Bilateral Teleoperation · Force Feedback

Teleoperating Leader Arm
for Dual Arm

Seunghwan Um, Chemin Ahn, Jeong Hwan Park, Sung Jun Lee, Hyouk Ryeol Choi (Advisor)

Command remote motion. Feel remote contact.

Bilateral Teleoperation with Robotic Hands

Coordinated arm and hand motion for interacting with objects.

The leader interface is used to command the dual-arm robot and its hands during object interaction.

Increased force-feedback gain. The system was operated with a higher force-feedback gain in this video to make the reflected force response more visible.

KRoC 2026Special-session presentation

Development of a Dual-Arm Robot Teleoperation System with Leader-Arm Force Feedback

Presented by Seunghwan Um at the 21st Korea Robotics Society Annual Conference, [TA6] Special Session 5 — Humanoid Technology for Logistics Center Tasks.

KRoC 2026 official program ↗

Overview

A backdrivable leader arm provides an intuitive interface for controlling a remote follower. Leader joint motion sets the follower’s target pose, while contact forces at the follower are reflected as joint torques on the leader. This closes the loop between motion commands and force feedback.

The leader remains in current (torque) control with gravity compensation. Jacobian-transpose mapping renders the remote wrench, with a Cartesian virtual impedance formulation that can add damping and optional inertia. The demonstrations below show dual-arm motion, external interaction, and teleoperation with robotic hands.

Motion commands →
01Operator / 작업자OperatorHand motion
02Backdrivable dual leader arms / 양팔 리더암Leader armBackdrivable · gravity compensated
03Dual-arm follower robot / 양팔 팔로워 로봇Follower robotRemote motion · contact
← Force feedback
A bilateral interface connects the operator’s motion to the robot’s interaction with the environment.
Dual-Arm Force Feedback

Leader–follower motion and external interaction in a dual-arm setup.

The operator guides the follower arms through the leader. External interaction with the follower is demonstrated during teleoperation.
Leader–Follower Motion in Simulation

A development demonstration of physical leader motion and the corresponding virtual dual-arm robot.

Leader–follower motion visualized in simulation.
Control Principle & Force Reflection

Motion travels from the leader to the follower. The remote wrench returns through the leader’s Jacobian as a torque command. The following equations describe the force-feedback formulation used in the C++ implementation.

Notation: L = leader, F = follower. Spatial vectors use angular/moment components first: \(\boldsymbol{w}=[\boldsymbol{m}^{T},\boldsymbol{f}^{T}]^{T}\). Wrench and Jacobian must use matching axes, reference origins, and row order.

01Leader motion → follower pose

\[\boldsymbol{T}_{F}^{\mathrm{des}}=\operatorname{FK}_{F}(\boldsymbol{q}_{F}^{\mathrm{des}})\]

Forward kinematics converts the follower target joint angles corresponding to leader motion into a target pose.

02Wrench frame transformation

\[\begin{aligned}\boldsymbol{f}_{\mathrm{base}}&=\boldsymbol{R}_{\mathrm{base,sensor}}\boldsymbol{f}_{\mathrm{sensor}},\\ \boldsymbol{m}_{\mathrm{base}}&=\boldsymbol{R}_{\mathrm{base,sensor}}\boldsymbol{m}_{\mathrm{sensor}}+\boldsymbol{p}_{\mathrm{base},A}\times\boldsymbol{f}_{\mathrm{base}}.\end{aligned}\]

sensor denotes the input sensor frame, base the robot base frame, and A the force application point. The rotation matrix expresses force and moment in the base frame; the cross-product term transfers the moment to the base origin.

03Jacobian-transpose force reflection

\[\boldsymbol{\tau}_{\mathrm{JT}}=\boldsymbol{G}_{\tau}\,\boldsymbol{J}_{L}^{T}(\boldsymbol{q}_{L})\,\boldsymbol{w}_{\mathrm{ref}}\]

The leader’s spatial Jacobian converts remote wrench into joint torques.

04Cartesian virtual impedance

\[\begin{aligned}\boldsymbol{v}_{L}&=\boldsymbol{J}_{L}\dot{\boldsymbol{q}}_{L},\qquad \boldsymbol{a}_{L}=\dot{\boldsymbol{v}}_{L},\\ \boldsymbol{W}_{L}&=\boldsymbol{G}_{w}\boldsymbol{w}_{\mathrm{ref}}-\boldsymbol{D}_{c}\boldsymbol{v}_{L}-\boldsymbol{M}_{c}\boldsymbol{a}_{L},\\ \boldsymbol{\tau}_{\mathrm{cart}}&=\boldsymbol{J}_{L}^{T}\boldsymbol{W}_{L}.\end{aligned}\]

An alternative to per-joint reflection: scale the wrench in Cartesian space, then add velocity damping and optional virtual inertia. \(\boldsymbol{v}_{L}\) is the spatial velocity associated with the same Jacobian. This formulation renders torque while preserving the leader’s backdrivability.

05Gravity compensation & motor current

\[\boldsymbol{\tau}_{\mathrm{cmd}}=\boldsymbol{\tau}_{g}+\operatorname{clip}(\widetilde{\boldsymbol{\tau}}_{\mathrm{fb}})+\boldsymbol{\tau}_{\mathrm{fric}}\]
\[u_{j}=\operatorname{sat}\!\left(\operatorname{round}\!\left(\frac{\tau_{\mathrm{cmd},j}}{K_{t,j}\,c_{I,j}}\right)\right)\]

Filtered feedback, including optional joint damping, is limited per joint and combined with gravity compensation and optional friction compensation. Torque constants \(K_{t,j}\) and current-unit scales \(c_{I,j}\) convert torque to bounded Dynamixel goal-current units \(u_j\).

Optional friction feedforward
\[\tau_{\mathrm{fric},j}=\operatorname{clip}\!\left(s_j\left[F_{c,j}\tanh\!\left(\frac{\dot q_j}{v_{\epsilon,j}}\right)+F_{v,j}\dot q_j\right]\right)\]

A smoothed Coulomb-plus-viscous model assists motion to compensate for leader joint friction. Its parameters need to be identified for the hardware; the term is optional.