Design and Experimental Validation of Sensorless 4-Channel Bilateral Teleoperation for Low-Cost Manipulators

arXiv:2507.06174 · cs.RO, cs.AI, cs.SY, eess.SY · Submitted 2025-07-08 · 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: "Design and Experimental Validation of Sensorless 4-Channel Bilateral Teleoperation for Low-Cost Manipulators".

Title and authors: Rosa: So we're looking at this paper, "Design and Experimental Validation of Sensorless four-Channel Bilateral Teleoperation for Low-Cost Manipulators <ref:2507.06174#pg1,Design and Experimental Validation of Sensorless 4-Channel Bilateral Teleoperation for Low>." It’s by Koki Yamane and his team, which is interesting because it tackles a real problem: how to get stable control when you can't afford expensive force sensors on a robot.

Dev: Exactly. The core idea here is that they've designed this sensorless four-channel bilateral teleoperation framework, and the big takeaway for us as engineers is showing that you can achieve stable position-force interaction on low-cost manipulators without those expensive force sensors, which is a huge win for hardware constraints.

Taro: I'm curious about how they handle the limitations of using only low-resolution encoder measurements and input torque to estimate both joint velocity and external force simultaneously, because that sounds like a real bottleneck when you're running high-speed tasks.

Rosa: Well, the paper explains their system by starting with a rigid serial link model where the dynamics are written as M(q) times acceleration plus C times velocity plus D times velocity plus g equals the commanded torque plus external torque, which is just standard stuff, but they build on top of that to handle this sensorless thing.

Dev: They use a disturbance observer or an extended state observer to estimate both the joint velocity and the external force at the same time using only position measurements and input torque, which they call tau ext, and their state space includes a disturbance term d defined as M times inverse q times the external torque plus some compensation term, which is basically how they model what's happening inside the robot.

Taro: So, when you look at this structure in the frequency domain, what’s actually new about how they tune these two estimators—the velocity one and the force one?

Rosa: That’s a key part of their contribution. They interpret the observer structure in the frequency domain, and they clarify that there's a direct coupling between how fast those velocity estimation bandwidths are running and how fast the external force estimation bandwidth is running, which lets them derive practical tuning guidelines based on just one cutoff angular frequency and the damping ratio ζ.

Title and authors: Dev: That sounds like a huge simplification for anyone trying to tune these things in the real world, because it means you don't have to juggle two separate tuning parameters for velocity and force estimation bandwidths; you can focus on that single cutoff frequency.

Taro: If they simplify the tuning freedom down to just one cutoff frequency, what does that actually tell us about how stable the system is when we push it into high-speed or contact-rich scenarios?

Rosa: Their experiments on a real low-cost manipulator show that this framework provides practically useful force estimates and enables stable teleoperation even in high-speed and contact-rich scenarios where you’d normally expect instability, provided you pick the right parameters.

Dev: They also mention some specific values that are practical defaults: for free motion comparison, a cutoff frequency of fifty radians per second showed that the velocity estimate matched the offline differentiated velocity pretty closely.

Taro: And for those contact-rich tasks where things get messy, what's the recommended damping ratio ζ they found to be most useful?

Rosa: For contact-rich tasks, critical damping, which is a damping ratio of one point zero, is presented as a practical default choice because it gives the fastest response without any overshoot in the external force estimate.

Dev: That’s smart—critical damping for fast response without instability in that specific force estimation context. Now, what about the control law itself? How do they actually translate all those estimates into torque commands?

Taro: I want to know more about the final torque command they derive; does it still rely on compensating for nonlinear dynamics or is that compensation handled differently once they have these estimates?

Rosa: The final torque command incorporates the estimated disturbance, which means you have a term tau cmd that includes proportional-derivative control for the relative motion mode and a forcedriven motion for the common motion mode, plus this full disturbance compensation term involving the estimated force and velocity terms.

Dev: That structure is important because it shows they are explicitly compensating for those nonlinear dynamics—the inertia matrix M times acceleration plus Coriolis and gravity forces—while simultaneously feeding in the estimated external torque to stabilize the interaction.

Taro: If we look at what this means for imitation learning, how much does incorporating this force information into the demonstration data actually improve performance in complex tasks?

Title and authors: Rosa: The paper shows that incorporating force information into the input significantly improves task success rates across multiple manipulation tasks. For instance, in a dual-arm pick-and-place task, when they used force information for both the input and output, both picking operations were successful in all trials.

Dev: So, it's not just about getting a better position estimate; it’s that having the estimated force helps the learning policy succeed where it otherwise wouldn't have on those contact-rich tasks.

Taro: It sounds like they are showing that this approach is robust enough to handle model errors and external disturbances well, even when the underlying dynamics aren't perfectly known or when you're dealing with noisy sensor data.

Rosa: That’s the goal of using the disturbance observer—it suppresses high-frequency oscillatory behavior while still keeping the force-feedback teleoperation stable, which is crucial for practical deployment.

Dev: And on a practical note, they did some parameter identification by converting their nonlinear model into linear equations and using data sampled at five hundred hertz, then downsampling it to twenty-five hertz to estimate the gravity parameters and inertia parameters.

Taro: That parameter identification step is important because it shows how much you can actually extract usable dynamics from low-resolution measurements, even if you're starting from a simple model like a rigid serial link.

Rosa: The overall implication of this paper, "Design and Experimental Validation of Sensorless four-Channel Bilateral Teleoperation for Low-Cost Manipulators," is that high-fidelity teleoperation with force feedback is now achievable on low-cost hardware without needing expensive force sensors by using this specific framework <ref:2507.06174#pg1,Design and Experimental Validation of Sensorless 4-Channel Bilateral Teleoperation for Low>.

Dev: It moves the barrier away from just building a super expensive robot; it shows that the control strategy can compensate for hardware limitations by using clever observer design and frequency domain analysis to get stable performance.

Taro: For someone looking at this, it means you can collect much richer demonstration data for AI learning because you aren't limited by what kind of sensors your robot has to use.

Rosa: Exactly. The paper lays out a concrete path: use the disturbance observer, tune based on that single cutoff frequency and damping ratio, and you get a system that performs well in real-world scenarios like contact manipulation.

Title and authors: Dev: It's about making complex control problems manageable for low-cost hardware by linking the mathematical structure of the observer to practical tuning rules derived from frequency analysis.

Taro: So, for the listener who might be driving or cooking, it means that advanced motion planning and interaction skills can be learned from demonstrations even if the robot hardware is cheap, as long as you use a framework like this to get stable force feedback.

Rosa: That’s the practical application. The paper validates that this specific design works under real-world constraints, showing how sensorless systems can be made robust enough for complex interaction tasks.

Dev: It really highlights the power of combining nonlinear dynamics compensation with a disturbance observer to achieve high tracking performance and operability with a low-cost manipulator that cannot use high feedback gains.

Taro: So, to wrap up on this paper, "Design and Experimental Validation of Sensorless four-Channel Bilateral Teleoperation for Low-Cost Manipulators," it proves that practical high-speed bilateral teleoperation with force feedback is achievable on low-cost manipulators without force sensors by employing a sensorless four-channel bilateral control framework <ref:2507.06174#pg1,Design and Experimental Validation of Sensorless 4-Channel Bilateral Teleoperation for Low>.

Rosa: It's a really solid piece of work because it gives us the tuning guidelines, and the experimental validation with both free motion and contact tasks shows exactly where this system shines.

Dev: And we've seen how incorporating estimated force information into demonstrations improves success rates in those contact-rich manipulation tasks, which is a big win for the imitation learning side of things.

Taro: My final thought is that the coupling between the velocity and external force estimation bandwidths, clarified through the frequency domain analysis, gives us a much clearer way to tune these estimators than just guessing parameters.

Rosa: It definitely provides practical hardware-oriented parameter tuning guidelines based on a damping ratio and a single cutoff frequency, which makes it useful for anyone building something low-cost.

Dev: It’s about suppressing high-frequency oscillatory behavior while maintaining stable force-feedback teleoperation, which is the hard part of making this work in real time loops.

Taro: That suppression of oscillations is what makes the whole system reliable when the world misbehaves, and that stability under those constraints is what I find most compelling about this research.

The paper's summary: Rosa: So, to recap, this paper shows they built a sensorless four-channel bilateral teleoperation system that works on low-cost manipulators without needing expensive force sensors.

Dev: Right, and what's really interesting is how they manage the whole process—they combine identifying the robot's nonlinear dynamics with a disturbance observer to estimate both velocity and external force at the same time.

Taro: I mean, when you look at this framework, it's designed for these low-cost manipulators that are inherently limited in what sensors they can use.

Rosa: Exactly. They show that by looking at the system structure in the frequency domain, they clarify how the velocity estimation and force estimation bandwidths are linked together.

Dev: And that’s a big deal because it lets them create practical tuning guidelines based on just one cutoff angular frequency and the damping ratio.

Taro: So, if I'm out in the field—driving a rig or something—and I need to tune this system, I don't have to guess a bunch of parameters; I just follow these guidelines?

Rosa: That’s what they did. They used real-robot experiments, including comparing different force sensors and doing teleoperation tasks.

Dev: The results showed that this framework gives you practically useful force estimates and lets you do stable teleoperation even when the robot is moving fast or bumping into things, all while using cheap hardware.

Taro: They found specific settings that work well in different situations, like critical damping for contact-rich tasks to get a fast response without any overshoot in the force estimate.

Rosa: And they also looked at how this system helps with imitation learning. By feeding the estimated force information back into the demonstrations, they significantly improved how successful the AI becomes when it tries to learn those complex manipulation tasks.

Dev: So, it’s not just about getting a good position estimate; it's about having that force data in your demonstration set to make sure the AI learns how to interact with things correctly.

Taro: It suggests that this approach is robust enough to handle modeling errors and external disturbances well, because they used a disturbance observer specifically designed for that.

Rosa: And finally, they did parameter identification by converting the nonlinear model into linear equations and using data sampled at five hundred hertz to find things like gravity parameters and motor inertia parameters.

Dev: It’s a solid way to get those dynamics from low-resolution encoder measurements, even when you aren't running high feedback gains on the robot.

Taro: So, the overall point is that you can achieve stable position-force interaction on low-cost manipulators without expensive sensors by using this specific sensorless control framework.

The paper's improvements: Taro: So, to wrap up, the paper shows how they can make this sensorless teleoperation system better by adding some specific tuning rules and showing how it handles real-world messiness.

Rosa: Yeah, they are suggesting that you really focus on that frequency domain analysis we talked about earlier; it’s not just a nice theoretical concept.

Dev: They say the improvements come from using those guidelines based on the damping ratio and that single cutoff frequency to make the tuning much more practical for real hardware.

Taro: So, if we look at future work, what are they suggesting they should do next to take this beyond just a proof of concept?

Rosa: They mentioned improving data collection and parameter identification using more advanced identification techniques to get better dynamics models.

Dev: And they want to extend the framework to handle a wider range of manipulation tasks, not just the ones they tested, and look at more expressive learning architectures that can really use that force information.

Taro: That makes sense; moving from testing specific pick-and-place tasks to general skill acquisition using force awareness.

Rosa: Exactly. They are also looking into how to get better data collection methods overall, which is always a challenge in robotics when you need high fidelity demonstrations for AI.

Dev: So, the next step seems to be making the whole system more adaptable and robust across different types of physical interactions rather than just perfecting one specific task.

Taro: It sounds like they are trying to build something that can gracefully recover when the real world throws a curveball, which is what I care about most.

Rosa: And that's the big picture—moving from just making it work in the lab to making it reliably useful outside, handling whatever misbehavior happens out there.

Conclusion: Rosa: So, to wrap up, this paper on "Design and Experimental Validation of Sensorless four-Channel Bilateral Teleoperation for Low-Cost Manipulators" shows that high fidelity teleoperation with force feedback is achievable on low-cost hardware without needing expensive force sensors.

Dev: Exactly, and the main contribution is integrating that disturbance observer with nonlinear dynamics compensation to get stable velocity and external force estimates simultaneously.

Taro: It moves us toward a system that can maintain high tracking performance even when there are errors in the robot's actual model or when it bumps into something unexpected.

Rosa: And they showed this works outside the lab, in real-world scenarios, including contact-rich tasks where stability is usually a big problem.

Dev: The safety case here is that by using that disturbance observer and tuning based on the frequency domain analysis—using just one cutoff frequency and the damping ratio—you get reliable force feedback teleoperation.

Taro: I think it’s important because it means we can use these low-cost manipulators for collecting really rich demonstration data for AI learning, since you're not limited by expensive sensor requirements.

Rosa: That’s right; incorporating that estimated force into the AI demonstrations significantly boosts the success rate in those complex manipulation tasks.

Dev: The paper also shows how they can extract useful dynamics from low-resolution encoder measurements using parameter identification based on a five hundred hertz sampling rate.

Taro: It’s solid work because it gives us a practical framework for tackling the hardware limitations that plague real-world robotic deployment.

Rosa: So, to summarize, this work proves that sensorless four-channel bilateral teleoperation is viable on low-cost manipulators when you use a disturbance observer and tune smartly with frequency domain insights.

Dev: It’s about achieving stable force feedback teleoperation without high feedback gains, which is a huge win for real-time control loops.

Taro: We definitely need to keep watching this because the next step involves extending this framework to more general tasks and even more advanced learning architectures that use force awareness better.

KOKI YAMANE, YUNHAN LI, MASASHI KONOSU, KOKI INAMI, JUNJI OAKI, TOSHIAKI TSUJI, SHO SAKAINO

University of Tsukuba · Saitama University

cs.RO, cs.AI, cs.SY, eess.SY

Submitted: 2025-07-08

Updated: 2026-11-18

Comments: 22 pages, 12 figures. Published in IEEE Access

Journal ref: IEEE Access, vol. 14, pp. 109407-109428, 2026

DOI: 10.1109/ACCESS.2026.3713604

Project page: https://kyo-kutsuzawa.github.io/technicalnote/acceleration_based_bilateral_control.html

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Teleoperation of low-cost manipulators is attracting increasing attention as a practical means of collecting demonstration data for imitation learning By interpreting the observer structure in the…

Key concepts

Disturbance Observer (DOB)
A mathematical tool used to estimate unknown external forces or disturbances acting on a system by comparing the actual system behavior with its predicted model. In this paper, it is crucial for simultaneously estimating joint velocity and external torque using only position and input torque measurements, which is necessary for sensorless operation.
Frequency Domain Interpretation
Analyzing the control system's structure in terms of frequencies helps understand how different components—like velocity estimation and force estimation—interact. This analysis revealed a direct coupling between their bandwidths, allowing researchers to set practical tuning parameters based on a single cutoff frequency instead of needing complex, independent settings.
Imitation Learning Data Enhancement
The system collects demonstration data where estimated force information is included in the input and output commands. Incorporating this force awareness into the demonstrations significantly boosts the success rate of subsequent imitation learning policies, especially for delicate contact-rich manipulation tasks like pick-and-place.
Rigid Serial Link Model
This is a simplified mathematical representation of how a multi-joint robot moves, treating the links as rigid bodies connected in a line. This model is used to define the nonlinear dynamics that the control system must compensate for, allowing for accurate torque commands even when high feedback gains are not feasible.

Terminology

Summary

Teleoperation of low-cost manipulators is attracting increasing attention as a practical means of collecting demonstration data for imitation learning By interpreting the observer structure in the frequency domain, we clarify the coupling between the velocity- and external-force-estimation bandwidths and derive practical tuning guidelines based on the damping ratio and a single cutoff frequency Real-robot experiments, including force-sensor comparison and teleoperation tasks, demonstrate that the proposed framework provides practically useful force estimates and enables stable teleoperation in high-speed and contact-rich scenarios under low-cost hardware constraints

Framework Overview

The paper presents a sensorless 4-channel bilateral teleoperation framework that integrates identified nonlinear dynamics compensation with a disturbance-observer-based velocity and external force estimation scheme This system is designed for low-cost, force-sensorless manipulators by requiring simultaneous estimation of joint velocity and external force from low-resolution encoder measurements and input torque The core of the control system involves a rigid serial link model where the dynamics are expressed as M(q)¨q + C(q, q˙)q˙ + Dq˙ + g(q) = τcmd + τext.

Observer Design and Tuning

The velocity and external force estimation is achieved using a disturbance observer (DOB) or extended state observer (ESO) to simultaneously estimate both velocity and external torque using only position measurements and input torque The state-space model includes the acceleration-domain disturbance term d defined as d:= M(q)−1(τext + ∆comp). The observer is designed to be minimal-order, and its parameters are parameterized by the cutoff angular frequency parameter ωc and damping ratio ζ. p. 5 Interpreting the observer structure in the frequency domain clarifies that the design coupling between velocity and external force estimation bandwidths can be understood, allowing for practical parameter constraints based on a single cutoff frequency. p. 5

Control Law Derivation

The control law is formulated by first defining an impedance-shaping objective for 4-channel bilateral teleoperation to shape the leader–follower impedance matrix ZLF. The desired dynamics are expressed in terms of relative-motion and common-motion coordinates, leading to a control law that incorporates proportional–derivative (PD) control for the relative-motion mode and forcedriven motion for the common-motion mode. p. 4 The final torque command is set as τcmd = M˜ (q)n Kp(qd − q) + Kd (q˙d − ˆq˙) + Kf (τd + τˆext) o− τˆext + C˜(q, ˆq˙)ˆq˙ + D˜ˆq˙ + ˜g(q), which incorporates disturbance compensation. p. 4

Experimental Validation and Results

Experiments on a low-cost manipulator demonstrated that the proposed framework provides practically useful force estimates and enables stable teleoperation in high-speed and contact-rich scenarios under low-cost hardware constraints In the free motion comparison, the cutoff frequency determines the trade-off between tracking performance and noise sensitivity, with a middle cutoff frequency (ωc = 50 rad/s) showing that the velocity estimate closely matches the offline-differentiated velocity. In contact-rich tasks, critical damping (ζ = 1. p. 00) is a practical default choice because it provides the fastest response without overshoot in the external force estimate. p. 6 Furthermore, incorporating estimated force information into demonstration data improves task success rates in tested contact-rich manipulation tasks. p. 0

Imitation Learning Applications

The proposed teleoperation system enables the collection of highfidelity demonstration data and that incorporating force information significantly improves the success rate of imitation learning across multiple manipulation tasks For instance, in the dual-arm pick-and-place task, using force information for both input and output resulted in successful picking in all trials. The success rate of imitation learning tasks is shown to be significantly improved when force information is included in the input. p. 1 Future work includes improving data collection and parameter identification by adopting more advanced identification techniques, extending the proposed framework to a wider range of manipulation tasks, and investigating more expressive learning architectures that can better exploit force-aware demonstrations. p. 0

Parameter Identification

The rigid serial link model is converted to linear equations using a variable transformation, where the regressor matrix Y(q, q˙, ¨q) is used to estimate the parameter vector ϕ using linear least squares when external force τext is zero. Data collected by unilateral teleoperation are sampled at 500 Hz and downsampled to 25 Hz for parameter identification using the ‘‘resample’’ function in MATLAB. p. 1 The identified parameters are shown in Table 4, which includes values for gravity parameters, inertia parameters, and motor inertia. p. 8

Cross Structure Gripper

An improved version of the Cross Structure Gripper is used, which allows it to close down to the base of the fingers for powerful grasping of tools This hand has a narrower finger spacing of 5 mm, which allows it to grip even more delicate objects and reduces interference between hands during dual-arm tasks. The implementation details involve using C++ for the low-level control system with each manipulator controlled by a dedicated thread, executed under a high-priority First-In, First-Out (FIFO) real-time scheduling policy to achieve a 1 kHz soft real-time control loop. p. 1

Conclusion

In this study, we demonstrated that practical high-speed bilateral teleoperation with force feedback is achievable on low-cost manipulators without force sensors by employing a sensorless 4-channel bilateral control framework The key contribution of this work lies in integrating nonlinear dynamics compensation with a disturbance-observerbased velocity and external force estimation framework, along with a clear frequency-domain interpretation of the observer structure. This analysis revealed the intrinsic coupling between the velocity and force estimation bandwidths and reduced the observer tuning freedom to a single cutoff frequency, providing practical, hardware-oriented parameter tuning guidelines suitable for low-cost implementations. p. 1 The proposed velocity and disturbance observer suppresses high-frequency oscillatory behavior while maintaining stable force-feedback teleoperation. p. 0

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. Improved AI system can perform stable position-force interaction on low-cost manipulators without force sensors by employing a sensorless 4-channel bilateral control framework. This allows the system to achieve stable and accurate position-force interaction under severe sensing and bandwidth constraints, which are common limitations of low-cost robotic hardware.

  2. Improved AI system can generate high fidelity demonstration data suitable for imitation learning by incorporating estimated force information into demonstrations, which improves task success rates in the tested contact-rich manipulation tasks.

  3. Improved AI system can perform accurate control even when a certain amount of model error exists by combining a parameter-identified, accurate nonlinear manipulator dynamics model with a disturbance observer to achieve high tracking performance and operability with a low-cost manipulator that cannot use high feedback gains.

  4. Improved AI system can optimize tuning parameters for the velocity and external force estimators by interpreting the observer structure in the frequency domain to clarify the coupling between the velocity- and external-force-estimation bandwidths and deriving practical tuning guidelines based on a damping ratio and a single cutoff frequency.

  5. Improved AI system can suppress high-frequency oscillatory behavior during teleoperation by using a disturbance observer that addresses modeling errors, as it suppress[es] this high-frequency oscillatory behavior while maintaining stable force-feedback teleoperation.

  6. Improved AI system can enhance imitation learning performance across multiple manipulation tasks by utilizing force information in the policy output, as shown in Task 1 where When both the input and the output used force information, both pick-andplace operations were successful in all trials.

Sources

Related papers