Force-Feedback Imitation Learning Approach

• Force-feedback imitation learning integrates force measurements into demonstrations to enable accurate force modulation during robotic manipulation.
• It leverages sensor-based data, sensorless estimations, and haptic feedback to capture essential force signals for robust policy learning.
• This approach significantly improves manipulation performance and adaptability in tasks such as assembly, delicate grasping, and insertion.

Force-feedback imitation learning is an approach in robot learning that explicitly incorporates real or estimated interaction forces during the demonstration and control phases of imitation learning. Rather than relying solely on position or kinematic data, this class of methods leverages force feedback—whether rendered to a human demonstrator during demonstration capture, collected explicitly with force/torque sensors, or estimated via model-based or sensorless techniques—to enable more robust, compliant, and adaptable manipulation in contact-rich tasks. Force-feedback mechanisms inform both the quality of collected demonstration data (by enabling demonstrations that reflect expert force modulation) and the learning/execution of control policies that adapt to and exploit force signals for success in fine manipulation, assembly, or adaptive grasping.

1. Key Concepts and Rationale

Traditional imitation learning approaches are often position-centric, recording and replaying joint, pose, or trajectory data generated by a human demonstrator. However, in many manipulation tasks—especially those requiring contact, precise force control, or compliance—force data is essential for faithful skill transfer and safe, robust operation. Force-feedback imitation learning addresses this by:

• Providing real-time haptic cueing to the human teacher during data collection, e.g., via force-feedback gloves, robot-augmented haptics, or bilateral teleoperation (Rueckert et al., 2015, Li et al., 2023, Ge et al., 21 Sep 2025);
• Explicitly recording force-torque signals (from end-effector, finger, or wrist sensors) during demonstrations for reproduction or for learning force-informed action policies (Chen et al., 17 Jan 2025, Le et al., 2021, Ablett et al., 2023);
• Incorporating force predictions or commands into learned policies (via behavior cloning, diffusion models, or reinforcement/imitation learning hybrids) that condition on and output force information (Liu et al., 10 Oct 2024, Tian et al., 22 Jan 2024);
• Employing model-based or sensorless force estimation (e.g., using joint torques and kinematics via Jacobian pseudoinverse) when explicit F/T sensors are not available (Ge et al., 21 Sep 2025, Yamane et al., 8 Jul 2025).

The inclusion of force feedback allows robots to adapt to object pose uncertainty, variable contact surfaces, reproducibly exert target contact forces, and avoid over-gripping or damaging objects. The approach is especially critical for tasks involving object insertion, assembly, grasping deformable or delicate items, and dynamic manipulation.

2. Demonstration Collection with Force Feedback

Force-feedback demonstration systems are constructed to both capture high-fidelity force interactions and improve the quality of the demonstration through haptic cueing. Key implementations include:

• Sensor gloves with force feedback: Operator fingers are equipped with flex sensors for kinematic data and vibration motors for haptic feedback, with the robot’s tactile sensor data mapped in closed loop to guide the grasp force during demonstrations (Rueckert et al., 2015).
• Bilateral teleoperation: 4-channel bilateral controllers synchronize master (human-control input) and slave (robot) positions, enforcing action-reaction force duality (e.g., via $\theta_m - \theta_s = 0$ and $\tau_m^\mathrm{res} + \tau_s^\mathrm{res} = 0$ (Adachi et al., 2018)). Force sensors or estimated wrenches are used to render physical feedback to the operator, and the system records separated acting and reaction forces for learning.
• Hand-over-hand demonstration (HIRO Hand): A wearable robotic hand allows the operator to use their own tactile feedback via springs, potentiometers, and mechanical coupling, directly inferring the amount of force applied through systematic joint deformation (Wei et al., 2023).
• Sensorless force estimation: When direct force sensing is absent, methods employ calibrated dynamic models, Jacobian-based torque mapping, and observers to estimate the external forces acting on the robot from its joint torques and kinematic state (Ge et al., 21 Sep 2025, Yamane et al., 8 Jul 2025).
• Custom teleoperation controllers and haptic devices: Real-time force signals from instrumented grippers or end-effectors (e.g., force sensors integrated with custom robotic fingers and linearized via op-amp circuits) are fed back to the human operator via haptic controllers actuated by electric motors, with force mapped according to physical principles (e.g., Newton's third law) to provide resistance or compliance (Bazhenov et al., 10 Apr 2025, Satsevich et al., 1 Oct 2025).

These mechanisms ensure that demonstration data are physically meaningful with respect to contact forces, and that the demonstrated trajectories reflect expert-level force modulation.

3. Learning Algorithms Integrating Force Information

Learning frameworks for force-feedback imitation learning combine positional/kinematic data with force profiles in policy representation. Representative approaches include:

• Probabilistic trajectory representations: Demonstrations are modeled as sequences $\tau = \{y_1, ..., y_T\}$, where each state $y_t$ includes mapped finger joint commands; linear regression with Gaussian basis functions is used to fit trajectory weights, and variances are interpreted as possible adaptive control gains (Rueckert et al., 2015).
• Hybrid trajectory and force learning: Hierarchical architectures combine goal-conditioned imitation learning for nominal trajectory following with reinforcement learning-based force control modules (e.g., tuned via Soft Actor-Critic to adapt controller gains for force regulation) (Wang et al., 2021).
• Behavioral cloning with force-informed targets: In DexForce, demonstration actions are computed by augmenting the observed fingertip position $x_o$ with the measured force $f$ via $x_f = x_o + K_f f$, and policies are trained on the resulting force-informed targets (Chen et al., 17 Jan 2025). Policies may also condition on force data directly to improve adaptability.
• Diffusion and transformer models: Diffusion-based policies and Transformer architectures take multimodal sensory input, including force/torque readings, and output both position and wrench (force) predictions. Self-attention mechanisms enable filtering critical force cues for manipulation tasks (e.g., opening a bottle cap without slip) (Kim et al., 2022, Liu et al., 10 Oct 2024).
• Force-matching and impedance control: Kinesthetic teaching is augmented with tactile force measurements, and replay trajectories are generated by inverting impedance relations (e.g., $$x^{d}_{\mathrm{rep},t} = K^{-1} \mathcal{F}_{\mathrm{raw},t} + x_{\mathrm{raw},t}$$) to synthesize trajectories that produce both demonstrated positions and contact forces (Ablett et al., 2023). Impedance-controlled torque tracking further assures compliance during task execution (Ge et al., 21 Sep 2025).

Such architectures enable robots to generalize beyond pose-only skill reproduction, achieving task goals that depend on specific force application patterns (critical for insertion, sliding, and sequential manipulation).

4. Control Strategies and Policy Execution

During policy execution, force feedback is leveraged in several ways:

• Closed-loop impedance control: Learned trajectories (including force information) are executed on robots using low-level impedance or hybrid force-position control. The controller ensures compliance and safe force application, typically following a law of the form

$F = K_p (x_d - x_c) - K_v \dot{x}_c$

with gains adaptive to modeled or observed demonstration variance (Rueckert et al., 2015, Ablett et al., 2023).

• Hybrid force–position control primitives: For contact-rich manipulation, a hybrid controller selects between position control and active force control based on the magnitude of predicted force (e.g., use force control when $F > 6$N) (Liu et al., 10 Oct 2024). Force commands may be orthogonalized to motion directions for decoupling.
• Adaptive motion planning with joint torque feedback: Real-time joint torques sensed during hand-object contact are used to infer object positions and adjust grasping or manipulation trajectories, exploiting torque signals as a virtual tactile modality (Tian et al., 22 Jan 2024).
• Force estimator integration: For robots without F/T sensors, force at the end-effector is estimated by

$$F_{\mathrm{ext}} = J(q)^{T+} (\tau_{\mathrm{obs}} - \tau_{\mathrm{inv}})$$

where the pseudoinverse Jacobian maps joint torque deviations (after compensating model dynamics via a digital twin) into Cartesian force estimates (Ge et al., 21 Sep 2025).

These strategies enable the robot to remain compliant, maintain critical contact forces during assembly or manipulation, and adapt to variability in task conditions or object properties.

5. Experimental Findings and Applications

Empirical studies across multiple platforms and tasks demonstrate the strong benefits of force-feedback imitation learning:

• Enhanced task performance and generalization: Including force information in both demonstration and policy improves success rates—e.g., DexForce achieves 76% average success with force-informed actions versus near-zero without; tactile force matching improves policy success by over 62.5% in door-opening tasks (Chen et al., 17 Jan 2025, Ablett et al., 2023).
• Task adaptability and robustness: Force-feedback allows rapid on-the-fly adaptation to new object properties (e.g., using pre-trained VAE sponge representations for adaptive wiping with nearly 96% reference force accuracy over 40 scenarios of sponge/surface variations) (Tsuji et al., 9 May 2025). Hybrid learning approaches enable robots to bridge simulation–reality gaps in insertion or peeling tasks (Liu et al., 10 Oct 2024).
• User experience and data quality: Force feedback in demonstration interfaces—whether sensor gloves, haptic controllers, or bilateral teleoperation—produces demonstrations with lower and less variable applied forces, shorter execution times, and more robust resultant policies (e.g., 98.75% insertion success in assembly with force-guided GP correction) (Li et al., 2023, Jha et al., 2021).
• Low-cost, scalable dataset collection: Open-source, affordable teleoperation systems (Echo, GELLO, Prometheus) have integrated force feedback for collecting high-quality datasets suitable for bimanual and delicate manipulation, using hardware ranging from UR manipulator-compatible joysticks to HTC Vive-based tracking and custom force-sensing robotic fingers (Bazhenov et al., 10 Apr 2025, Sujit et al., 18 Jul 2025, Satsevich et al., 1 Oct 2025).

A representative table summarizing these findings:

Platform / Method	Force Feedback Mechanism	Reported Performance / Impact
HIRO Hand (Wei et al., 2023)	Wearable robotic hand (springs, mechanical sensors)	21 grasp types; 72.2% seq. task success
DexForce (Chen et al., 17 Jan 2025)	F/T sensors at finger base; kinesthetic teaching	76% avg. task success (near-zero w/o force)
FILIC (Ge et al., 21 Sep 2025)	Joint torque-based virtual F/T estimation, dual-loop IL	80-90% real-world/sim success with force
ForceMimic (Liu et al., 10 Oct 2024)	Handheld force-capture demo, hybrid force-motion IL	+54.5% vs. vision-only; 85% long-peel success
Prometheus (Satsevich et al., 1 Oct 2025)	Robotiq gripper, silicone-padded FSR, motorized feedback	35.77% reduction in grip force; 90% P+F policy success

These results highlight that integrating physical force signals—either via direct sensing or high-fidelity estimation—substantially improves the robustness and generalization of manipulation skills, especially in scenarios characterized by contact uncertainty or non-rigid interactions.

6. Limitations and Open Challenges

Despite clear benefits, force-feedback imitation learning faces several challenges:

• Platform dependence and transferability: Many systems employ demonstration interfaces (e.g., gloves or teleoperation) specifically designed or calibrated for a given robot's kinematics; skill transfer across morphologically dissimilar hands or arms may require nontrivial remapping (Rueckert et al., 2015, Wei et al., 2023).
• Sensor cost and scalability: Accurate F/T sensors remain costly for multi-DOF hands; although sensorless estimation methods exist, model mismatch and real-world nonlinearities can impact force estimation fidelity (Ge et al., 21 Sep 2025, Yamane et al., 8 Jul 2025).
• Adaptive control leveraging learned uncertainty: While probabilistic models (e.g., Bayesian regression, diffusion models) provide variance information, integration of demonstration uncertainty into adaptive low-level control (e.g., dynamic impedance adaptation) is not always realized and remains an area for future work (Rueckert et al., 2015).
• Integration with multi-modal and high-level policies: Learning robust, scalable policies that optimally combine force feedback with vision, proprioception, and possibly language remains an active frontier, particularly for tasks with deformable or multi-contact objects (Tsuji et al., 9 May 2025, Liu et al., 10 Oct 2024).
• Human factors and demonstration quality: In some user studies, force feedback was most beneficial to experienced users, while the effect on novice demonstrators was less pronounced, suggesting the need for better haptic interfaces or adaptive feedback strategies (Sujit et al., 18 Jul 2025).

These challenges motivate ongoing development of general, scalable frameworks for force-centric imitation learning and demonstration collection pipelines adaptable to diverse robotic embodiments and task structures.

7. Broader Impacts and Prospective Directions

Force-feedback imitation learning frameworks have a broad impact on both fundamental robotics research and practical automation:

• Manufacturing and assembly: Hybrid force–motion learning policies offer robust operation in uncertain, contact-rich industrial environments, with demonstrated success across insertion, assembly, and high-precision manipulation tasks (Wang et al., 2021, Jha et al., 2021).
• Service robots and home automation: Improved adaptation to deformable, fragile, or unknown objects (e.g., wiping, laundry folding, food preparation) is facilitated by force/torque-informed controllers trained via force-feedback demonstrations (Tsuji et al., 9 May 2025, Liu et al., 10 Oct 2024).
• Educational, open-source robotics: Low-cost, scalable force-feedback platforms democratize access to high-quality data collection and learning, benefiting academic research groups and emerging commercial entities (Bazhenov et al., 10 Apr 2025, Satsevich et al., 1 Oct 2025).
• Human–robot interaction: Tactile and force-aware demonstration collection systems that render haptic feedback improve both the safety and the reliability of human-guided training and teleoperation setups, promoting more intuitive programming and safe robot autonomy (Li et al., 2023, Sujit et al., 18 Jul 2025).

This suggests that future research integrating multimodal perception (force, vision, proprioception), adaptive control, and large-scale behavior cloning will further advance force-feedback imitation learning, particularly by enabling robots to generalize across tasks, objects, and real-world settings with diverse and uncertain contact dynamics.