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

This paper presents a sensorless 4-channel bilateral teleoperation framework that enables stable, high-speed force feedback control on low-cost manipulators through disturbance-observer-based estimation and simplified tuning, ultimately demonstrating that such force-enhanced data significantly improves imitation learning performance.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Picture: Teaching Robots to "Feel" Without Sensors

Imagine you are trying to teach a robot how to peel a cucumber or turn a screw. Usually, you'd need a robot with expensive, high-tech "fingers" that have built-in pressure sensors to feel how hard it's pushing. But those sensors are costly and fragile.

This paper asks a bold question: Can we teach a cheap robot to "feel" and handle delicate tasks without those expensive sensors?

The answer is yes. The authors built a system that lets a low-cost robot arm (called CRANE-X7) perform high-speed, delicate tasks and even teach itself new skills, all without a single force sensor.

---

The Problem: The "Blind" Robot

Most cheap robots used for learning (like the ones in the "ALOHA" series) work like a blindfolded person holding a stick.

• How they work: You move your hand (the "Leader"), and the robot (the "Follower") tries to copy your position.
• The flaw: If the robot hits a wall or a soft object, it doesn't know it's hitting anything until it crashes or gets stuck. It has no sense of touch. This makes it terrible at tasks requiring speed or contact, like peeling fruit or tightening a nut.

The Solution: The "Telepathic" Twin

The authors created a 4-Channel Bilateral Teleoperation system. Think of this as giving the robot a "telepathic" connection to the human operator.

Instead of just copying positions, the system does two things simultaneously:

1. Position Control: "Move your hand to where I am."
2. Force Control: "Push with the same amount of strength I am feeling."

But here's the magic trick: The robot doesn't have sensors to feel the force. So, how does it know?

The Secret Sauce: The "Mathematical Crystal Ball"

Since the robot can't feel the force, the authors built a mathematical crystal ball (an observer) inside the computer.

1. The Physics Model: They taught the computer the exact physics of the robot (how heavy its arms are, how friction works, how gravity pulls). This is like knowing exactly how heavy a backpack is before you put it on.
2. The Disturbance Detective: When you move the robot, the computer calculates exactly how much force should be needed.
   • If the robot moves exactly as predicted, great!
   • If the robot moves slower than predicted, the computer realizes, "Hey, something is pushing back!" (Maybe it hit a wall).
   • If the robot moves faster, it realizes, "Something is pulling it!"

By constantly comparing the expected movement (based on math) with the actual movement (measured by cheap sensors), the computer can estimate the force being applied. It's like a driver feeling the wind resistance on a car without a wind gauge; they just know the engine is working harder than usual, so there must be a headwind.

The Innovation: Tuning the "Crystal Ball"

The paper's biggest technical breakthrough is making this "crystal ball" easy to tune.

• The Old Way: Tuning these systems was like trying to balance a stack of Jenga blocks while riding a unicycle. You had to adjust many different knobs (gains) for speed and force, and if you messed up one, the whole thing would shake or crash.
• The New Way: The authors realized that the speed and force estimation are actually linked. They showed that you only need to tune one single knob (a frequency cutoff) to get everything right. It's like realizing that instead of tuning the bass, treble, and volume separately, you just need to turn one "Master Tone" knob to get perfect sound.

The Results: From "Clumsy" to "Master Chef"

They tested this on a low-cost robot with three tasks:

1. Pick and Place: Picking up blocks of different sizes.
2. Nut Turning: Rubbing a nut onto a screw quickly.
3. Cucumber Peeling: Peeling a cucumber without squishing it.

The Findings:

• Without Force Info: The robot was clumsy. It dropped small blocks, couldn't turn the nut, and squashed the cucumber.
• With the "Mathematical Crystal Ball": The robot became stable and precise. It could feel the resistance of the cucumber skin and adjust its grip instantly.
• The Imitation Learning Bonus: They used this system to record "demonstrations" (human experts doing the tasks) and taught a robot AI to do it alone. The AI that learned from the "force-feeling" demonstrations succeeded 100% of the time, while the AI that learned from the "blind" demonstrations failed miserably.

The Takeaway

This paper proves that you don't need expensive, fragile sensors to make robots feel. By using smart math and a "disturbance observer" (a digital detective), we can turn cheap, low-cost robots into skilled, high-speed workers that can handle delicate, contact-heavy tasks.

In short: They taught a cheap robot to "feel" its way through the world using nothing but a calculator and some clever physics, making it a perfect teacher for future AI robots.

Technical Summary

Here is a detailed technical summary of the paper "Design and Experimental Validation of Sensorless 4-Channel Bilateral Teleoperation for Low-Cost Manipulators."

1. Problem Statement

The paper addresses the challenge of enabling high-speed, contact-rich bilateral teleoperation using low-cost manipulators that lack force sensors and high-resolution encoders.

• Context: Imitation Learning (IL) and "Vision-Language-Action" (VLA) models require massive datasets of human demonstrations. Low-cost hardware (e.g., CRANE-X7) is increasingly used to collect this data due to cost and safety benefits.
• Limitations of Current Systems:
  • Most low-cost systems use unilateral control (position-only), which lacks force feedback, making contact-rich tasks (e.g., wiping, peeling) difficult or impossible.
  • Existing 4-channel bilateral control (simultaneous position and force control) typically relies on simplified linear dynamics models and phase-lagged velocity estimation (using low-pass filtered derivatives).
  • On low-cost hardware with slow control cycles and low-resolution encoders, these simplifications lead to significant performance degradation, instability, and poor tracking during high-speed or contact-heavy operations.
• Core Challenge: How to achieve stable, transparent, high-performance bilateral control on hardware with limited sensing capabilities (no torque sensors, 12-bit encoders) and computational constraints, without relying on expensive high-fidelity sensors.

2. Methodology

The authors propose a sensorless 4-channel bilateral control framework that integrates nonlinear dynamics compensation with a Disturbance Observer (DOB) for simultaneous velocity and external force estimation.

A. Control Architecture

• 4-Channel Bilateral Control: Both master (leader) and slave (follower) robots exchange position and force information. The control law is derived using a coordinate transformation that separates the system dynamics into:
  • Difference coordinates ($q_-$): Controlled for position synchronization (stiffness).
  • Average coordinates ($q_+$): Controlled for force reflection (transparency).
• Nonlinear Dynamics Compensation: Unlike previous methods using simplified linear models, this system utilizes a parameter-identified nonlinear rigid-body dynamics model ($M(q)\ddot{q} + C(q,\dot{q})\dot{q} + D\dot{q} + g(q)$). The identified parameters (inertia, Coriolis, gravity, friction) are used to linearize the system via feedback linearization, reducing the burden on the observer.

B. Velocity and External Force Estimation (The Core Innovation)

Since the manipulator lacks torque sensors and has low-resolution encoders, the system estimates joint velocity ($\dot{q}$) and external torque ($\tau_{ext}$) using a minimal-order observer based on the Disturbance Observer (DOB) principle.

• State-Space Formulation: The observer treats the external force (normalized by inertia) as a state variable.
• Frequency Domain Interpretation:
  • The authors analyze the observer in the frequency domain, revealing an intrinsic coupling between the velocity estimation bandwidth and the external force estimation bandwidth.
  • Complementary Filtering: The velocity estimate acts as a complementary filter, combining a prediction based on the input torque (using the identified inertia model) and a feedback term based on the derivative of the position measurement.
  • Lead Compensation: By using the identified inertia model to predict velocity, the system avoids the phase lag inherent in simple low-pass filtered derivatives. This allows for higher control gains (specifically the D-gain in PD control), improving stability and responsiveness.
• Tuning Guidelines: The analysis demonstrates that the observer's tuning freedom can be reduced to a single cutoff frequency ($\omega_c$). The damping ratio is fixed to critical damping ($\zeta=1$) for the fastest response without overshoot. This simplifies hardware-oriented parameter tuning.

C. Cascade Control Structure

The controller is interpreted as a cascade of acceleration, velocity, and position loops:

1. Acceleration Layer: Uses velocity feedback (derived from the observer) rather than direct acceleration feedback, allowing implementation with low-resolution encoders. It inherently includes integral action to eliminate steady-state errors.
2. Velocity/Position Layer: The estimated velocity provides lead compensation, increasing phase margin and suppressing resonant peaks caused by inertia mismatches.

3. Key Contributions

1. Practical Sensorless 4-Channel System: A novel control framework enabling stable, high-speed bilateral teleoperation on low-cost manipulators without force sensors, overcoming the limitations of phase-lagged velocity estimation.
2. Frequency-Domain Observer Analysis: A theoretical clarification of the coupling between velocity and force estimation bandwidths. This leads to a systematic tuning guideline where two observer gains are reduced to a single cutoff frequency parameter.
3. Integration of Nonlinear Dynamics: The successful application of a parameter-identified nonlinear dynamics model to a low-cost system, significantly improving tracking performance compared to linear approximations.
4. Imitation Learning Validation: Demonstration that incorporating estimated force information into demonstration data significantly improves the success rate of imitation learning policies for complex tasks.

4. Experimental Results

Experiments were conducted on a CRANE-X7 (7-DOF, low-cost) manipulator with 12-bit encoders.

A. Teleoperation Performance

The proposed method was compared against:

• Unilateral Control
• Symmetric Position Bilateral Control
• Force Feedback Bilateral Control
• 4-Channel with Fixed Inertia
• 4-Channel with Phase-Lagged Velocity (standard DOB)

Findings:

• Free Motion: The proposed method achieved the lowest Mean Absolute Error (MAE) in position ($0.61^\circ$) and angular velocity ($9.2^\circ/s$), outperforming the "Phase-Lagged Velocity" method which suffered from high-frequency vibrations due to phase lag.
• Contact-Rich Tasks (Wiping): The proposed method maintained stability and accuracy ($0.28^\circ$ position error) during wiping tasks, whereas methods with phase-lagged velocity exhibited oscillations and instability.
• Robustness: The system successfully handled high-speed motions and contact interactions where other methods failed or required significantly lower gains (reducing stiffness).

B. Imitation Learning (IL) Application

The system was used to collect demonstration data for training Action Chunking with Transformer (ACT) policies on three tasks:

1. Dual-Arm Pick-and-Place:
   • Without Force: Failed to pick small objects (10-20mm) and dropped objects frequently.
   • With Force Input: 100% pick success.
   • With Force Input & Output: 100% success for both pick and place, with stable holding.
2. Nut Turning: Required significant force and speed.
   • Without Force: 0% success.
   • With Force Input: 100% success.
3. Cucumber Peeling:
   • Without Force: 0% success (gripper missed the peeler/cucumber).
   • With Force: Improved success rates (3/5 and 2/5), demonstrating that force information is critical for contact-rich manipulation.

Conclusion: Force information in the demonstration data is crucial for the robot to learn the necessary contact forces and timing, significantly boosting task success rates.

5. Significance

• Democratizing Robot Learning: This work proves that high-fidelity teleoperation data for training advanced AI models (VLA/LBM) can be collected using low-cost, sensorless hardware, removing the barrier of expensive force-torque sensors.
• Theoretical Advancement: The frequency-domain analysis of the DOB provides a rigorous mathematical foundation for tuning observers in low-cost systems, moving beyond heuristic tuning.
• Practical Impact: The proposed method enables stable, high-speed contact tasks on affordable robots, making it a viable solution for scaling up data collection for the next generation of general-purpose robots.

In summary, the paper bridges the gap between the theoretical ideal of 4-channel bilateral control and the practical constraints of low-cost hardware, providing a robust, tunable, and high-performance solution that directly enhances the capabilities of imitation learning.