FP2: Equipping Robotic Foundation Models with Force Control

Hongjie Fang; Shirun Tang; Junjian Hu; Shidong Zhang; Derek Zhang; Linhao Chen; Dehai Li; Mingyu Mei

cs.ROSep 29, 2026

FP2: Equipping Robotic Foundation Models with Force Control

Authors: Hongjie Fang, Shirun Tang, Junjian Hu, Shidong Zhang, Derek Zhang, Linhao Chen, Dehai Li, Mingyu Mei, +3 more

Organizations: FORTE Lab. · Noematrix. · SJTU. · UPenn. · FDU. · UIUC. · ZJU. · Flexiv. · SII. · B

Abstract

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

Figures & tables

Fig. 1: FP2 Framework. (Top left) The task-adapted RFM serves as a foundation policy for action generation, while FP2 compresses its contextual tokens and combines them with wrench and proprioceptive histories to predict structured force-control parameters. (Top right) The contextual token compressor is trained self-supervisedly through reconstruction; the reconstruction decoder is discarded after training. (Bottom) Training proceeds in three stages: (1) standard RFM task adaptation, (2) self-supervised contextual token compression with the RFM frozen, and (3) force-control policy training with both the RFM and compressor frozen.

Fig. 2: Contact-Rich Manipulation Tasks. We evaluate FP2 on 4 real-world tasks covering diverse physical interactions. Flip Box requires maintaining effective contact as the contact point and force direction vary during flipping; Insert EV Charger requires large insertion forces along the correct direction; Insert Peg demands precise alignment under a tight tolerance; and Wipe Curve requires sustained contact while following a curved surface. Red labels highlight the contact-rich phases, and the right panels illustrate representative challenges.

Policy	Flip Box		Insert EV Charger		Insert Peg		Wipe Curve		Average
Policy	SR $\uparrow$	NFE $\downarrow$	SR $\uparrow$	NFE $\downarrow$	SR $\uparrow$	NFE $\downarrow$	CR $\uparrow$	NFE $\downarrow$	Score $\uparrow$
HybridIL [ 39 ]	40% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	5.496	0% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	N/A	0% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	N/A	48% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	0.276	22% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$
ACP [ 23 ]	80% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	14.355	20% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	5.202	52% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	3.794	77% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	1.528	57% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$
Force Policy [ 17 ]	84% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	1.522	68% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	2.157	0% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	N/A	49% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	0.406	50% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$
GR00T N1.7 [ 4 ]	12% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	5.836	36% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	1.559	0% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	N/A	1% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$	N/A	12% ${}^{\text{\color[rgb]{0.66,0.66,0.66}}}$
GR00T N1.7 + FP2 (ours)	76% ${}^{\text{\color[rgb]{0.76,0.23,0.13}+64}}$	2.076	76% ${}^{\text{\color[rgb]{0.76,0.23,0.13}+40}}$	1.061	0% ${}^{\text{\color[rgb]{0.66,0.66,0.66}0}}$	N/A	3% ${}^{\text{\color[rgb]{0.76,0.23,0.13}+2}}$	N/A	39% ${}^{\text{\color[rgb]{0.76,0.23,0.13}+27}}$

TABLE I: Real-World Evaluation of Task Performance and Physical Interaction Quality. The average score is the arithmetic mean of the four task metrics (SR and CR). Small numbers denote absolute percentage-point changes relative to the corresponding task-adapted RFM. NFE is reported only for valid successful executions; “N/A” indicates that no such execution is available.

Fig. 3: Qualitative Analysis of Physical Interaction with and without FP2. Representative rollouts and measured force profiles are shown for (A) Flip Box , (B) Insert EV Charger , (C) Insert Peg , and (D) Wipe Curve . The foundation policy alone often fails due to lost contact, insufficient force, or unstable interaction, whereas FP2 maintains appropriate force regulation, enabling successful execution.

Fig. 4: Failure Analysis on Flip Box and Insert EV Charger . Each bar summarizes 25 trials and decomposes the outcomes into successful executions and different failure modes, grouped into position errors and interaction-related causes.

Force Control Policy $\pi_{F}$ Design		Latency	SR	SR @ Novel Object
Input	Output	Latency	SR	color	texture	stiffness	geometry	Average
(A) RFM Context and Physical Feedback
$\mathbf{z}$	$(\mathbf{\Sigma},\mathbf{S},\hat{\mathbf{W}})$	2.27ms	4%	0/10	2/10	0/10	0/10	5%
$(\mathbf{W},\mathbf{P})$	$(\mathbf{\Sigma},\mathbf{S},\hat{\mathbf{W}})$	4.28ms	36%	4/10	2/10	1/10	0/10	18%
(B) RFM Context Representation
$\text{AvgPool}({\mathbf{C}}),(\mathbf{W},\mathbf{P})$	$(\mathbf{\Sigma},\mathbf{S},\hat{\mathbf{W}})$	4.49ms	84%	10/10	8/10	7/10	5/10	75%

TABLE II: Ablation of FP2 Force Control Policy Design on the Flip Box Task with

\pi_{0.5}

. (Left) We compare different inputs (A) , RFM context representations (B) , wrist vision usage (C) , and action re-generation (C) , and report both training object and novel object success rates. Latency is measured per inference on an RTX 5090 GPU. (Right) Illustrations of the training object and novel objects. These novel objects differ from the training object in terms of color, texture, stiffness, and geometry, respectively.

Explore similar work

IMPACT: Learning Internal-Model Predictive Control for Forceful Robotic Manipulation
Jun 9, 2026Jiawei Gao, Chaoqi Liu, Peilin Wu +2Robotic Manipulation Impedance Control
Harvard University · Stanford University
FAWAM: Force-Aware World Action Models for Closed-Loop Contact-Rich Manipulation
Jun 7, 2026Haotian He, Zeyu Yan, Qipeng Liu +2Tactile World Model Efficient World-Action Model
School of Mathematical Sciences, Peking University · School of Artificial Intelligence, Shanghai Jiao Tong University.
Wrench-ACT: Enhancing Robot Policies for Contact Rich Behavior Using Direct Wrench Control
Sep 29, 2026Johannes Hechtl, Yannik Blei, Simon Ball +5Robotic Manipulation Policies Contact Wrench Estimation
Siemens Research and Predevelopment · Department of Computer Science and Artificial Intelligence, University of Technology Nuremberg, Germany

Jun 9, 2026, cs.ROJ/K move · Enter open · S save

IMPACT: Learning Internal-Model Predictive Control for Forceful Robotic Manipulation

Jiawei Gao, Chaoqi Liu, Peilin Wu, Haonan Chen+1

Harvard University · Stanford University

Robotic Manipulation Impedance Control End-Effector Pose Model Predictive Control Implicit Impact Analysis

Open paper AskPDF arXiv

Real-world robotic manipulation tasks often involve forceful interactions with the environment, such as using tools of varying weights, transporting objects with different masses, and performing contact-rich tasks like table wiping. Previous learning-based approaches typically employ imitation learning policies that output target end-effector poses tracked by low-level impedance controllers. In these systems, forceful interactions are either implicitly realized through steady-state tracking errors or explicitly commanded using wrist force/torque or tactile sensors. However, implicit approaches generalize poorly across object weights, while explicit approaches require specialized hardware and increase system complexity. In this work, we propose IMPACT, a framework that decouples these forceful tasks into task-planning and internal-model-based predictive control. Extensive simulation and real-world experiments demonstrate that the proposed framework achieves higher success rates and improved generalization to unseen object weights, as well as better safety and energy efficiency.

Abstract

Explore similar work

IMPACT: Learning Internal-Model Predictive Control for Forceful Robotic Manipulation

FAWAM: Force-Aware World Action Models for Closed-Loop Contact-Rich Manipulation

Wrench-ACT: Enhancing Robot Policies for Contact Rich Behavior Using Direct Wrench Control