acceptodds
Under review as a conference paper at ICLR 2027

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

34 views

Abstract

Expert demonstrations show how to complete contact-rich tasks, but provide limited coverage of how interaction forces respond to different actions. Task-agnostic play provides broader experience of contact dynamics without specifying task goals. To fully leverage these two types of data, we introduce ForceWAM, a force world action model that decouples knowledge of task-compatible interaction outcomes from knowledge of how local actions change contact. A world model learns from successful demonstrations to jointly predict future video and six-axis force/torque trajectories. Together, these futures describe the expected evolution of motion and contact during the task. A separate inverse dynamics model (IDM) learns from both task and play, mapping current observations and recorded visual-wrench futures to their associated robot actions. Play data thus provide knowledge of force responses for goal-directed execution. At deployment, the IDM reactively translates generated futures into actions, reusing them with fresh wrench feedback between world-model updates. Across real-robot force-regulated sliding, box flipping with and without disturbances, and ring placement, ForceWAM achieves – mean success, surpassing world-modeling and reactive baselines. Our experiments further show that adding task-agnostic play data improves success by – percentage points, demonstrating its value for learning contact-rich execution.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Contact-rich Manipulation, Robotics and Control

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Qy1kOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

ForceWAM presents a framework that leverages world-model predictions to guide robot action generation in contact-rich manipulation tasks. Incorporating predicted force information provides additional physical cues for action generation, leading to substantial improvements in task success rates. The effectiveness and robustness of the framework are further demonstrated through real-world experiments. However, I find the methodological novelty relatively incremental, and the experiments do not sufficiently establish the specific advantages of the proposed architectural design. Furthermore, the experimental evaluation is limited to three task families using a self-collected dataset, raising concerns about generalization. There are also several minor presentation issues that make the technical details unnecessarily difficult to follow. Overall, I would consider increasing my score if the authors could more convincingly demonstrate that the contribution extends beyond incorporating force information into an existing world-model-based action generation framework.

Critical strengths
  1. Well-motivated framework for contact-rich manipulation. The integration of force information into action generation is important for improving the physical understanding and interaction capabilities of robotic policies. Using successful task demonstrations for future prediction while incorporating task-agnostic play data into inverse dynamics learning is intuitive and practically meaningful.

  2. Promising experimental performance. The experimental comparisons, while not entirely comprehensive, cover several representative baselines. ForceWAM demonstrates substantial improvements in success rates across the three evaluated task families, providing encouraging evidence of its effectiveness in contact-rich manipulation.

  3. Real-world experimental validation. I particularly appreciate that all experimental evaluations are conducted on physical robots, with supplementary videos illustrating the execution process. These demonstrations provide useful evidence of the practical feasibility of ForceWAM and help readers understand how the framework operates in contact-rich scenarios.

Critical weaknesses
  1. Presentation and mathematical notation.
  • In Section 3.1, when introducing the observation (o_t), I suggest explicitly defining its constituent variables and their dimensions to improve mathematical clarity and rigor.

  • Line 152: It would be clearer to explicitly state the meaning of the first factor in Equation (1), rather than requiring readers to identify the corresponding component from the equation.

  • Line 241: Since the observation has already been defined mathematically, I suggest reusing the established notation for consistency and simplicity.

  • Line 242: The physical meanings of (C_t^{\mathrm{slow}}) and (C_{t,k}^{\mathrm{fast}}) are not immediately clear. I recommend introducing these variables together with their corresponding definitions and explaining what information each represents.

  • The connection between Figure 2 and Section 3.4 is not sufficiently explicit. Referring directly to the relevant components of Figure 2 when explaining the slow-fast execution mechanism would make the methodology easier to follow.

  • Line 267: I suggest explicitly defining (\mathbf{u}_{s+k}) and clarifying its relationship to the predicted action and the command sent to the robot controller.

  1. Limited methodological novelty.
  • My primary concern is the novelty of the proposed framework. As I understand it, ForceWAM first employs a world model to predict future observations and then uses a diffusion-based inverse dynamics model to generate actions conditioned on those predictions. Similar formulations have already been explored in previous works, such as Video Prediction Policy (ICML 2025) and Video Generators are Robot Policies (arXiv:2508.00795). Although ForceWAM additionally incorporates force information, force-aware action generation has also been investigated in recent work, such as Force Policy (arXiv:2602.22088). Therefore, I am not fully convinced that the proposed framework introduces a sufficiently substantial methodological contribution beyond combining existing world-model-based action generation with force-aware policy learning. I would appreciate a more explicit discussion and comparison to establish its novelty.

  • Another concern is the selection of baselines. The authors primarily compare against existing world-action models and several force-aware policies, but do not include certain recent methods that directly incorporate force measurements into action generation without relying on a world model, such as Force Policy. A comparison with this category of methods would help demonstrate whether the proposed combination of world prediction and force information provides advantages beyond force-conditioned action generation alone.

  • The authors also motivate their method by arguing that conventional end-to-end policies entangle the learning of task-compatible outcomes and action responses. However, actions and world evolution are inherently coupled: actions change the environment, while changes in the environment influence subsequent actions. I am therefore interested in understanding the motivation for explicitly decoupling these learning processes. In particular, how does the proposed architecture handle the temporal dependencies and potential asynchrony between world evolution and action generation during execution? A clearer explanation of why this decomposition is beneficial would strengthen the methodological contribution.

  1. Insufficient experimental validation and ablation studies.
  • I did not find an ablation study that directly demonstrates the advantages of separating world-model prediction from action generation. In principle, these components could be integrated into a single model that directly predicts future actions while jointly learning world dynamics. A controlled comparison between the proposed decoupled architecture and a jointly trained alternative would provide stronger evidence that the decomposition itself contributes to the reported performance improvements.

  • The experimental evaluation is relatively limited in task diversity. The authors evaluate ForceWAM on a self-collected dataset covering three task families and four conditions, without demonstrating its effectiveness on additional contact-rich manipulation benchmarks. For example, the publicly available FAWAM dataset includes other manipulation tasks, such as fruit peeling and board cleaning. Evaluating ForceWAM on such tasks, where compatible observation and control interfaces are available, would help establish whether the proposed framework generalizes beyond the specific scenarios considered in this paper.

Questions for authors

Please refer to the concerns raised in the Weaknesses section, particularly regarding the methodological distinction from existing world-model-based action generation frameworks, the motivation for decoupling world prediction and inverse dynamics learning, and the experimental evidence supporting the benefits of this architectural design.

7 comments
AuthorsOpenReview ↗

Response to Reviewer Qy1k (1/5)

Thank you for the constructive review, and for saying clearly what would change your assessment. We respond to each point below. We have uploaded a revised PDF; changes are marked in blue and the new experiments are in App. G.

W1: presentation and notation. We revised all six points in the paper.

  • Sec. 3.1, definition of oto _ t. The components and dimensions are now defined together before Eq. 1: ot=(It,wt)o _ t = (I _ t, w _ t), where It∈RH×W×3I _ t \in \mathbb{R}^{H\times W\times 3} is the wrist RGB image and wt=(ft,τt)∈R6w _ t = (f _ t, \tau _ t) \in \mathbb{R}^6 is the TCP force (N) and torque (N·m) measured by the end-effector F/T sensor. The future reference gt=(zt,Wt)g _ t = (z _ t, W _ t) consists of a video latent zt∈R16×56×56z _ t \in \mathbb{R}^{16\times 56\times 56} of future wrist frames and the normalized wrench trajectory Wt∈R16×6W _ t \in \mathbb{R}^{16\times 6} over the next 16 steps. At=(at,0,…,at,Ha−1)A _ t = (a _ {t,0}, \dots, a _ {t,H _ a-1}) with Ha=18H _ a = 18, of which the first six actions are executed before the reference is refreshed.
  • Eq. 1, the two factors. The text now names them: "The first factor, pϕ(gt∣ot,ℓ)p _ \phi(g _ t \mid o _ t, \ell), is the task-conditioned future generator ForceWM, which specifies what a successful interaction should look and feel like. The second factor is the inverse dynamics model (IDM), which we approximate by pθ(At∣ot,gt)p _ \theta(A _ t \mid o _ t, g _ t): given the target future, the action is assumed to be conditionally independent of the instruction ℓ\ell. This assumption is what allows task-agnostic play to train the IDM."
  • Reusing the notation. oto _ t is now used throughout: pϕ(gt∣ot,ℓ)p _ \phi(g _ t \mid o _ t, \ell) in Sec. 3.2, the condition of the IDM loss in Sec. 3.3, Eslow(ot,g~t)E _ {\mathrm{slow}}(o _ t, \tilde g _ t) in Sec. 3.4, and pϕ(gr∣or,ℓ)p _ \phi(g _ r \mid o _ r, \ell) and Eslow(or,g^r)E _ {\mathrm{slow}}(o _ r, \hat g _ r) in Eq. 4.
  • CslowC^{\mathrm{slow}} and CfastC^{\mathrm{fast}}. Both are defined where they first appear: "The slow tokens Ctslow=Eslow(ot,g~t)C^{\mathrm{slow}} _ t = E _ {\mathrm{slow}}(o _ t, \tilde g _ t) encode what should happen: the current observation and the future reference, computed once per reference refresh. The fast tokens Ct,kfast=Efast(wt:t+k)C^{\mathrm{fast}} _ {t,k} = E _ {\mathrm{fast}}(w _ {t:t+k}) encode what has happened so far: the wrench prefix measured up to action offset kk, updated once per action."
  • Fig. 2 and Sec. 3.4. The caption now labels panels (a)–(c), and Sec. 3.4 refers to the components of Fig. 2(c) by name. The "Slow reference refresh" branch produces g^r\hat g _ r, which is cached ("Cached reference"). In the "Fast action updates" branch, the "Measured wrench feedback" yields the fast tokens, and only the "Current token" is decoded and executed.
  • us+ku _ {s+k}. It is now defined after Eq. 5: "The command us+k=(Ts+kcmd,γs+k)u _ {s+k} = (T^{\mathrm{cmd}} _ {s+k}, \gamma _ {s+k}) sent to the low-level Cartesian impedance controller at control step s+ks+k consists of an absolute TCP-pose target and a gripper command. Ts+kcmdT^{\mathrm{cmd}} _ {s+k} composes the relative pose in a^s,k\hat a _ {s,k} with the cached base pose TrT _ r; the force-derived auxiliary targets in a^s,k\hat a _ {s,k} are used only for training and discarded at execution (App. C)."
AuthorsOpenReview ↗

Response to Reviewer Qy1k (2/5)

W2.1: novelty.

Concern: Predicting the future with a world model and generating actions with a diffusion-based IDM has been explored (VPP, Video Policy), and force-aware action generation also exists (Force Policy), so the contribution beyond combining the two is unclear.

ForceWAM is more than a world-model policy with an added force input. Its contribution is a way to use play data for contact-rich manipulation, which covers the following three aspects.

  1. Play data matters for contact-rich manipulation. Demonstrations stay near a nominal contact trajectory (about 10 N in Slide), and play adds harder and lighter pressing, slipping and loss of contact. With demonstrations and ForceWM fixed, play raises success by 15–45 pp, and success rises with the amount of play (Table 1, Fig. 4).
  2. A visual-wrench future as the interface makes play usable. We treat contact execution as inverse dynamics conditioned on a visual-wrench future. An IDM conditioned on the recorded future can be supervised by any recorded interaction, so task-agnostic play trains it while ForceWM learns from demonstrations only. Mixing the same play into a policy or a joint world-action model hurts (+Play rows of Table 1; Joint-WAM in Post 5). The wrench part is needed because a visual future does not determine force. On Slide, a visual-only IDM reaches 15% and a wrench-only IDM 20%, against 85% with both (Table 2).
  3. A reactive IDM designed for the force signal. Contact force changes within an action chunk and cannot be read from the image, so the IDM is causal in the measured wrench: action token kk sees measured wrench 0:kk, and the reference wrench only at offsets larger than kk. This gives within-chunk reaction with no intervention data. Without it, success drops from 85/95/85 to 40/50/40 (App. G.1).

Table R1. Differences from VPP, Video Policy and Force Policy

Prior workWhat it providesWhat differs in ForceWAMWhy it matters
VPP, Video PolicyAction module conditioned on a predicted visual future, trained on demonstrationsA visual-wrench future, and an IDM that also learns from playSliding at 5 N looks the same as sliding at 15 N, so a visual future leaves force open
Force PolicyInteraction frame, per-axis force or position control, and a reference wrench, executed by a hybrid force-position controllerA predicted visual-wrench future as the condition of the action modelIt does not predict how the interaction should evolve, and it learns from demonstrations only

The revised related work (Sec. 2) now discusses VPP and Video Policy.

Three results show that the gain does not come from the force input alone.

  1. A force input is not enough. The IDM with live wrench feedback and no future condition reaches 15/30/35% (Table 2), and force-reactive baselines without a world model reach at most 60% (Table 1).
  2. Adding force to a jointly trained WAM is not enough. On the same backbone and data, a model that jointly predicts video, wrench and actions averages 28% (Table R2, Post 5).
  3. Play helps through the decoupled IDM. It lowers Joint-WAM by 5 pp and raises ForceWAM by 28 pp (Table R2).
AuthorsOpenReview ↗

Response to Reviewer Qy1k (3/5)

W2.2: choice of force-conditioned baselines.

Concern: The baselines do not include recent methods that use force measurements directly in action generation without a world model, such as Force Policy.

The category you describe, methods that feed force measurements directly into action generation without a world model, is represented in Table 1 by RDP, ImplicitRDP and ForceVLA. Force Policy itself uses RDP and ForceVLA as baselines. We did not include Force Policy because its setting differs from ours in three respects.

  • Control interface. Force Policy's outputs are executed by a hybrid force-position controller. ForceWAM, like every method in Table 1, sends TCP-pose targets to the same Cartesian impedance controller and regulates force indirectly through pose targets.
  • Supervision. Force Policy needs per-segment interaction-frame and control-mode labels, obtained from demonstrations by a power-based analysis and a vision-language model. Our data has no such labels.
  • Perception. Its global policy relies on 3D perception from a global RGB-D camera, and our models use a wrist RGB image.

Running Force Policy on our platform would change the controller, the labels and the perception at once, so the comparison would not isolate the effect of the predicted future. The force-conditioned baselines in Table 1 share our controller and inputs. They reach at most 60%, and ForceWAM reaches 85–95%. Force Policy is complementary to our method, and its interaction-frame action space could serve as the output space of our IDM.

W2.3: actions and world evolution are coupled, so why decouple?

Concern: Actions change the environment and the environment influences later actions. Why decouple the two learning processes, and how are temporal dependencies and asynchrony handled during execution?

We agree that they are coupled, and ForceWAM models the coupling in two places. What we decouple is the source of supervision.

  1. World evolution and contact are predicted jointly. ForceWM generates future video and future wrench in the same denoising pass over the same time window, so the visual and contact futures are synchronized descriptions of one interaction.
  2. The IDM maps this joint future to actions. pθ(At∣ot,gt)p _ \theta(A _ t \mid o _ t, g _ t) outputs the actions that realize the predicted visual-wrench evolution from the current state, and it closes the loop on the measured wrench.

Inverse dynamics suits contact for a physical reason. Compliant contact can be modeled as a spring: in the quasi-static case, contact force is proportional to the offset between the target and the actual pose, f=K(xvt−xreal)f = K(x _ {\mathrm{vt}} - x _ {\mathrm{real}}) (Chen et al., 2026, Eqs. 3–4). Under the Cartesian impedance controller we use, the same relation links the commanded TCP target to the contact force it produces, f≈K(xcmd−x)f \approx K(x _ {\mathrm{cmd}} - x). The action is then determined by where the robot should be and how much force it should apply, and these are given by the visual future and the wrench future. Both futures are needed for this reason (Table 2, Slide: visual only 15%, wrench only 20%, both 85%). The mapping is set by the robot, the controller and the contact. It does not depend on the task instruction, so learning it needs no task labels.

Reviewer_Dg8MOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ForceWAM for contact-rich robotic manipulation. A world model learns task-compatible visual–wrench futures from successful demonstrations, while an inverse dynamics model (IDM) learns to realize such futures using both demonstrations and task-agnostic play. During execution, the IDM uses measured wrench feedback between reference updates. Experiments cover three real-robot task families and four evaluation conditions. The main contribution is a visual–wrench interface that separates learning task outcomes from learning local action–contact responses, allowing play data to improve execution without training the future generator on off-task behavior.

Critical strengths
  1. The separation between task-outcome learning and contact-response learning is well motivated and provides a useful way to incorporate play data.
  2. Play improves success by 15–45 percentage points with task demonstrations, ForceWM, and IDM optimizer updates held fixed. The paper also examines increasing play-data amounts.
  3. Real-robot results are strong across the four conditions. Ablations support the complementary value of visual and wrench futures beyond current observations and measured wrench feedback.
Critical weaknesses
  1. The methodological advance is incremental. Video planning with a separately trained IDM and self-play already appears in VERA, while future-wrench prediction and reactive force correction appear in FA-WAM. ForceWAM makes a useful extension by connecting these ideas through a visual–wrench future interface and separating demonstration-based outcome learning from response learning with play. Nevertheless, its contribution is primarily an effective combination and adaptation of established approaches.
  2. The scope of play-data reuse remains unclear. Play is collected within each evaluated task family, although without an evaluation-task instruction. The shared multi-task IDM and play-data scaling results demonstrate utility within these families, but do not establish whether a trained IDM transfers to new tasks or whether play from other task families benefits a target task without its own play data. It therefore remains unclear how much of the benefit reflects reusable contact knowledge versus experience closely related to the target task. This distinction matters for understanding the broader value of task-agnostic play.
  3. The effect of prediction errors on execution is insufficiently characterized. Figure 5 and Appendix F show predictions from the same box-flipping rollout, but do not connect prediction discrepancies to downstream performance. Future-conditioning augmentation, including temporal perturbations, improves success by 15–20 percentage points, and Flip-D provides evidence of disturbance tolerance. These are useful results, but they do not explain which inaccuracies the reactive IDM can compensate for and when an inaccurate future reference causes execution to fail.
Questions for authors
  1. Appendix E reports sequential ForceWM and IDM calls of approximately 100 ms and 60 ms at reference refreshes. The previous controller target remains active, and the first new action uses the request-time wrench. How does this refresh process affect the continuity of action commands and contact forces, particularly when contact changes during computation? Reusing noise within a window reduces sampling variation, but continuity across reference updates is not discussed.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Dg8M (1/4)

Thank you for the positive assessment and the careful questions. We respond to each point below. We have uploaded a revised PDF; changes are marked in blue and the new experiments are in App. G.

W1: novelty.

Concern: The method is seen as an effective combination and adaptation of established approaches, in particular VERA and FA-WAM.

ForceWAM's contribution is a way to make task-agnostic contact play improve contact-rich execution. This had not been achieved for force-centric manipulation. Existing methods (RDP, ImplicitRDP, ForceVLA, FA-WAM) learn from task demonstrations, and adding play to them hurts: ImplicitRDP drops from 20/45/25 to 10/25/10 (Table 1). With ForceWAM, the same play raises success by 15–45 pp.

Putting VERA and FA-WAM together does not give this result. We tested both halves.

  1. A self-play IDM does not carry over to contact. VERA's IDM inverts a visual future, which works when motion determines the action. In contact the robot barely moves while the force changes a lot, so the same visual future is consistent with pressing at 5 N or at 15 N, and the inverse problem has no unique answer. Trained on our tasks with the same play, VERA reaches 10/25/20 (App. G.2), and our own IDM with a visual-only future reaches 15/40/30 (Table 2). Play starts to help once the wrench future is part of the condition. Under impedance control the commanded target and the contact force are tied by f≈K(xcmd−x)f \approx K(x_{\mathrm{cmd}} - x), so motion and force together determine the action. Neither future is enough alone (Slide: visual 15%, wrench 20%, both 85%).

  2. The way FA-WAM uses wrench leaves no room for play. FA-WAM predicts wrench together with actions in one model trained on task data, and the predicted wrench serves as the target of a residual corrector trained on human interventions and policy rollouts. Play has no valid task target, so it cannot enter a joint model. In our jointly trained variant, adding play lowers success from 20/40/25 to 15/35/20 (App. G.1). In ForceWAM the predicted wrench is the condition of the action model, so a recorded play segment can supervise it. Reaction comes from the IDM itself: action token kk sees the measured wrench up to kk and the reference wrench only after kk, so it acts on the difference between what has happened and what should happen next. This needs offline data only.

The parts also depend on each other, which independent modules would not. Starting from the decoupled IDM trained on task data without feedback (33% on average), play alone adds 10 pp and within-chunk feedback alone adds 27 pp, and the two together add 55 pp (App. G.1).

AuthorsOpenReview ↗

Response to Reviewer Dg8M (2/4)

W2: the scope of the benefit of task-agnostic play, and whether the trained IDM generalizes beyond its training conditions.

Concern: Play is collected within each evaluated task family. The results show that it is useful within these families, but not whether a trained IDM transfers to new tasks or whether play from other families helps a target task, so the scope of the benefit of task-agnostic play is unclear.

In this paper, task-agnostic means that play is collected without a downstream task instruction and without requiring task success. It does not mean that the data is independent of the objects, environment or task family. As Sec. 4.1 states, play is collected in the environment of each evaluated task family and consists of exploratory pressing, sliding, and making and breaking contact.

We use contact-rich play to cover a wider range of contact states than demonstrations do, which makes the IDM a stronger executor. Demonstrations stay near the nominal contact trajectory. Play adds harder and lighter pressing, slipping, and making and breaking contact, and it is cheap because it needs no task label or success check. Three results support this. With demonstrations and ForceWM fixed, play raises success from 40/75/65 to 85/95/85. Success rises steadily with the amount of play (Fig. 4). At an equal budget of 200 episodes, replacing half of the demonstrations with play raises success from 40/75/65 to 65/85/75 (App. G.2).

The three task families we evaluate differ strongly in visual geometry, contact phases and force response: holding a normal force while sliding, pushing a box over its edge, and seating a ring on a pole. Sec. 4.3 validates the benefit of play within these families. Transfer of contact knowledge across task families and the effect of scaling up play are left to future work. The focus of this paper is how to use play for contact-rich execution: play covers the contact states that demonstrations miss, the visual-wrench interface lets a separately trained IDM learn from it, and the reactive IDM uses the measured force within each chunk to carry out the predicted interaction.

On whether the trained IDM generalizes beyond its training conditions: we ran it, with no new data, under contact conditions that appear in none of the training data (App. G.4). This tests new contact conditions within a task family. Flip-D in the submission is already such a test under strong random disturbance, where the box is pushed by hand from first contact to the end of the task, with offsets of 0–45°.

Table R1. Unseen contact conditions (ForceWAM; success %, 20 trials each)

ConditionUnseen contact conditionNominal (Table 1)Unseen
Slidecompliant pad under the force scale8570
Flipunseen box with different appearance and weight9585
Flip-D (Table 1)random hand pushes, 0–45°, throughout contact95 (Flip)85

ForceWAM keeps 70% on Slide with a compliant pad and 85% on Flip with an unseen box of different appearance and weight. Under random pushes, Flip-D reaches 85%, and the strongest baseline reaches 40% (Table 1).

AuthorsOpenReview ↗

Response to Reviewer Dg8M (3/4)

W3: how prediction errors affect execution.

Concern: The paper does not connect prediction discrepancies to downstream performance, or explain which inaccuracies the reactive IDM can compensate for and when an inaccurate reference causes failure.

We first quantify the prediction error. Closed-loop rollouts are not expert demonstrations, so they provide no ground-truth label for the future state. We therefore evaluate ForceWM offline on the validation set of our dataset, with one ForceWM sample per window (App. G.5). Errors are in physical units. On Slide, the force RMSE is 1.8 N, against a normal-force standard deviation of 1.5 N in the human demonstrations, so the prediction error of ForceWM is comparable to the variation in the human demonstration data.

Table R2. Prediction quality on the validation set

FamilyForce RMSE (N)Torque RMSE (N·m)PSNR ↑SSIM ↑LPIPS ↓
Slide1.80.0621.30.690.27
Flip2.40.0719.80.640.31

To relate these errors to downstream performance, we compare ForceWAM without play and with play. The two systems use the same ForceWM, and only the IDM differs. Success rises from 40/75/65 to 85/95/85. The predicted futures are the same in both systems, so this gain comes from play improving how the IDM executes them. Most failures of the task-only system are execution failures, and play removes them. This is the contribution of the paper.

Which inaccuracies the IDM can compensate for. During training, the future condition is perturbed (App. D, Table 3): scale jitter in [0.85, 1.15], additive noise (std 0.075 on the normalized wrench), dropout of the visual or wrench part, and temporal offsets of up to 12 steps. The IDM has therefore seen amplitude errors, noise, a missing modality and a reference that lags the actual contact. Without this augmentation, success drops from 85/95/85 to 65/80/70 (Table 2). Within-chunk feedback then corrects the remaining gap using the measured wrench; removing it while keeping the same future condition lowers success from 85/95/85 to 40/50/40 (App. G.1). In Flip-D the box is pushed throughout contact, so the actual contact keeps departing from the reference, and ForceWAM still reaches 85% (Table 1).

When an inaccurate reference causes failure. The failures that remain in the full system come from both ForceWM prediction error and IDM error. In the failure cases we observed, the phase matters. In Flip, for example, a large prediction error during the contact phase (in the predicted wrench, the predicted video, or both) leads to failure, while errors before contact or near completion usually do not.

AuthorsOpenReview ↗

Response to Reviewer Dg8M (4/4)

Q1: continuity at a reference refresh.

Concern: At a refresh, ForceWM and the IDM run in sequence (about 100 ms and 60 ms) while the previous target stays active. How does this affect the continuity of action commands and contact forces, especially when contact changes during computation?

Within a chunk, actions follow the measured contact. Between refreshes the IDM re-denoises from cached noise with deterministic DDIM, so only the measured wrench changes its output, and each action is conditioned on the latest measurement. Removing this within-chunk feedback while keeping the same reference lowers success from 85/95/85 to 40/50/40 (App. G.1). Without feedback, the actions of a chunk are fixed at the refresh and cannot follow contact changes. With feedback, the commands stay consistent with the actual contact state throughout the chunk.

Across chunks, execution is synchronous. ForceWM and the IDM run in sequence at a refresh, and no other plan runs in parallel that could conflict with the new one.

  1. Contact force is held during the refresh. The command is a TCP-pose target tracked by a Cartesian impedance controller. Holding the previous target holds the commanded indentation, so during the extra about 60 ms (a 160 ms period instead of 100 ms) the controller keeps applying force and does not release contact.
  2. Every chunk is re-anchored to the current state. ForceWM generates the new reference from the current image and wrench, and the IDM's relative actions are decoded with the TCP pose recorded at request time (App. C). The first action after a refresh is a small step from where the robot currently is. Consecutive references also overlap in time, since each covers 16 steps and is replaced after 6.
  3. Contact changes during computation are corrected within one step. The first action after a refresh uses the request-time wrench, which is at most about 160 ms old. From the next action on, the IDM conditions on the latest measurement. Training uses time-shifted references (App. D, Table 3) so that the IDM tolerates a reference that lags the current contact state.

We also checked this on the existing evaluation logs. In the contact phase, we compare the first step after a refresh with the other steps within a window.

Table R3. Continuity at refresh steps vs within-window steps

Commanded translation increment (mm)Commanded rotation increment (°)Force change per step (N)Force change per 100 ms (N)
First step after a refresh2.10.450.80.50
Other steps within a window1.80.400.60.60

Command increments at refresh steps are close to those within a window. The per-step force change is slightly larger at a refresh (0.8 vs 0.6 N). Normalized to 100 ms it is not larger (0.50 vs 0.60 N), so the difference comes from the longer step.

Flip-D covers the case you describe, since pushes continue throughout the contact phase and also fall inside refresh computation. ForceWAM reaches 85% on Flip-D, and the strongest baseline reaches 40% (Table 1).

Thank you again for your time and detailed feedback. We are happy to discuss any further questions.

Reviewer_DMBbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ForceWAM separates contact-rich manipulation into a task-conditioned model that generates future video and six-axis wrench trajectories, and an inverse dynamics model (IDM) that translates these futures into actions. The future generator learns from successful demonstrations, while the IDM additionally learns from task-agnostic play and incorporates fresh force feedback during execution. Experiments cover sliding, box flipping with and without disturbances, and ring placement.

Critical strengths
  1. Well-motivated use of complementary data. The distinction between learning successful task outcomes and learning local contact responses is convincing. Conditioning the IDM on recorded outcomes provides a sensible way to learn from play that does not accomplish the downstream task.

  2. Encouraging real-robot results. The method consistently outperforms the evaluated baselines, including with a task-only IDM. The play-scaling experiment fixes the future generator and optimizer updates, providing useful evidence that additional play improves execution.

  3. Useful multimodal ablations and clear presentation. Section 4.5 supports the complementary value of visual and wrench futures. Figure 2 clearly explains the training–deployment interface, and Section 3.4 carefully distinguishes predicted references from causally available force measurements. The connection between video prediction and physical interaction is relevant to embodied vision.

Critical weaknesses
  1. The methodological advance is primarily an integration of established ideas. UniPi already separates video planning from action inference [1], and VERA explicitly trains an independent IDM using task-agnostic self-play [2]. ImplicitRDP provides causal force feedback and cached-noise diffusion execution [3], while FAWAM uses predicted wrench trajectories to guide online correction [4]. These precedents are acknowledged, but the contribution should be framed more narrowly around the joint visual–wrench interface and its use for demonstration/play learning. arxiv.org

  2. The play comparisons do not establish an advantage over stronger ways of using play. Section 4.1 adds exploratory actions directly to the behavioral supervision of ImplicitRDP and Cosmos Policy. Their deterioration is consistent with conflicting task and exploration targets; it does not demonstrate that competing architectures cannot benefit from play through goal relabeling or pretraining followed by task fine-tuning. An equal-budget comparison against additional successful demonstrations would also help distinguish broader contact coverage from simply more interaction data.

  3. Generalization and statistical reliability remain insufficiently characterized. The four evaluation conditions comprise three task families, with play collected separately for each family. No explicit held-out-object, material, or cross-task-play evaluation establishes transferable contact knowledge. The reported means lack confidence intervals and clearly specified independent training runs. The large performance margins are promising, but robustness across seeds and play subsets remains uncertain.

  4. Some mechanism claims are only partially isolated. All future-conditioning ablations retain live wrench feedback, so they do not quantify the contribution of within-chunk reactivity. Moreover, removing all futures also removes the IDM’s explicit task-goal channel because it receives no language instruction. Figure 5 illustrates predictions from one rollout, but does not establish prediction quality across tasks or explain failures when generated futures differ substantially from recorded training futures.

Questions for authors
  1. Add a stronger play-learning baseline using outcome/goal conditioning or play pretraining followed by demonstration fine-tuning. Describe baseline tuning and training budgets sufficiently for reproduction.

  2. Clarify how FA-WAM’s residual corrector was trained under the stated task-only protocol. Its original method requires intervention and policy-rollout supervision [4], making this an important comparison detail. arxiv.org

  3. Ablate fresh within-chunk force feedback while retaining the same future references. Give the no-future baseline an explicit task condition to separate goal information from future prediction.

  4. Report repeated-run uncertainty, quantitative force-regulation and recovery metrics, and performance under controlled changes in contact properties. Specify disturbance magnitude/timing and whether the sliding force criterion applies continuously or through an aggregate measure.

  5. Quantify future-prediction errors on held-out episodes and relate them to execution failures. Include effective control rates and observation-to-command latency in the main experimental description.

[1] Du et al. Learning Universal Policies via Text-Guided Video Generation. NeurIPS, 2023.

[2] Li et al. Turning Video Models into Generalist Robot Policies. arXiv, 2026.

[3] Chen et al. ImplicitRDP: An End-to-End Visual-Force Diffusion Policy with Structural Slow-Fast Learning. arXiv, 2025.

[4] He et al. FAWAM: Force-Aware World Action Models for Closed-Loop Contact-Rich Manipulation. arXiv, 2026.

8 comments
AuthorsOpenReview ↗

Response to Reviewer DMBb (1/6)

Thank you for the detailed review. We respond to each point below. We have uploaded a revised PDF; changes are marked in blue and the new experiments are in App. G.

W1: the contribution and its relation to prior work.

Concern: The method is seen mainly as an integration of established ideas (UniPi, VERA, ImplicitRDP, FA-WAM), and you suggest framing the contribution around the visual-wrench interface and its use for learning from demonstrations and play.

ForceWAM is related to the four works you list. Its contribution is a way to use play data for contact-rich manipulation, which covers the following three aspects.

  1. Play data matters for contact-rich manipulation. Demonstrations stay near a nominal contact trajectory (about 10 N in Slide), and play adds harder and lighter pressing, slipping and loss of contact. With demonstrations and ForceWM fixed, play raises success by 15–45 percentage points (pp), and success rises with the amount of play (Table 1, Fig. 4).
  2. A visual-wrench future as the interface makes play usable. We treat contact execution as inverse dynamics conditioned on a visual-wrench future. An IDM conditioned on the recorded future can be supervised by any recorded interaction, so task-agnostic play trains it while ForceWM learns from demonstrations only. Mixing the same play into a policy hurts (+Play rows of Table 1). The wrench part is needed because a visual future does not determine force. On Slide, a visual-only IDM reaches 15% and a wrench-only IDM 20%, against 85% with both (Table 2).
  3. A reactive IDM designed for the force signal. Contact force changes within an action chunk and cannot be read from the image, so the IDM is causal in the measured wrench: action token kk sees measured wrench 0:kk, and the reference wrench only at offsets larger than kk. This gives within-chunk reaction with no intervention data. Without it, success drops from 85/95/85 to 40/50/40 (Post 4).

Table R1. Differences from the works you list

Prior workWhat it providesWhat differs in ForceWAMWhy it matters
UniPi, VERAVideo planning with a separate IDM; VERA trains its IDM on self-playThe IDM is conditioned on a visual-wrench futureA visual future leaves force open, so play cannot teach how much contact force to apply (VERA: 10/25/20, Post 2)
ImplicitRDPCausal within-chunk force feedbackA predicted wrench reference, and an IDM trained on playIt does not predict how contact should evolve, and adding play makes it worse (Table 1)
FA-WAMJoint wrench and action prediction from task data; learned residual correctorA separately trained IDM, with reaction built into the causal IDMIts corrector needs human interventions and policy rollouts; ours needs neither

Revision. Following your suggestion, the contribution list in Sec. 1 is now written around the visual-wrench interface. It states that this formulation lets both task demonstrations and task-agnostic play supervise action prediction, and that the IDM updates actions within each chunk from measured wrench against the generated reference.

AuthorsOpenReview ↗

Response to Reviewer DMBb (2/6)

W2 / Q1: adding a baseline with a stronger way of using play; equal-budget comparison with demonstrations.

Concern: The +Play rows only add play to the behavioral supervision of ImplicitRDP and Cosmos Policy, so they do not show an advantage over stronger ways of using play such as outcome or goal conditioning. An equal-budget comparison against additional demonstrations is also missing.

As requested, we added both comparisons: VERA as an outcome-conditioned play baseline, and an equal-budget comparison between additional demonstrations and play. With the same play, VERA averages 18% and ForceWAM 88%. With 200 episodes per task family, 100 demonstrations + 100 play averages 75%, against 60% for 200 demonstrations.

VERA as an outcome-conditioned play baseline. VERA (Li et al., 2026b) decouples a video planner from a Jacobian-based IDM that can learn from task-agnostic interaction, which makes it the closest existing alternative. We adapted the official code to our tasks and kept VERA's visual inputs and Jacobian IDM. The data sources match ours, and each method keeps its own input modalities and objective.

Table R2. VERA training setup

Video plannerIDM
Initializationreleased WAN 1.3BJacobian IDM with VGGT-1B backbone
Training datatask demonstrations (same as ForceWM)task + play (same pool and split as ours)
Optimizer steps20,00020,000
Learning rate6e-65e-5
Shared settingsAdamW, BF16, official batch size, 4× 48 GB RTX 4090, seed 42same

Equal budget. We fix the IDM budget at 200 episodes per task family and compare 200 demonstrations with 100 demonstrations + 100 play. ForceWM is unchanged.

Table R3. Ways of using play (success %)

MethodPlaySlideFlipFlip-D
VERAvisual-future IDM, task + play102520
ForceWAM IDM, 200 demosnone407565
ForceWAM IDM, 100 demos + 100 playequal episode budget658575
ForceWAM IDM, 200 demos + 200 playvisual-wrench IDM, task + play859585

With the same play, VERA averages 18%. Its future specifies motion and leaves force open, so its IDM cannot learn from play how much contact force to apply. It is weakest on Slide, where the robot has to hold 10 N. Our visual-only IDM (Table 2: 15/40/30) has the same limitation even though it receives measured wrench feedback. Play improves execution when the future also specifies how contact should evolve.

At an equal episode budget, replacing half of the demonstrations with play raises the average from 60% to 75% and Slide from 40% to 65%. The gain comes from broader contact coverage, since the amount of data is the same.

Revision. Both comparisons and the VERA training setup are in App. G.2.

AuthorsOpenReview ↗

Response to Reviewer DMBb (3/6)

W3 / Q4: statistical reliability.

Concern: The statistical reliability of the reported success rates is not sufficiently characterized.

Each cell of Table 1 is one training run evaluated on 20 trials. Each trial is a binary success or failure, so for a success rate pp over nn trials the variance is fixed by the mean, p(1−p)/np(1-p)/n. We report the resulting standard error of each ForceWAM success rate (App. G.3).

Table R4. ForceWAM success rates with uncertainty (Table 1 results, 20 trials per condition; %)

SlideFlipFlip-D
ForceWAM, success ± standard error85 ± 8.095 ± 4.985 ± 8.0
Strongest baseline (Table 1)206040

The standard error of ForceWAM is 4.9–8.0 pp, while its margin over the strongest baseline is 35–65 pp. The improvement is therefore much larger than the uncertainty of the evaluation. These statistics are in App. G.3.

On play subsets: in Fig. 4 success rises steadily with the amount of play, and each intermediate point uses its own independently drawn subset (Sec. 4.3), so the gain does not depend on one particular subset.

W3 / Q4: generalization.

Concern: There is no cross-task play evaluation that establishes transferable contact knowledge, and no generalization test on unseen objects or materials.

On cross-task play, we first state what task-agnostic means in this paper and what the paper focuses on. Task-agnostic means that play is collected without a downstream task instruction and without requiring task success. As Sec. 4.1 states, it is collected in the environment of each evaluated task family. We use contact-rich play to cover a wider range of contact states than demonstrations do, which makes the IDM a stronger executor. The experiments above support this. With ForceWM fixed, play raises success from 40/75/65 to 85/95/85, and at an equal episode budget, replacing half of the demonstrations with play still raises it to 65/85/75 (Post 2). The three task families we evaluate differ strongly in visual geometry, contact phases and force response. Transfer of contact knowledge across task families and the effect of scaling up play are left to future work. The focus of this paper is how to use play for contact-rich execution: play covers the contact states that demonstrations miss, the visual-wrench interface lets a separately trained IDM learn from it, and the reactive IDM uses the measured force within each chunk to carry out the predicted interaction.

On unseen objects and materials, we ran the trained IDM, with no new data, under contact conditions that appear in none of the training data (App. G.4). Flip-D in the submission is already such a test under strong random disturbance.

Table R5. Unseen contact conditions (ForceWAM; success %, 20 trials each)

ConditionUnseen contact conditionNominal (Table 1)Unseen
Slidecompliant pad under the force scale8570
Flipunseen box with different appearance and weight9585
Flip-D (Table 1)random hand pushes, 0–45°, throughout contact95 (Flip)85

ForceWAM keeps 70% on Slide with a compliant pad and 85% on Flip with an unseen box of different appearance and weight. Under random pushes, Flip-D reaches 85%, and the strongest baseline reaches 40% (Table 1).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.