Force Policy: Learning Hybrid Force-Position Control Policy under Interaction Frame for Contact-Rich Manipulation

Hongjie Fang; Shirun Tang; Mingyu Mei; Haoxiang Qin; Zihao He; Jingjing Chen; Ying Feng; Chenxi Wang

cs.ROFeb 25, 2026

Force Policy: Learning Hybrid Force-Position Control Policy under Interaction Frame for Contact-Rich Manipulation

Authors: Hongjie Fang, Shirun Tang, Mingyu Mei, Haoxiang Qin, Zihao He, Jingjing Chen, Ying Feng, Chenxi Wang,

Organizations: Noematrix · Shanghai Jiao Tong University · Zhejiang University · Shanghai Innovation Institute ·

Abstract

Contact-rich manipulation demands human-like integration of perception and force feedback: vision should guide task progress, while high-frequency interaction control must stabilize contact under uncertainty. Existing learning-based policies often entangle these roles in a monolithic network, trading off global generalization against stable local refinement, while control-centric approaches typically assume a known task structure or learn only controller parameters rather than the structure itself. In this paper, we formalize a physically grounded interaction frame, an instantaneous local basis that decouples force regulation from motion execution, and propose a method to recover it from demonstrations. Based on this, we address both issues by proposing Force Policy, a global-local vision-force policy in which a global policy guides free-space actions using vision, and upon contact, a high-frequency local policy with force feedback estimates the interaction frame and executes hybrid force-position control for stable interaction. Real-world experiments across diverse contact-rich tasks show consistent gains over strong baselines, with more robust contact establishment, more accurate force regulation, and reliable generalization to novel objects with varied geometries and physical properties, ultimately improving both contact stability and execution quality. Project page: https://force-policy.github.io/

Figures & tables

Fig. 1: Interaction Frame for Example Contact-Rich Tasks.

Interaction	Task Mode	Selection Mask $S$	Reference Wrench
Free	Free	$[0,0,0,0,0,0]$	$[-,-,-,-,-,-]$
Contact	Surface	$[0,0,1,0,0,0]$	$[-,-,\boldsymbol{f}_{z},-,-,-]$
	Insertion	$[0,1,1,0,1,1]$ *	$[-,0,0,-,0,0]$
	Rotation	$[0,0,1,0,0,0]$ *	$[-,-,\boldsymbol{f}_{z},-,-,-]$

TABLE I: Control Structure Selection based on Task Modes. The IF

x

-axis aligns with the primary motion/insertion direction, while the

z

-axis aligns with the dominant normal.

Fig. 2: Force Policy and Dual-Policy Asynchronous Scheduler. (Left) Force Policy consists of a global vision policy and a local force policy. The global policy provides task-level visual context and global actions, while the local policy predicts interaction structure and local actions to realize hybrid force-position control during contact. (Right) The dual-policy asynchronous scheduler switches between the two policies and reduces latency and jerk via model-agnostic chunk alignment using dynamic time warping (DTW) [ 59 ] .

Fig. 3: Tasks. We design three tasks spanning two categories (polishing and insertion) to evaluate different policies for contact-rich manipulation. The descriptions on the right highlight the key challenges of each task compared to similar tasks in prior literature. All tasks require highly accurate force regulation to be successfully completed. We randomize object placement within the workspace area during both data collection and evaluation for each task.

Policy	Push and Flip		Plug in EV Charger			Scrape off Sticker (Easy)			Scrape off Sticker (Hard)
Policy	push	flip	contact	match	plug in	contact	one-side	full off	contact	one-side	full off
RISE-2 [ 31 ]	100.0 %	42.5%	100.0 %	90.0 %	0.0%	100.0 %	80.0%	80.0%	100.0 %	20.0%	10.0%
$\pi_{0.5}$ [ 7 ]	80.0%	52.5%	100.0 %	30.0%	0.0%	70.0%	70.0%	70.0%	100.0 %	40.0%	20.0%
RDP [ 88 ]	100.0 %	57.5%	100.0 %	70.0%	5.0%	100.0 %	90.0%	70.0%	100.0 %	30.0%	20.0%
FoAR [ 36 ]	100.0 %	60.0%	100.0 %	80.0%	10.0%	100.0 %	70.0%	40.0%	100.0 %	60.0%	20.0%
ForceVLA [ 89 ]	50.0%	30.0%	20.0%	0.0%	0.0%	10.0%	0.0%	0.0%	10.0%	0.0%	0.0%

TABLE II: Evaluation Success Rates. We report success rates evaluated in an accumulative, stage-wise manner. Force Policy achieves the highest success rates compared to vision-based and force-aware baselines among all tasks.

Fig. 4: Visualization of Effective Forces during Deployment and from Demonstrations on the Scrape off Sticker (Hard) Task. All baselines fail to replicate the force behavior from demonstrations, resulting in degraded performance. Force Policy closely imitates the effective force in demonstrations, achieving higher success rates.

TABLE III: Force Regulation Evaluation on the Push and Flip Task. (Left) Pushing the heavy object requires approximately 45N, while demonstrations apply about 15N pushing force to flip the target object; we then measure its pushed distance $d$ as an indicator of force regulation. (Right) Statistics of the pushed distance for each method.

Panel of Table 8 — TABLE IV: Generalization Evaluation on the Push and Flip Task. We use unseen objects of different colors, geometries, and stiffnesses to evaluate the generalization ability of force-aware policies. Please refer to the supplementary material for detailed object comparisons.

Fig. 5: IF Recovery Evaluation on the Scrape off Sticker Task. Angular error is computed between the recovered and ground-truth force control axes; a failure is counted if it exceeds 20 ∘ .

Fig. 6: Asynchronous Scheduler Evaluation on the Push and Flip Task. The scheduler effectively reduces both motion and force jerks (left) , resulting in more smoother executed trajectories (right) .

Fig. 7: Sensor Signal Ambiguity. Similar sensor signals (wrench and twist) can lead to different interaction types and structures.

Fig. 8: Visualization of the Interaction Frame and the Task Mode on the Push and Flip Task.

Fig. 9: Visualization of the Interaction Frame and the Task Mode on the Plug in EV Charger Task.

Fig. 10: Visualization of the Interaction Frame and the Task Mode on the Scrape off Sticker Task.

Fig. 11: Visualization of Different IF Recovery Methods on the Scrape Off Sticker Task. Wrench-only IF recovery methods couple friction directions, whereas our IF recovery method better aligns with the ground-truth (the world $z$ -axis).

Policy	w/ Scheduler?	Success Rate $\uparrow$		Avg. $d$ (cm) $\downarrow$
Policy	w/ Scheduler?	push	flip	Avg. $d$ (cm) $\downarrow$
RISE-2 [ 31 ]		100.0 %	42.5%	0.86
RISE-2 [ 31 ]	✓	100.0 %	62.5 %	0.00
$\pi_{0.5}$ [ 7 ]		80.0%	52.5%	0.00
$\pi_{0.5}$ [ 7 ]	✓	95.0 %	77.5 %	0.00

TABLE V: Evaluation of Schedulers for Vision-Only Policies on the Push and Flip Task. Asynchronous scheduler improves the performance and make the contact more stable.

Fig. 12: Evaluation Metrics Explanation of Partial Success.

Fig. 13: Failure Analyses on the Push and Flip Task. Most failures arise from attempting to flip the object without first establishing contact with the wall. Force-aware policies using position control often exhibit unstable or unintended contacts.

Fig. 14: Failure Analyses on the Scrape off Sticker (Hard) Task. Force-aware VLA models typically fail to establish stable contact with the surface. All baselines usually cannot exert sufficient and stable force to scrape off the sticker.

Fig. 15: Contact Robustness of Force Policy under Disturbances in the Push and Flip Task. Force Policy is robust under human disturbances and can perform autonomous recovery during the contact phase.

Fig. 16: Objects in Generalization Evaluation. Unseen objects of different colors, geometries, and stiffnesses are selected to evaluate the generalization ability of the policies.

Fig. 17: Demonstration Overview for Each Task. We visualize the initial configurations from 50 demonstrations for each task.

Fig. 18: Evaluation Configurations for Each Task. These configurations are randomly initialized beforehand.

Policy	Flip	Charger	Scrape (Easy)	Scrape (Hard)
RISE-2	42.5 $\pm$ 21.8	0.0 $\pm$ 0.0	80.0 $\pm$ 30.2	10.0 $\pm$ 22.6
$\pi_{0.5}$	52.5 $\pm$ 17.8	0.0 $\pm$ 0.0	70.0 $\pm$ 34.6	20.0 $\pm$ 30.2
RDP	57.5 $\pm$ 19.0	5.0 $\pm$ 11.3	70.0 $\pm$ 34.6	20.0 $\pm$ 30.2
FoAR	60.0 $\pm$ 23.5	10.0 $\pm$ 15.1	40.0 $\pm$ 36.9	20.0 $\pm$ 30.2
ForceVLA	30.0 $\pm$ 19.2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0
TA-VLA	62.5 $\pm$ 22.6	0.0 $\pm$ 0.0	50.0 $\pm$ 37.7	10.0 $\pm$ 22.6

TABLE VI: Final-Stage Success Rates in Evaluations .

Contact-Rich Robotic Manipulation

Contact-rich robotic manipulation studies how robots plan and control physical interactions such as grasping, insertion, pushing, and dexterous object handling.

See all 399 papers

Learning Hybrid-Control Policies for High-Precision In-Contact Manipulation Under Uncertainty
Apr 21, 2026Hunter L. Brown, Geoffrey Hollinger, Stefan LeeRobotic Control Contact-Rich Robotic Manipulation
Collaborative Robotics and Intelligent Systems Institute, Oregon State University, Corvallis, OR
ForceVLA2: Unleashing Hybrid Force-Position Control with Force Awareness for Contact-Rich Manipulation
Mar 16, 2026Yang Li, Zhaxizhuoma, Hongru Jiang +11Contact-Rich Robotic Manipulation Force-Controlled Robotic Manipulation
Tongji University · Shanghai AI Laboratory · Lumos Robotics +6
ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching
May 11, 2026Shuoheng Zhang, Yifu Yuan, Hongyao Tang +7Contact-Rich Robotic Manipulation Robot Imitation Learning
Tianjin University · Shanghai AI Lab · Huawei Noah’s Ark Lab

Apr 21, 2026J/K move · Enter open · S save

Learning Hybrid-Control Policies for High-Precision In-Contact Manipulation Under Uncertainty

Hunter L. Brown, Geoffrey Hollinger, Stefan Lee

Collaborative Robotics and Intelligent Systems Institute, Oregon State University, Corvallis, OR

Robotic Control Contact-Rich Robotic Manipulation Robotic RL Robotic Insertion Force-Controlled Robotic Manipulation Robust RL

Open paper AskPDF arXiv

Reinforcement learning-based control policies have been frequently demonstrated to be more effective than analytical techniques for many manipulation tasks. Commonly, these methods learn neural control policies that predict end-effector pose changes directly from observed state information. For tasks like inserting delicate connectors which induce force constraints, pose-based policies have limited explicit control over force and rely on carefully tuned low-level controllers to avoid executing damaging actions. In this work, we present hybrid position-force control policies that learn to dynamically select when to use force or position control in each control dimension. To improve learning efficiency of these policies, we introduce Mode-Aware Training for Contact Handling (MATCH) which adjusts policy action probabilities to explicitly mirror the mode selection behavior in hybrid control. We validate MATCH's learned policy effectiveness using fragile peg-in-hole tasks under extreme localization uncertainty. We find MATCH substantially outperforms pose-control policies -- solving these tasks with up to 10% higher success rates and 5x fewer peg breaks than pose-only policies under common types of state estimation error. MATCH also demonstrates data efficiency equal to pose-control policies, despite learning in a larger and more complex action space. In over 1600 sim-to-real experiments, we find MATCH succeeds twice as often as pose policies in high noise settings (33% vs.~68%) and applies ~30% less force on average compared to variable impedance policies on a Franka FR3 in laboratory conditions.

Abstract

Contact-Rich Robotic Manipulation

Explore similar work

Learning Hybrid-Control Policies for High-Precision In-Contact Manipulation Under Uncertainty

ForceVLA2: Unleashing Hybrid Force-Position Control with Force Awareness for Contact-Rich Manipulation

ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching