Force-Feedback Imitation Learning

• Force-feedback imitation learning is a technique that explicitly incorporates sensed or estimated contact forces into imitation learning to achieve compliant and robust robotic manipulation.
• It employs multimodal policy architectures, including LSTMs, Transformers, and diffusion models, alongside hybrid position–force control strategies for adaptive skill transfer.
• The approach has demonstrated improved success in tasks like peg-in-hole and door opening by leveraging immersive teleoperation, precise force estimation, and data augmentation methods.

Force-Feedback Imitation Learning encompasses a broad class of algorithms, data collection methodologies, control strategies, and hardware interfaces that explicitly incorporate contact force information—either sensed, estimated, or retargeted—into the formulation, training, and deployment of imitation learning (IL) policies for manipulation tasks. By leveraging force feedback together with positional or visual data, such systems can achieve compliant, adaptive, and robust control in contact-rich environments. Critical advances include the separation of acting and reaction forces for skill transfer, multimodal policy architectures fusing force with vision or proprioception, closed- and hybrid-loop control schemes with active force execution, and new methods for teleoperation, dataset augmentation, and safe deployment.

1. Core Principles and Control Architectures

Force-feedback IL is grounded in several foundational control architectures. The most influential is the four-channel bilateral control law (Adachi et al., 2018, Yamane et al., 8 Jul 2025, Masuya et al., 4 Dec 2024), which utilizes paired master–slave manipulators to separate human-applied (acting) forces and environment-driven (reaction) forces. Mathematically, the bilateral transparency constraints are: $$\theta_m - \theta_s = 0 ; \quad \tau_m^{res} + \tau_s^{res} = 0$$ where $\theta$ are joint positions and $\tau^{res}$ are reaction torques on master (m) and slave (s) sides.

Controllers are typically hybrid position–force, combining impedance (PD) for position and admittance (P) for force: $$\tau_{imp} = K_p (\theta_{ref} - \theta) + K_d (\dot\theta_{ref} - \dot\theta), \quad \tau_{adm} = K_f (\tau_{ref} - \tau_{obs})$$ Bilateral control is augmented by disturbance observers and reaction-force observers to maintain safety and precise force tracking in real time (Masuya et al., 4 Dec 2024).

An alternative is sensorless estimation via accurate manipulator dynamics (Yamane et al., 8 Jul 2025). Here, model-compensated joint torques and observer-based filtering yield external force estimates: $$\tau_{ext} = M(\theta)\ddot\theta + C(\theta, \dot\theta)\dot\theta + g(\theta) - \tau_{cmd}$$ where $M, C, g$ are inertia, Coriolis, and gravity terms, with $\tau_{cmd}$ as the commanded torque.

2. Data Acquisition: Teleoperation with Force Feedback

High-quality force-feedback demonstration data is essential for successful IL in contact-rich tasks. State-of-the-art systems achieve this via immersive teleoperation, haptic gloves, and force-augmented demonstration interfaces:

• Immersive Demonstration: Fingertip and palm-level resistive feedback (SenseGlove + Panda arm) drive consistent, low-force demonstrations (Li et al., 2023). Duty cycles to fingertips are empirically mapped to commanded forces; palm-level wrenches are rendered by PD control at the user's palm.
• Open-Source Teleoperation: Prometheus (Satsevich et al., 1 Oct 2025) offers real-time haptics (based on handle-mounted force sensors) to modulate operator input according to instantaneous object contact forces.
• Low-Cost and Dexterous Gloves: DOGlove (Zhang et al., 11 Feb 2025) provides 21-DoF hand pose capture, multi-modal (torque + haptic) feedback via cable-driven servos, and precise fingertip force rendering, enabling both kinesthetic and vision-free force-rich teleoperation.
• Pseudo-tactile Signals for simulation learning: Filtered force/torque readings are thresholded into binary grasp-confirmation flags (Yang et al., 31 Mar 2025), enabling robust policy training on pure simulation data.

Dataset protocols standardize synchronized logging of RGB, pose, force, and gripper states at rates matching sensor capabilities (10 Hz–1 kHz), with calibration curves to map voltage/current readings to physical force.

3. Policy Representations and Learning Paradigms

Force-feedback IL leverages diverse policy architectures—recurrent neural networks (LSTM), Transformers, and diffusion policies—each adapted to fuse force, position, vision, and auxiliary modalities:

RNN and LSTM Architectures

RNNs ingest normalized time-series of joint angles, velocities, and forces, predicting either low-level torque commands (controller learning) or high-level position/force commands (command learning) (Adachi et al., 2018). Two-layer LSTM stacks (100 units each) with tanh activations are typical.

Transformer and Cross-Attention Models

Vision and force tokens are encoded by CNN/MLPs and fused via cross-attention before temporal self-attention in multi-layer Transformers (Ge et al., 21 Sep 2025, Sujit et al., 18 Jul 2025, Kim et al., 2022). The output is usually a sequence of pose or controller references, while force features are provided directly as inputs.

Diffusion Policy and Hybrid Action Spaces

Advanced methods (FARM (Helmut et al., 15 Oct 2025), ForceMimic (Liu et al., 10 Oct 2024)) employ denoising diffusion models in trajectory or hybrid pose-force spaces. Actions include absolute robot pose, grip width, and explicit force targets; tactile-conditioned FiLM layers inject high-dimensional force maps into every network block. Losses combine MSE trajectory matching with explicit force distribution objectives (Wasserstein-1 distance for force trace fidelity).

Bimodal Policy Inputs and Output Spaces

Policies often incorporate multimodal sensor inputs—visual features (RGB, point clouds, tactile images), proprioception, and explicitly measured or estimated force/torque readings. Some frameworks decouple force inputs and outputs, allowing force-only or force+motion learning for adaptive hybrid control (Yamane et al., 8 Jul 2025, Liu et al., 10 Oct 2024).

4. Integration of Force Feedback into Imitation Learning

Force feedback is incorporated in IL pipelines through both demonstration data and explicit policy execution. Key approaches include:

• Augmenting Demonstrations: Force-augmented demonstrations consistently yield lower contact force, reduced variance, and faster completion times. Learned policies trained only on position but sourced from force-informed demonstrations inherit these properties (Li et al., 2023, Satsevich et al., 1 Oct 2025, Zhang et al., 11 Feb 2025).
• Explicit Force in Policy Encoding: Policies that directly ingest force signals (full 6D wrench, tactile images, etc.) exhibit superior performance in high-precision or multi-contact tasks, notably unscrewing, nut turning, or soft object manipulation (Chen et al., 17 Jan 2025, Helmut et al., 15 Oct 2025).
• Hybrid Position–Force Outputs: Diffusion and hybrid policies predict both desired motion increments and force targets, which are tracked by impedance or admittance controllers in execution (Liu et al., 10 Oct 2024, Ge et al., 21 Sep 2025).
• Data Augmentation via Real-World Variation: Variable-speed playback augments real-world datasets with diverse contact-force profiles, improving generalization and interpolation accuracy across task durations and frequencies (Masuya et al., 4 Dec 2024).

5. Experimental Outcomes and Application Domains

Force-feedback IL frameworks have been validated on a range of manipulation and assembly tasks, with empirical evidence demonstrating marked gains in robustness, compliance, efficiency, and adaptability:

Task	Success Rate (Force-Augmented)	Baseline Rate	Notable Gains
Line Drawing (OOT)	70–90%	<75%	Unseen angles (Adachi et al., 2018)
Peg-in-Hole	80–90%	46–68%	Generalizes to offset holes (Ge et al., 21 Sep 2025)
Pick-and-Place	88–90%	30–80%	Softer, faster behaviors (Li et al., 2023, Satsevich et al., 1 Oct 2025)
Vegetable Peeling	85% (>10 cm length)	55%	Stable force tracking (Liu et al., 10 Oct 2024)
Drawer/Door Opening	93–100%	<62%	Stable adaptation (Sujit et al., 18 Jul 2025)
Screwing, Plant	95–100%	<10–65%	Dynamic force adaptation (Helmut et al., 15 Oct 2025)

Force-matched imitation (via tactile sensors) led to average policy gains of 62.5%, mode switching provided +30.3% (door tasks), and multimodal visuotactile input a further +42.5% (Ablett et al., 2023). Incorporating pseudo-tactile feedback eliminated gripper ambiguity, raising disturbance-resilience rates to 100% and overall success up to 90% (Yang et al., 31 Mar 2025).

6. Limitations, Open Issues, and Extensions

Despite advances, several limitations persist. Sensor accuracy (limited resolution or drift in force/torque/tactile sensors), hardware constraints (actuator range, device mass), and model errors can degrade feedback fidelity (Adachi et al., 2018, Ge et al., 21 Sep 2025, Helmut et al., 15 Oct 2025). Many systems remain restricted to planar or limited-DOF tasks, with vision integration and full 6D wrench sensing as ongoing challenges.

Trade-offs exist in controller design: end-to-end hybrid learning can be robust in-contact but may suffer instability or oscillations outside trained conditions, while command-only learning is more stable but less adaptive. Proper gain tuning, online stiffness/damping adaptation, and model compensation are active areas of research.

Extensions include online model residual learning for force estimation, richer multi-modal fusion (tactile–vision–language), joint adaptation of impedance parameters, and scaling to broader skill libraries and task families. Closed-loop learning with tactile feedback and reinforcement learning is being studied for adaptive compliance in unstructured environments (Liu et al., 10 Oct 2024, Wang et al., 2021).

7. Significance and Impact in Robotic Manipulation

Force-feedback imitation learning fundamentally improves robotic manipulation in contact-rich domains by equipping policies with the ability to regulate and adapt environmental interaction forces, mirroring human dexterity. This enables safer, more compliant insertion, assembly, object handling, and manipulation across rigid, soft, and deformable objects. The separation and explicit encoding of force data, together with multimodal training and hybrid control strategies, have raised policy robustness, adaptability, and sample efficiency—setting the foundation for scalable, high-precision skill transfer in industrial and everyday contexts. The domain continues to expand with open-source systems, cost-effective hardware, and data-rich teleoperation protocols, pushing the boundaries of robot learning in the real world.