A new control architecture for human-scale bilateral teleoperation eliminates force/torque sensors by using inverse dynamics modeling to estimate external forces in real time. This approach reduces system complexity and cost while achieving higher transparency—the operator feels like they are directly interacting with the remote environment.
What the Researchers Built
Researchers from the University of Alberta developed a sensorless four-channel teleoperation architecture (4c-DC) implemented on a custom WAM (Whole Arm Manipulator) bilateral system. The leader is a 4-DOF WAM arm with a 3-DOF haptic wrist that mirrors the follower’s kinematics, while the follower is a 7-DOF WAM arm. Instead of using expensive and fragile force/torque sensors, the system relies on model-based inverse dynamics to estimate external joint torques from the robot’s own motion data.
Key components include feedforward dynamic compensation on both leader and follower, real-time torque estimation, and a four-channel control loop that transmits position, velocity, force, and torque information bidirectionally. The system runs at 500 Hz over UDP, with all control software written in C++ using the open-source libbarrett library.
The architecture enables the operator to feel contact forces not only at the end-effector but along the entire arm body—a significant improvement over traditional end-effector-only force feedback. This is achieved without any additional sensors, making the system cheaper and mechanically simpler than sensor-equipped alternatives.
Key Results
The researchers evaluated their proposed 4c-DC architecture against three baselines: two-channel P-P control (2c), standard four-channel control (4c), and a transparency enhancement method (TE). All tests were conducted in joint space without a human operator to ensure objective reproducibility.
In a door-opening case study, 4c-DC consistently outperformed all other methods in both free motion and hard contact:
- Position tracking during contact improved significantly, reflecting higher maximum transmittable impedance. The operator could better perceive the door’s physical constraints.
- Force tracking in free motion was superior, indicating lower leader impedance and easier operator movement—meaning the operator felt less resistance when moving freely.
- External torque estimation error (quantified by NRMSE) was low for the most sensitive joints (Joints 2 and 4 of the WAM), validating that the inverse dynamics model accurately captures interaction forces without physical sensors.
The system also demonstrated the ability to detect contact along the manipulator body, not just at the tool tip. This is a key advantage over conventional teleoperation schemes that only measure end-effector forces.
How It Works
The system is built on a four-channel teleoperation architecture, which transmits four signals between leader and follower: position, velocity, force, and torque. Unlike simpler two-channel architectures that only exchange position or force, four-channel control enables full transparency—the operator feels exactly what the remote robot feels, both in free motion and during contact.
The central innovation is inverse dynamics modeling to estimate external torques. Instead of using force/torque sensors, the system computes the expected torques from the robot’s dynamic model (inertia, Coriolis, gravity), then subtracts those from the measured motor torques. The remainder is the estimated external torque applied by the environment or operator.
To obtain the dynamic model, the researchers performed parameter estimation using optimized excitation trajectories. For the 4-DOF leader, 26 base dynamic parameters were identified. The wrists were treated as rigid links due to their small dynamic contribution, reducing model complexity. The estimation was done offline using open-source tools, but the resulting model runs in real time with negligible latency—critical for stable haptic feedback.
The control law combines feedforward compensation on both sides with the estimated torques. On the leader side, the estimated operator torque is used to generate haptic feedback that reflects the follower’s interaction forces. On the follower side, the estimated follower joint torques are used for force reflection. The whole loop runs at 500 Hz, ensuring stability even during hard contact tasks like door opening.
The system also handles contact detection along the entire manipulator body because inverse dynamics estimates torques at every joint, not just at the end-effector. This is a unique advantage over sensor-based systems that can only measure forces at the mounting point.
Why This Matters for Robotics
Bilateral teleoperation is critical in hazardous environments—nuclear decommissioning, underwater inspection, bomb disposal, and telesurgery. Force feedback heavily impacts operator performance: poor transparency leads to fatigue, low precision, and even task failure.
Eliminating force/torque sensors reduces hardware cost and maintenance while increasing ruggedness. Sensors are notoriously delicate and add inertia to the robot. A sensorless approach makes teleoperation systems more affordable and deployable in dirty or high-radiation areas.
For used industrial robots and collaborative cobots, this architecture could be retrofitted to enable high-quality teleoperation without major hardware changes. Many existing robots already measure motor currents—combining those with accurate dynamic models could produce sensorless force feedback.
The ability to detect contact along the full arm length opens new applications for whole-arm manipulation (e.g., pushing objects with the forearm) and safer human-robot interaction, where any collision can be felt by the operator and reacted to immediately.
Limitations and Open Questions
The inverse dynamics model does not account for stiction or Coulomb friction, which become significant at low velocities and during fine manipulation. This degrades transparency when the operator moves slowly or performs delicate tasks.
The current implementation assumes both leader and follower have similar kinematics (both WAM arms). Extending to heterogeneous systems—different robot brands or configurations—would require independent dynamic identification for each side.
Communication delay was negligible in the experimental setup (both processes on the same computer). Real deployments over networks will introduce latency, and the architecture’s stability under delay remains untested. Finally, online adaptation (e.g., using disturbance observers) could compensate for dynamic model inaccuracies and varying loads—a promising but unexplored direction.
Frequently Asked Questions
What is a four-channel teleoperation architecture? It transmits four signals (position, velocity, and force/torque from both sides) bilaterally, enabling the operator to feel exactly what the remote robot feels—both in free motion and during contact.
How does sensorless force estimation work without force sensors? The system uses inverse dynamics: it computes the expected joint torques from a dynamic model of the robot, then subtracts those from measured motor torques. The remainder is the estimated external torque from contact or operator input.
Does this approach work on any robot, or only the WAM arm? It works on any robot with accurate joint position and torque sensing and a known dynamic model. The WAM’s cable-driven, back-drivable design makes it particularly suitable, but the method is generalizable.
How does whole-arm contact detection compare to end-effector only? By estimating torques at every joint, the system detects forces applied anywhere on the arm—not just at the tool tip. This provides richer haptic feedback for tasks like pushing with the forearm or unintended collisions.
Conclusion
The sensorless four-channel control architecture proves that high-quality bilateral teleoperation is possible without expensive force/torque sensors, using model-based inverse dynamics instead. Experimental results on a WAM system show superior transparency in both free motion and hard contact, opening the door to cheaper, more robust teleoperation for demanding applications.