CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation

Houzhu Ding; Xinda Qi; Luqi Wang; Ming Yang

cs.ROOct 21, 2026

CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation

Authors: Houzhu Ding, Xinda Qi, Luqi Wang, Ming Yang

Organizations: Ant Group

Abstract

Force-sensitive manipulation requires contact and load information that is not provided explicitly by visual observations. We present CW-WAM, a world action model that jointly predicts video, motion, and contact wrench. Inspired by hybrid motion-force control, it contains a unified interaction module that treats both motion and contact wrench as modeling targets and predicts time-aligned motion-wrench chunks as control references. A pair-causal attention mechanism structures the motion-wrench dependencies, while a long fixed-length wrench memory retains past interaction information. The predicted wrench chunk acts as a force reference for a 20-Hz outer-loop admittance controller within a slow-fast control system that maintains force feedback and smooth motion between model queries. CW-WAM improves contact prediction and increases the insertion success rate from 25% to 70% for square pegs with 1-mm clearance per side and from 40% to 95% for power adapters, compared with WAM without wrench input. It also improves lifting-height accuracy across different loads. Our framework with an admittance controller reduces force-reference RMSE from 4.8 to 2.7 N for peg insertion and from 2.0 to 1.3 N for wiping, respectively, with smaller force fluctuations during successful execution. Project website: https://cw-wam-project.github.io/

Figures & tables

Fig. 2: CW-WAM jointly predicts video, motion, and wrench from visual observations, language, robot state, and interaction history. Paired motion–wrench tokens capture recent interactions, while long-term wrench memory retains earlier contact information. Pair-causal attention links each future wrench to its paired action and earlier pairs. The model jointly denoises the three streams using separate prediction heads. Predicted motion and wrench trajectories provide paired references for robot control.

Fig. 3: Slow–fast motion–force control. Admittance corrects nominal base-frame poses before IK. Joint-trajectory updates preserve $C^{2}$ continuity. Wrench feedback is compensated; TCP-pose feedback optionally aligns wrench progress to measured TCP pose.

Task	Feedback weights	$M$	$C$	$K_{a}$	$k_{I}$	$\lambda$
Peg insertion	$s_{T}=\operatorname{diag}(1,1,0.05)$	5	75	200	0.5	0.5
	$s_{R}=\operatorname{diag}(1,1,0.2)$	0.1	1.5	4	0.5	0.5
Adapter insertion	$s_{T}=\operatorname{diag}(1,1,0.5)$	5	75	200	0.1	1
	$s_{R}=\operatorname{diag}(1,1,1)$	0.1	1.5	4	0.1	1
Whiteboard wiping	$s_{T}=\operatorname{diag}(0,0,1)$	3	34.64	100	0.1	1
	$s_{R}=\operatorname{diag}(0,0,0)$	–	–	–	–	–

TABLE I: Selected Admittance Gains

Fig. 4: Task outcomes and execution quality. Top: insertion or marker-removal success out of 20 trials; hatched bars show successful release. Quality rows compare all 20 trials (light, hatched) with successful subsets (dark bars); these groups coincide for wiping. A dash indicates no wrench prediction.

Model	$e_{p}$ [mm]	$e_{R}$ [ ∘ ]	$e_{F}$ [N]	$e_{I}$ [N s]
WAM	26.97	2.85	–	–
WAM + Wrench (padded)	26.32	2.76	3.15	2.33
WAM + Wrench (valid)	25.50	2.80	3.03	2.21
CW-WAM w/o LTM and PC	22.88	2.25	2.92	2.11
CW-WAM w/o LTM	19.92	2.02	2.84	2.17
CW-WAM	20.69	2.07	2.79	2.01

TABLE II: Open-loop prediction across four tasks. Padded and valid denote padded-width and valid-channel loss normalization. Bold: lowest error; –: no wrench output

Fig. 5: Contact-force curves across tasks. Colored curves show measured force magnitudes across successful trials. Thick blue and green curves show CW-WAM without and with admittance, respectively. The black curve and gray band show the demonstration reference. Curves are aligned by contact phase, with low-force intervals compressed. Adapter unloading includes only trials with successful release.

Fig. 6: Load-dependent lifting under light and heavy payloads. Curves show representative end-effector height trajectories. Dashed lines indicate approximate heights needed to clear the bin rim. The lower panels show peak heights from all ten trials per method and load. Diamonds mark the trials shown in the trajectories. Gray lines and bands show demonstration references.

Fig. 7: Force and torque tracking for peg insertion (left) and whiteboard wiping (right). (a–f) Measured and reference magnitudes, with three-axis force RMSE labels. Each example is selected from the trial whose force variation is closest to the group median. (g,h) Force variation (solid bars, left axis) and torque variation (hatched bars, right axis). Aggregate bars summarize 10 trials per condition.

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 410 papers

FAWAM: Force-Aware World Action Models for Closed-Loop Contact-Rich Manipulation
Jun 7, 2026Haotian He, Zeyu Yan, Qipeng Liu +2Robotic Control World Models for Robotics
School of Mathematical Sciences, Peking University · School of Artificial Intelligence, Shanghai Jiao Tong University.
ForceWAM: Decoupling Future and Response Learning from Demo and Play Data for Contact-Rich Manipulation
Oct 13, 2026Zijun Xu, Wendi Chen, Xinkai Wang +16Contact-Rich Robotic Manipulation Robot Learning
Fudan University · Shanghai Innovation Institute · Shanghai Jiao Tong University +3
TacWAM: Anchor-Guided World Action Model with Mechanics-Aware Tactile Prediction
Jul 30, 2026Lei Jin, Yiding Ma, Xin Zhang +3World Models for Robotics Contact-Rich Robotic Manipulation
1Tsinghua University · 2Manifold AI

Jun 7, 2026J/K move · Enter open · S save

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

Haotian He, Zeyu Yan, Qipeng Liu, Ning Guo+1

School of Mathematical Sciences, Peking University · School of Artificial Intelligence, Shanghai Jiao Tong University.

Robotic Control World Models for Robotics Contact-Rich Robotic Manipulation Robotic Manipulation Force-Controlled Robotic Manipulation

Open paper AskPDF arXiv

Force signals provide critical interaction cues for contact-rich robotic manipulation. However, existing methods mostly use force as an additional observation modality, without fully exploiting its role in modeling future interaction dynamics or guiding execution-time feedback correction. In this paper, we propose FAWAM, a force-aware world action model that incorporates force information at three levels: perception, prediction, and closed-loop execution. FAWAM first encodes historical 6-axis force/torque signals to modulate action generation, then jointly predicts future actions and end-effector wrenches to explicitly model contact evolution. It further introduces a residual correction module that uses the predicted wrench trajectory as an execution-time reference to refine actions online based on real-time force feedback. Real-world experiments across multiple contact-rich tasks show that FAWAM improves the average success rate by 36.25% over vision-only baselines and 21.25% over existing force-aware baselines, demonstrating the effectiveness of our force-aware framework for robust contact-rich manipulation.

Abstract

Contact-Rich Robotic Manipulation

Explore similar work

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

ForceWAM: Decoupling Future and Response Learning from Demo and Play Data for Contact-Rich Manipulation

TacWAM: Anchor-Guided World Action Model with Mechanics-Aware Tactile Prediction