Hartwig, J, et al. Input and Editing of Force Profiles of In-Contact Robot Motions via a Touch Graphical User Interface.

    In 2023 IEEE International Conference on Robotic Computing (to appear). p. 9
    Where the paper says it

    Input and Editing of Force Profiles of In-Contact Robot Motions via a Touch Graphical User Interface. In 2023 IEEE International Conference on Robotic Computing (to appear).

    IEEE.<ref:2507.15608#pg2>

  16. Cooley, JW, Tukey, JW (1965). p. 9

    Where the paper says it

    In 2023 IEEE International Conference on Robotic Computing (to appear). Cooley, JW, Tukey, JW (1965).

    An algorithm for the machine calculation of complex Fourier series. p. 9
    Where the paper says it

    Cooley, JW, Tukey, JW (1965). An algorithm for the machine calculation of complex Fourier series.

    Mathematics of computation, 19(90). p. 9
    Where the paper says it

    An algorithm for the machine calculation of complex Fourier series. Mathematics of computation, 19(90).

    <ref:2507.15608#pg7>

  17. Savitzky A, et al. p. 1

    Where the paper says it

    , using Fourier Analysis [17]), averaging over filter windows (e. , Savitzky-Golay filter [18]), or fit primitives (e.

    (1964). Smoothing and Differentiation of Data by Simplified Least Squares Procedures. p. 9
    Where the paper says it

    Savitzky A, et al. Smoothing and Differentiation of Data by Simplified Least Squares Procedures.

    Analytical Chemistry 36. p. 9
    Where the paper says it

    Smoothing and Differentiation of Data by Simplified Least Squares Procedures. Analytical Chemistry 36.

    <ref:2507.15608#pg7>

  18. Fischler M, et al. p. 9

    Where the paper says it

    Analytical Chemistry 36. Fischler M, et al.

    (1981). Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. p. 9
    Where the paper says it

    Fischler M, et al. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.

    Communications of the ACM 24. p. 9
    Where the paper says it

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24.

    <ref:2507.15608#pg7>

  19. Joseph, SB, et al. p. 9

    Where the paper says it

    Communications of the ACM 24. Joseph, SB, et al.

    (2022). Metaheuristic algorithms for PID controller parameters tuning: Review, approaches and open problems. p. 9
    Where the paper says it

    Joseph, SB, et al. Metaheuristic algorithms for PID controller parameters tuning: Review, approaches and open problems.

    Heliyon, 8(5). p. 9
    Where the paper says it

    Metaheuristic algorithms for PID controller parameters tuning: Review, approaches and open problems. Heliyon, 8(5).

    <ref:2507.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Peak detection in force signals can be improved by using a relative threshold to set the difference in magnitude between forces ku and kw to total peak height ku − kv ≤ δ, which helps identify unwanted short peaks more accurately during demonstrations.

  2. The system can be enhanced by applying different smoothing algorithms—Sinc-in-time filter, Savitzky-Golay filter, and RANSAC line approximation—and selecting the critical parameter that minimizes the error criterion J(S, Sˆ) to find an optimized signal S˜ that corresponds better to human intention.

  3. The improved AI system can perform playback programming or serve as a signal pre-processing for machine learning methods by minimizing the error between a force signal S from a human demonstration and an ideal signal Sˆ that meets the intentions of the user.

Abstract

For non-robot-programming experts, kinesthetic guiding can be an intuitive input method, as robot programming of in-contact tasks is becoming more prominent. However, imprecise and noisy input signals from human demonstrations pose problems when reproducing motions directly or using the signal as input for machine learning methods. This paper explores optimizing force signals to correspond better to the human intention of the demonstrated signal. We compare different signal filtering methods and propose a peak detection method for dealing with first-contact deviations in the signal. The evaluation of these methods considers a specialized error criterion between the input and the human-intended signal. In addition, we analyze the critical parameters' influence on the filtering methods. The quality for an individual motion could be increased by up to percent concerning the error criterion. The proposed contribution can improve the usability of robot programming and the interaction between humans and robots.

Related papers