Learning Force-Regulated Manipulation with a Low-Cost Tactile-Force-Controlled Gripper

Abstract: Successfully manipulating many everyday objects, such as potato chips, requires precise force regulation. Failure to modulate force can lead to task failure or irreversible damage to the objects. Humans can precisely achieve this by adapting force from tactile feedback, even within a short period of physical contact. We aim to give robots this capability. However, commercial grippers exhibit high cost or high minimum force, making them unsuitable for studying force-controlled policy learning with everyday force-sensitive objects. We introduce TF-Gripper, a low-cost (~$150) force-controlled parallel-jaw gripper that integrates tactile sensing as feedback. It has an effective force range of 0.45-45N and is compatible with different robot arms. Additionally, we designed a teleoperation device paired with TF-Gripper to record human-applied grasping forces. While standard low-frequency policies can be trained on this data, they struggle with the reactive, contact-dependent nature of force regulation. To overcome this, we propose RETAF (REactive Tactile Adaptation of Force), a framework that decouples grasping force control from arm pose prediction. RETAF regulates force at high frequency using wrist images and tactile feedback, while a base policy predicts end-effector pose and gripper open/close action. We evaluate TF-Gripper and RETAF across five real-world tasks requiring precise force regulation. Results show that compared to position control, direct force control significantly improves grasp stability and task performance. We further show that tactile feedback is essential for force regulation, and that RETAF consistently outperforms baselines and can be integrated with various base policies. We hope this work opens a path for scaling the learning of force-controlled policies in robotic manipulation. Project page: https://force-gripper.github.io .

• The paper introduces RETAF, a decoupled policy that separates high-frequency force regulation from low-frequency pose prediction using tactile and visual feedback.
• It presents TF-Gripper, a cost-effective ($150) tactile-force-controlled gripper achieving up to 68% stable grasp success on force-sensitive tasks.
• Experimental results show that integrating direct tactile feedback with reactive force control significantly outperforms conventional position-based strategies.

Learning Force-Regulated Manipulation with a Low-Cost Tactile-Force-Controlled Gripper

Introduction and Motivation

This paper addresses the challenge of force-regulated robotic manipulation, focusing on everyday objects highly sensitive to applied force, such as potato chips or tomatoes. Conventional commercial grippers either lack sufficiently low minimum force or are prohibitively expensive, impeding accessible research on learning robust force-control policies. Moreover, prior policy architectures predominantly rely on gripper position or width control, which serve as indirect proxies and exhibit high sensitivity to object and task-specific variations.

To overcome these constraints, the authors introduce TF-Gripper, a cost-effective ($\sim$ \$150) tactile-force-controlled parallel-jaw gripper with an effective force range of 0.45–45 N. TF-Gripper integrates tactile sensors and supports cross-robot compatibility. Additionally, a teleoperation interface is proposed that enables collection of human-applied grasping force, providing high-quality data with ground-truth force profiles for learning.

Beyond hardware, the central contribution is RETAF (Reactive Tactile Adaptation of Force), a decoupled policy architecture where force regulation is handled separately at high frequency based on wrist-view and tactile feedback, while pose prediction is managed by a base policy at lower frequency. This framework aims to resolve latency and modality fusion challenges, enabling robust, reactive manipulation.

Figure 1: TF-Gripper with tactile sensing enables learning force-regulated manipulation, mimicking human capability to modulate contact force.

TF-Gripper Hardware System

TF-Gripper employs a dual-motor timing-belt actuation mechanism, translating actuator current to fingertip force with minimal backlash and consistent moment arm. The main structure is entirely 3D printed, leveraging cost-efficiency and adaptability. Soft-fingertip pads and piezo-resistive tactile sensors (FlexiTac) are integrated for compliant contact and real-time tactile feedback.

The actuator exhibits a near-linear current-to-force mapping, empirically validated, with a force resolution of $\sim$ 0.2 N and temperature-compensated drift below 4%. An interchangeable adapter ensures compatibility across robot arms such as Franka, UR5, and KUKA.

Figure 2: Key mechanical and sensing components of TF-Gripper illustrating full 3D-printed setup and integrated wrist camera.

Teleoperation for force data collection is achieved via finger rings connected to actuators mimicking spring-like resistance. Force input from operators is measured by motor current, calibrated, and directly mapped to the robot gripper's actuation, facilitating accurate demonstration transfer.

Figure 3: Teleoperation interface integrating VR controller and force-sensing finger rings for recording human-applied grasping force.

RETAF Policy Framework

RETAF addresses two principal obstacles: frequency mismatch (slow pose prediction bottlenecks reactive force regulation) and unstable modality fusion (directly combining tactile and global visual context impedes learning). It decouples the control policy into (i) a base policy predicting end-effector pose and gripper open/close action at low frequency, and (ii) a Force Adaptation Policy operating at $\geq$30 Hz, triggered upon gripper closure, to predict continuous force based on wrist-view and tactile inputs via joint-attention.

Figure 4: RETAF architecture decouples base pose/gripper action prediction from high-frequency reactive force adaptation using wrist-view and tactile inputs.

Base policy training follows conventional behavior cloning objectives on pose and open/close actions. The force adaptation policy is trained via supervised regression (MSE) against human-applied force profiles, using a lightweight network that enables real-time execution.

Experimental Setup and Tasks

Experiments are conducted on five manipulation tasks requiring precise force regulation: tofu grasping, chip picking, cherry tomato picking, liquid transfer, and cherry tomato harvest. Each task enforces criteria for stable grasp and task success, sensitive to excessive or insufficient force.

Figure 5: Evaluation environment setup showing representative task instrumentation.

Data collection utilizes the teleoperation interface, capturing trajectories at 15 Hz with full pose, force, position, tactile, and visual data. Each task includes 50 human demonstrations, ensuring comprehensive coverage of force-sensitive dynamics.

Baseline Policies and Implementation

Baselines include Diffusion Policy (DP), with and without tactile input, and ViTac-MAE, which pretrains tactile encoders via masked autoencoding for improved contact representation. RETAF is compared against variants representing gripper actions as position, width, open/close, or continuous force. All methods encode visual observations using CLIP, tactile data with CNNs, and employ 1D U-Net diffusion backbones.

Experimental Results and Analysis

Force vs. Position Control

Force-controlled policies consistently achieve higher stable grasp rates than position-controlled counterparts across all tasks and architectures. For example, RETAF with force control attains a stable grasp rate of 68%, compared to 44% for position control. This improvement is pronounced during contact, not object reaching, underscoring the centrality of direct force regulation.

Figure 6: RETAF tracks ground-truth force in demonstrations more accurately than DP/Force, especially capturing correct open timing.

Role of Tactile Sensing

Direct fusion of tactile input provides only marginal improvements in baseline DP models. ViTac-MAE improves grasp stability via stronger tactile encoding but remains limited. RETAF, explicitly leveraging high-frequency tactile feedback for reactive control, demonstrates consistent performance gains across all stages, validating its architectural decoupling.

RETAF Policy Effectiveness

RETAF outperforms baselines in reach, stable grasp, and task success rates. Restricting gripper action prediction to open/close simplifies learning and improves pose precision, enabling more consistent object reaching. RETAF’s high-frequency adaptation substantially increases task success in force-sensitive scenarios.

RETAF with Different Base Policies

RETAF demonstrates robust integration with various base policies, including DP and $\pi_{0.5}$. When paired, reach, stable grasp, and task success rates improve substantially across representative tasks. This confirms the model-agnostic utility of RETAF’s force adaptation module.

Qualitative Failure Analysis

Figure 7: RETAF achieves stable grasps; DP variants show common failures including excessive force (breakage), insufficient width (slip), and inaccurate pose prediction.

Failure modes in DP often arise from overly large widths leading to slip, or excessive force causing breakage, while RETAF predicts more nuanced and appropriate forces, evidenced in both quantitative and qualitative rollouts.

Force-Tactile Dynamics

RETAF closely reproduces the characteristic force modulation patterns observed in human demonstrations around initial contact, in contrast to noisier and less precise predictions from DP. This fidelity is critical for tasks demanding rapid, precise adaptation.

Figure 8: RETAF captures human-like force modulation during contact, outperforming baseline diffusion policy in time-aligned correlation.

Dataset Diversity

Figure 9: Representative diversity of gripper-view images in training and evaluation; four manipulation tasks shown prior to grasping.

Data distribution consistency is verified; observed performance differences are attributed to policy capabilities rather than domain shifts.

Implications and Future Directions

This work empirically substantiates the advantages of explicit force control and tactile feedback in learning-based manipulation, particularly for force-sensitive objects and tasks. RETAF’s decoupled architecture is theoretically motivated by control latency and modality fusion challenges, and demonstrates practical robustness across tasks and policy backbones.

The introduction of TF-Gripper coupled with RETAF establishes an accessible platform for scalable force-regulated manipulation research. Practical implications extend to automated handling in food processing, gentle assembly, and biological specimen manipulation. Future work includes theoretical analysis of force-vs-position control dynamics, broader object/task coverage, and scaling with comprehensive force/tactile datasets.

Conclusion

The paper demonstrates that low-cost, tactile-force-controlled hardware combined with reactive policy architectures significantly advances robust force-regulated manipulation. By decoupling pose and force prediction, RETAF enables high-frequency adaptation to contact dynamics, validated across diverse tasks and policy frameworks. The findings provide both practical tools and theoretical insights for advancing manipulation in force-sensitive domains.