Force-Feedback Telemanipulation

Updated 22 June 2026

Force-feedback telemanipulation is a bilateral architecture that provides real-time haptic feedback through physical actuation for enhanced situational awareness and precision.
It employs varied control strategies such as position–force hybrid control and Cartesian impedance coupling to maintain stability and ensure precise sensorimotor integration.
Advanced sensing modalities including six-axis sensors, vision-based tactile methods, and neural network estimators enable reliable force rendering and improved task outcomes.

Force-feedback telemanipulation refers to remote manipulation systems that render real-time haptic information—typically contact forces and sometimes torques—back to a human operator by physically actuating a master manipulator, haptic device, or wearable interface. By closing the sensorimotor loop between human and remote robot, force-feedback telemanipulation provides situational awareness and task precision that otherwise are not achievable via visual-only or unilateral teleoperation. This capability underpins safe and effective handling of contact-rich, dynamic, or deformable objects in domains ranging from surgical robotics, hazardous-material handling, and space operations to the collection of high-quality demonstration data for imitation learning.

1. Fundamental Architectures and Control Strategies

Force-feedback telemanipulation systems universally adopt a bilateral architecture, mapping operator inputs to remote (“follower”) robots, and mapping environmental reaction forces back to the operator. Canonical realizations include:

Position–force hybrid control: The operator commands positions or velocities, while sensed or estimated forces at the slave are rendered to the leader as force cues. For example, the Omega.7 master and DIGIT-instrumented parallel gripper system implements $x_{fd} = x_l$ and $f_{ld} = f_s$ , with the slave providing only force feedback (no reflected inertia or gravity), ensuring light, stable free-space operation (Zhu et al., 2022).
Cartesian impedance coupling: Both master and slave execute impedance control in task space; the robot follows the operator, and the operator is “pulled” or “pushed” by the remote environment according to measured (or estimated) end-effector wrenches. This approach, prevalent in avatar-style bimanual teleoperation (e.g., Panda–SenseGlove–Schunk systems (Lenz et al., 2021, Schwarz et al., 2021, Lenz et al., 2023)), enables natural, transparent interaction, provided damping and null-space components are properly tuned to manage time delay and limit avoidance.
Joint-space bilateral PD or impedance: Reduction to joint variables is standard in systems with high-fidelity, symmetric hardware or when exact spatial correspondences are not required. For instance, symmetric bilateral PD loops are used in low-inertia, proprioceptive telemanipulators for sub-newton force tracking (SaLoutos et al., 2022).

Bilateral stability is ensured via passivity—requiring that environmental, master, and slave subsystems (including transfer delays and haptic rendering) do not produce net energy. Formulations range from direct passivity observation, time-domain oscillation detection and damping (Lenz et al., 2023), to carefully chosen impedance parameters and saturations (Kanai et al., 10 Sep 2025).

2. Force and Tactile Sensing Modalities

Multiple force, torque, and tactile estimation modalities have been integrated into telemanipulation pipelines:

Six-axis force/torque sensors: Commonly used at robot wrists for high-bandwidth, direct force feedback (Lenz et al., 2021, Schwarz et al., 2021, Lenz et al., 2023).
Joint torque estimation: Utilized when external F/T sensors are absent. Joint-torque residuals ( $\tau_{\mathrm{ext}}$ ) are computed via model-based subtraction of expected dynamics/gravitation from measured actuator torques, as in GELLO – Panda bilateral schemes (Sujit et al., 18 Jul 2025).
Vision-based tactile sensors: Low-cost, high-resolution contact sensing (DIGIT, GelSight Mini) is achieved by photometric measurement of elastomer-gel deformations. Depth mapping (via MLP regression to surface normals and Poisson solvers), followed by polynomial or neural regression, yields accurate normal force estimation (Zhu et al., 2022, Becker et al., 2024, You et al., 5 Mar 2026). Shear and slip inference require additional features or deep models (Zhu et al., 2022).
Neural-network force estimators: Imaging or state-based networks, including ResNet- and fully-connected models, can regress contact forces from either vision ( $V$ ), state ( $S$ ), or multimodal ( $VS$ ) input (Chua et al., 2021). Notably, vision-only networks yield consistent passivity but underestimate stiffness; state-based networks achieve better transparency but may destabilize human-in-the-loop feedback, especially laterally.
Sensorless virtual force estimation: In ACE-F, follower end-effector deviations under task-space PD control are interpreted as virtual spring–damper forces, eliminating the need for physical F/T sensors and enabling cross-platform feedback (Yan et al., 25 Nov 2025). Similar approaches are also realized in input-gated bilateral teleoperation (IGBT), where follower actuation current is repurposed as a contact force estimate (Kanai et al., 10 Sep 2025).

3. Haptic Rendering Devices and Wearable Interfaces

The actuation medium for force feedback is critical to system fidelity, safety, and task suitability:

Desktop haptic devices: Omega.7 and Virtuose 6D deliver 3–6 DoF force/torque rendering at the hand or wrist. Their high backdrivability is essential for stable bilateral teleoperation (Zhu et al., 2022, Adjigble et al., 2023).
Exoskeleton arms: Torque-controlled industrial robot arms (e.g., Panda) are paired with operator-side force sensors and joint-torque actuation, enabling full-arm haptic feedback in avatar-style setups (Lenz et al., 2021, Schwarz et al., 2021, Lenz et al., 2023).
Haptic gloves: Cable-driven (CDF-Glove) and exoskeleton-based (MFE, SenseGlove) gloves provide multi-DoF finger–level kinesthetic, vibrotactile, and (in MFE) thermohaptic feedback (Liang et al., 6 Mar 2026, Tang et al., 3 Apr 2026, Lenz et al., 2021). Cable mechanisms allow independent control of flexion/abduction with sub-degree repeatability and closed-loop force response. Advanced exoskeletons achieve over 3.5–8.1 N fingertip forces (MFE), and integration of localized pressure and temperature cues for material property discrimination.
Continuum soft actuators: High-power-density soft pneumatic devices enable compact, 3-DoF kinesthetic feedback at the fingertip, achieving up to 6 N force in a 13 g form factor with 3–4 Hz bandwidth (Su et al., 2024).
Wearable vests and torso displays: For users with limited fine-tactile perception (e.g., in assistive robotics), force/tactile metrics (e.g., Contact Concentration Index, Effective Deformation Area) are rendered vibrotactile to the torso, closing the loop between intent (EMG) and feedback (You et al., 5 Mar 2026).
Pedal haptics: TriPilot-FF introduces bipedal pedal devices, using foot-position for mobile base control and rendering resistive feedback derived from proximity sensors (lidar), thereby distributing control to upper and lower limbs (Li et al., 10 Feb 2026).

4. Sensing-to-Feedback Pipelines: Estimation, Control, and Rendering

The signal processing and control pipeline in force-feedback telemanipulation comprises:

Sensing/deformation measurement: Raw data (RGB images, joint-torques, pressure) is processed into local force estimates.
- Vision-based tactile sensors: MLPs or deep CNNs map per-pixel color deformation to surface normals and ultimately to depth ( $D(x,y)$ ). Max indentation yields a scalar deformation, which is mapped to force via polynomial regression (e.g., $F_s = p(d_{current})$ with $R^2=0.9987$ fit) (Zhu et al., 2022).
- Optical flow and marker tracking are applied to dense arrays (GelSight Mini) for normal and shear force estimation (RMSE: 0.24 N normal, 0.30 N shear) (Becker et al., 2024).
Machine learning force predictors: Neural networks ingest vision and/or robot state to produce 3D force vectors; transparency versus stability is carefully characterized via rendered impedance and passivity analysis (Chua et al., 2021).
Feedback mapping and scaling: Forces are rendered as
- Kinesthetic forces (continuous or discretized, e.g. via cable tension, resisted by springs/servos or actuators)
- Vibrotactile feedback (amplitude/frequency mapped to force magnitude, e.g., $A = \frac{\log(1+\alpha F_{th})}{\log(1+\alpha F_{max})}$ ) (Becker et al., 2024)
- Thermohaptic signals (via Peltier elements for temperature (Tang et al., 3 Apr 2026))
Closed-loop control: Teleoperation loops run at 30–1 000 Hz (hardware dependent), synchronizing sensor streams with low-latency communication (RS485/Modbus, Ethernet, or ROS shared memory), and guarantee passivity through damping, explicit monitoring, or filter–saturation structures.

5. Performance Evaluation, Benchmarks, and Task Outcomes

Performance metrics span force-tracking accuracy, task completion time, success rates, error reduction, and subjective usability:

Force-tracking: Systems using physical wrist F/T sensors or high-fidelity joint torque measurement attain RMSE below 0.5 N at >500 Hz (SaLoutos et al., 2022, Lenz et al., 2021). Vision-based tactile estimation reaches $f_{ld} = f_s$ 00.03–0.1 N RMSE with appropriate calibration (Zhu et al., 2022, Becker et al., 2024).
Task performance: Empirical studies consistently show that force feedback increases grasp success and reduces error. For example:
- In-hand pivoting with force feedback: 86.7% success vs. 40% visual-only, and reduced completion time (Zhu et al., 2022).
- Blind grasping tasks with CDF-Glove: 5× improvement in success rate (1/10 to 5/10) (Liang et al., 6 Mar 2026).
- Grasping soft objects with GelSight–MANUS integration: 48% reduction in deformation (Becker et al., 2024).
- Sim2real data collection pipelines with force-feedback: >5–15% improvement in simulated task success, and up to 4 s reduction in execution time (Zou et al., 3 Mar 2025).
- Imitation learning tasks with force-augmented demonstrations (GELLO, ACE-F, TriPilot-FF): up to 95% demonstration success, 47% reduction in completion time, and significant transfer to real-robot deployment (Sujit et al., 18 Jul 2025, Yan et al., 25 Nov 2025, Li et al., 10 Feb 2026).

Subjective metrics (NASA-TLX, Likert scale) consistently indicate improved confidence, object presence, and handling intuitiveness with force feedback. Notably, novice operators benefit most from haptic cues for contact and slip detection.

6. Limitations, Robustness, and Research Challenges

Despite progress, force-feedback telemanipulation systems exhibit characteristic constraints:

Bandwidth limits: Low-cost vision-based tactile sensors are restricted to 30–100 Hz by camera and control-loop rates, making dynamic (>10 Hz) tasks sluggish (Zhu et al., 2022, Liang et al., 6 Mar 2026). State-of-the-art electromechanical arms and exoskeletons can attain >1 kHz, essential for high-frequency contacts.
Sensing coverage: Many approaches estimate only normal forces; shear and slip detection require more sophisticated features or inference (Zhu et al., 2022, Becker et al., 2024).
Contact generalization: Sensor calibration and deformation–force mappings are often contact-geometry or material dependent, introducing systematic errors when object surface or contact point varies (Zhu et al., 2022).
Passivity and stability: Complex or high-bandwidth bilateral couplings risk limit-cycle oscillation or non-passive operation, especially under network delays or when learning-based estimation is used. Explicit passivity layers or damping, as well as oscillation observers, are required for robust operation (Lenz et al., 2023, Chua et al., 2021).
Sensor calibration and drift: Vision-based tactile estimators may require batch calibration for each gel or sensor batch due to manufacturing variances (Becker et al., 2024).
Cross-embodiment/portability: Achieving general, user-transparent force feedback across diverse robot platforms often necessitates sensorless virtual force estimation (e.g., ACE-F), robust cross-kinematic mapping, and platform-specific gain control (Yan et al., 25 Nov 2025).

A plausible implication is that robust, transparent force-feedback in broad, real-world telemanipulation will require architectural advances in sensor fusion, closed-loop observer design, and higher-rate, multi-modal actuation/sensing.

7. Emerging Directions and Application Domains

Ongoing and future research in force-feedback telemanipulation is evident across multiple axes:

Multi-modal haptic rendering: Integration of force, pressure, skin deformation, and temperature in lightweight, dexterous gloves/exos and vests shows substantial gains in object recognition and manipulation with minimal cognitive overhead (Tang et al., 3 Apr 2026, You et al., 5 Mar 2026).
Wearable, low-cost haptic interfaces: Open-source, cable-driven gloves (CDF-Glove), foldable 3-DoF arms (ACE-F), or even finger-mounted soft continuum displays democratize high-fidelity force feedback and demonstration data collection at low price points ( $f_{ld} = f_s$ 1250$) (Liang et al., 6 Mar 2026, Yan et al., 25 Nov 2025, Su et al., 2024).
Sim-to-real learning and data collection: High quality, force-augmented teleoperation datasets transfer more robustly to real systems, reducing reliance on laborious physical demonstration, especially when paired with vision models of sufficient fidelity and chunked-action imitation architectures (Zou et al., 3 Mar 2025, Li et al., 10 Feb 2026, Sujit et al., 18 Jul 2025).
Whole-body and mobile manipulation: TriPilot-FF's integration of foot-operated haptics and bimanual force reflection enables collision-safe, dexterous, and energy-efficient telemanipulation in large, mobile platforms, a departure from conventional hand-centric paradigms (Li et al., 10 Feb 2026, Purushottam et al., 2024).
Sensorless and minimal-retuning architectures: Techniques like input-gated bilateral teleoperation (IGBT) achieve robust, low-latency, sensor-free force feedback suitable for embedded, resource-constrained hardware—broadening the deployable base for force-feedback telemanipulation (Kanai et al., 10 Sep 2025).

Continued innovation is expected to arise in areas including passivity-preserving learning-based force estimation, multi-point and high-bandwidth tactile rendering, and seamless cross-embodiment feedback generalization. These trajectories promise increasingly transparent, flexible, and scalable telemanipulation for both expert and non-expert operators.