Organizations: Department of Computer Science, University of Virginia, Charlottesville, VA, USA. · Department of Computer Science, Columbia University, New York, NY, USA.
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 0.45-45 N and is compatible with different robot arms. Additionally, we designed a teleoperation device paired with TF-Gripper to record human-applied grasping forces. While we can train standard low-frequency policies with the collected force data, achieving reliable performance remains challenging due to the reactive and contact-dependent nature of force-regulated manipulation. 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. Our experiments show that, compared to position control, direct force control with TF-Gripper improves grasp stability and overall task performance across six real-world tasks. 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 .
Figures & tables
Fig. 1: To mimic the human ability to regulate force through tactile feedback to manipulate everyday objects, we develop a force-controlled gripper with tactile sensing for learning reactive and gentle manipulation policies.
Fig. 2: Gripper position vs. force control on objects with similar appearance but different physical properties. (a) The same type of potato chips but with a small variation in size. (b) The same cherry tomato after two days of softening.
Fig. 3: Hardware design of the TF-Gripper. (a) An adapter connects the gripper to a robot and two motors actuate the fingertips along linear rails by pulling a timing belt. (b) The overall TF-Gripper setup. (c) A camera provides a wrist view, with soft fingertips integrated with tactile sensors.
Fig. 4: PWM–force actuation of the TF-Gripper. Left: the PWM command controls the motor rotational torque, which drives the fingertip through the pulley and timing-belt transmission to regulate the applied contact force. Right: the measured PWM–force relationship at the fingertip contact surface, showing the effective grasping force.
Fig. 5: Design of the teleoperation device. A VR controller controls the robot’s end-effector pose. Two finger rings are connected to the motors; they detect the force applied by the human operator to drive the TF-Gripper.
Fig. 6: Overview of the RETAF framework. The base policy predicts end-effector pose and gripper open/close from visual and proprioceptive observations at a low frequency. When the gripper closes, the force adaptation policy is activated to predict grasping force at a high frequency using joint attention over wrist-view and tactile inputs.
Fig. 7: Overview of the evaluation environment and tasks.
Fig. 8: Performance of six manipulation tasks measured in three stages: Reach, Stable Grasp, and Task Success. Bars show the mean ± standard error over three seeds. Hatched bars use position control and solid bars use force control.
Fig. 9: Ablation study. Left: per-task Grasp/Success rates (mean ± standard error over three seeds, 10 rollouts each). Right: averages over the 3 tasks (arrows: absolute change vs. RETAF). † ViTac-MAE: pretrained tactile encoder added to RETAF.
Fig. 10: Force prediction on a validation demonstration. For each frame, we compare DP/Force and RETAF predictions with the human-applied force (scaled to [−1,1] , where −1 denotes gripper open) and plot the summed tactile value (scaled to [0,1] ). RETAF tracks the ground truth and the open timing more closely. Appendix -B shows more trajectories.
Fig. 11: (a) Two successful rollouts, where RETAF correctly predicts appropriate forces and achieves stable grasps. (b) Failure cases: (i) When DP predicts force or position, it shows a higher failure rate during the reach-to-target stage than DP predicting open/close actions. (ii) Regardless of the gripper action type, DP often fails to predict accurate end-effector poses. (iii) We observe that DP–position tends to open too wide, causing slip, while (iv) DP–force tends to predict excessive force, leading to object breakage. In contrast, (v) RETAF predicts more appropriate grasping forces.
Fig. 12: Generalization test on the Chip Pick task. Left: in-distribution chips (A–D). Right: out-of-distribution chips (E–F). Bars show mean ± standard error of Grasp (G) and Success (S) rates over three seeds, with 10 trials per seed.
Fig. 13: Rigid-object contact-rich tasks. (a) Block Insertion and Peg Insertion. (b) Success rates over 30 rollouts per task.
Task
10 demos
25 demos
50 demos
100 demos
Cherry Tomato Pick
30.0 / 6.7
73.3 / 36.7
83.3 / 73.3
90.0 / 80.0
Peg Insertion
36.7 / 6.7
66.7 / 16.7
83.3 / 33.3
96.7 / 46.7
TABLE III: Data scaling of RETAF: Grasp / Success rates (%) vs. number of demonstrations, 30 rollouts per setting.
Commanded (N)
TF-Gripper (N)
Franka Hand (N)
32
32.36±1.07
31.15±0.55
35
35.12±1.14
33.94±0.68
TABLE V: Measured fingertip force (mean ± std over 10 repetitions) of TF-Gripper and the Franka Hand at commanded forces in their overlapping range.
Fig. 14: Force and tactile signals before and after contact in human demonstrations (left) and in rollouts (right), for 10 randomly selected trajectories of Cherry Tomato Pick. The tactile sum is min-max scaled to [0,1] . The force is scaled to [−1,1] , where values above 0 give the magnitude of the grasping force.
Fig. 15: Randomly selected wrist-view images captured before grasping, for four tasks. For each task, we show 15 training samples and 6 rollout frames. Liquid Transfer is not shown because the cup and the tube are mounted on a fixed rig, and its wrist views are largely identical.
Cherry Tomato Pick
Liquid Transfer
Method
R
G
S
R
G
S
π0.5
80%
30%
20%
60%
20%
20%
π0.5 + RETAF
100%
90%
90%
70%
60%
50%
TABLE VI: RETAF with π0.5 as the base policy. Tasks are evaluated in three stages: Reach (R), Stable Grasp (G), and Task Success (S), with 10 rollouts per setting.
RETAF
RETAF w/o open/close in base policy
Task
Reach
Stable Grasp
Task Success
Reach
Stable Grasp
Task Success
Cherry Tomato Pick
100%
90%
80%
100%
20%
0%
Liquid Transfer
70%
60%
60%
80%
0%
0%
TABLE VII: Ablation on removing the open/close prediction from the base policy, with 10 rollouts per setting. These single-run results are not directly comparable to the three-seed results in Fig. 8 .
Method
Peg Insertion
Block Insertion
DP w/ Force
8/30 (26.7%)
11/30 (36.7%)
RETAF
10/30 (33.3%)
13/30 (43.3%)
TABLE VIII: Task success on the two rigid-object tasks with TF-Gripper, over 30 rollouts per task.
Fig. 16: Everyday tasks performed with TF-Gripper.
Fig. 17: Cross-gripper setup, shown with the Block Insertion task. TF-Gripper (left) and the Franka Hand (right) are mounted on the same Franka arm with the same tool center point, wrist camera, and tactile fingertips. The insets show the wrist-camera views.
Method
TF-Gripper
Franka Hand
DP
7/20 (35.0%)
4/20 (20.0%)
RETAF
13/20 (65.0%)
11/20 (55.0%)
TABLE IX: Cross-gripper comparison on Cherry Tomato Pick under position control, which both grippers support. Values are task success over 20 trials per setting.
How should a robot learn to manipulate objects so fragile that sub-Newton contact forces can cause irreversible damage? Existing visuo-tactile policy learning typically treats tactile sensing as an additional policy input. In direct-contact force-sensitive manipulation, however, the bottleneck can arise earlier, during data collection: manual gripper control is too delayed and coarse-grained to reliably maintain the narrow force range required for stable grasping. We therefore use a deterministic 25 Hz tactile reflex controller as a collection-time teacher, producing demonstrations with controller-shaped grasping behavior for tactile-free policy learning. On Action Chunking with Transformers (ACT), policies trained from reflex-shaped demonstrations recover the teacher's grasping profile and achieve 95% stable grasps on the nominal plastic-cup task, substantially outperforming visually screened manual demonstrations. The same intervention improves in-distribution stability on π0.5 and shows a favorable exploratory trend on an unseen paper-cup variant. Under randomized external disturbance, however, the reflex-data π0.5 policy still fails in 45% of policy-only trials, whereas a deployment-time reflex arbiter retains all grasps. These results reveal a new role for tactile feedback in force-sensitive manipulation: rather than integrating tactile into the policy, we use it as a collection-time teacher that shapes grasping behavior in demonstrations for policy learning, while disturbance rejection remains controller-dependent, revealing the boundary of tactile-free policy.
Ziyan Feng, Zizhao Yuan, Yulong Fu +6
The Hong Kong University of Science and Technology (Guangzhou)
Regulating grasping force to reduce slippage during dynamic object interaction remains a fundamental challenge in robotic manipulation, especially when objects are manipulated by multiple rolling contacts, have unknown properties (such as mass or surface conditions), and when external sensing is unreliable. In contrast, humans can quickly regulate grasping force by touch, even without visual cues. Inspired by this ability, we aim to enable robotic hands to rapidly explore objects and learn tactile-driven grasping force control under motion and limited sensing. We propose a physics-informed energy abstraction that models the object as a virtual energy container. The inconsistency between the fingers' applied power and the object's retained energy provides a physically grounded signal for inferring slip-aware stability. Building on this abstraction, we employ model-based learning and planning to efficiently model energy dynamics from tactile sensing and perform real-time grasping force optimization. Experiments in both simulation and hardware demonstrate that our method can learn grasping force control from scratch within minutes, effectively reduce slippage, and extend grasp duration across diverse motion-object pairs, all without relying on external sensing or prior object knowledge. (Video: https://youtu.be/l3TJV29Mo6w)
Cheng-Yu Kuo, Hirofumi Shin, Takamitsu Matsubara
Graduated School of Science and Technology, Nara Institute of Science and Technology, Nara 630-0192, Japan · Honda R&D, Ltd., Saitama 351-0114, Japan
Robotic foundation models (RFMs) are increasingly capable of general-purpose manipulation, yet reliable physical interaction remains challenging in contact-rich settings. We present FP2, a lightweight downstream interface that equips task-adapted RFMs with explicit force control while preserving their action-generation capability. FP2 adopts an action-regulation decomposition: the task-adapted RFM serves as a foundation policy responsible for task-level action generation, while a high-frequency force control policy focuses solely on interaction regulation. To condition force regulation on the ongoing manipulation, FP2 compresses foundation-policy contextual representations and combines them with wrench and proprioceptive histories to predict structured force-control parameters. We evaluate FP2 with four RFM backbones across four real-world contact-rich manipulation tasks. FP2 consistently improves task performance and force regulation quality over the corresponding foundation policies, while comparing favorably with representative force-aware and force-control baselines. Ablations further show that foundation-policy context and physical feedback are complementary for effective force regulation, while preserving foundation-policy action generation improves both efficiency and novel-object generalization. Project website: http://force-policy.github.io/fp2