acceptodds
Under review as a conference paper at ICLR 2027

Actions Without Consequences: Binding World Action Model Policies to What They Decide

Abstract

World action models (WAMs) build on pretrained video models to jointly predict future observations and robot actions. Demonstration data, however, provides only one action for each observed state and therefore does not directly constrain how the predicted future should change under alternative actions. We study this gap with CF-BENCH, a counterfactual benchmark for measuring action sensitivity in WAM policies. \anchor is introduced to compare the robot pose represented in the predicted future with the pose implied by the predicted action through forward kinematics and camera calibration. This connection also enables counterfactual training without ground-truth future observations: we perturb the action, enforce the corresponding kinematic change in the robot region, and constrain regions outside the arm's swept volume to remain unchanged. With counterfactual training, ANCHOR improves the command-normalised keypoint response to 0.79 for in-distribution perturbations and 0.71 for held-out perturbations. Under matched evaluation settings, it also improves task success over base-FT by 2.3–6.6 percentage across RoboCasa, RoboTwin 2.0, and LIBERO-Plus.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Counterfactual Action Sensitivity, Kinematic Anchoring, Counterfactual Training

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MfcBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces ANCHOR to align the robot body in a WAM's generated future with the body implied by its generated action. Forward kinematics and camera calibration provide projected robot keypoints, which are compared with a learned readout from video latents. Counterfactual action perturbations extend this supervision beyond demonstrations, together with an outside-swept-region locality regularizer. The paper evaluates counterfactual responsiveness, policy success, and best-of-N (BoN) candidate selection (in appendix, which I think better to emphasize more in main paper) on two WAM backbones and three simulation benchmarks.

Critical strengths
  1. The method is sensible and uses a physically meaningful constraint. Known robot geometry supplies partial supervision for unexecuted actions without requiring ground-truth counterfactual rollouts for training. Separating keypoint-head calibration from the anchor gradient is a useful design choice: the measurement head cannot simply adjust its parameters to satisfy the agreement loss. This makes the proposed consistency constraint more credible.

  2. The reported results support practical value. ANCHOR+CF improves policy success over base-FT by 3.4, 2.3, and 6.6 percentage points on RoboCasa, RoboTwin, and LIBERO-Plus. The additional BoN experiments show that the geometric score can also be used for candidate selection. Thus, the benefits extend beyond improved scores on the paper's counterfactual-response metric.

Critical weaknesses
  1. The generalization capacity and adaptation requirements remain insufficiently characterized. The method relies on a known robot model, camera calibration, and a reliable keypoint readout from generated video. Section 4.4 states that a new embodiment requires target-data fine-tuning of the head and adapters, and that zero-shot success is near chance. The few-shot adaptation and LIBERO-Plus results are encouraging, but they leave open how robust the measurement head and score are to new robot appearances, views, and occlusions, and what adaptation overhead is required. The commanded-pose approximation and outside-swept-region prior introduce further applicability limits under contact or action-dependent camera motion. These dependencies matter because the constraint and score are only useful when their geometric measurements and assumptions remain reliable.

  2. The decision-making value of action-video agreement needs a more complete BoN evaluation. Policy SR improvements are useful, but alone do not isolate the benefit of matching imagined futures to actions from the benefit of additional geometric training supervision. I consider BoN planning a particularly direct test of the paper's motivation. Figure 7 and Table 6 do provide this test, but it is confined to the appendix and the comparison is limited: on RoboCasa, the anchor score gives 74.5% SR at N=8 versus 74.2% for the value head, without uncertainty for that pairwise difference; the corresponding value-head comparison is absent for LingBot. Table 2 also omits most alternative training methods on RoboTwin. A matched BoN evaluation across the base-FT and ANCHOR+CF generators, with clearer baseline coverage and inference budgets, would better establish how much the improved coupling contributes to planning. The current presentation emphasizes single-candidate policy gains more clearly than this central decision-making evidence.

Questions for authors
  1. Under a matched BoN candidate budget and inference-cost protocol, how do base-FT and ANCHOR+CF compare using the same selection strategies? Please clarify the missing LingBot value-head comparison and the uncertainty in the small RoboCasa ranking difference. This would materially affect my assessment of the decision-making contribution.
1 comment
AuthorsOpenReview ↗

Response to Reviewer MfcB

We thank the reviewer for recognizing ANCHOR’s physical grounding and policy gains. We agree that BoN is a direct test. We have moved the BoN results into the main text. We distinguish improved coupling from candidate-selection benefits.

Q1/W2: Matched BoN and scope

On Cosmos/RoboCasa, we cross Base-FT/ANCHOR+CF generators with Value/Anchor selectors at N=8, with 60 episodes/task/cell (3 evaluation seeds × 20 matched initial states). This independent follow-up uses one checkpoint pair trained for 1,000 updates on TurnOffMicrowave; the tasks below are exploratory transfer tests, not the submitted multi-task evaluation. We use unit anchor/CF/locality weights and CF rate 0.5 where numerical settings were unspecified.

TaskBase-FT: Value / Anchor (%)ANCHOR+CF: Value / Anchor (%)
PnPCounterToMicrowave25.0 / 26.725.0 / 30.0
OpenDrawer80.0 / 85.080.0 / 86.7
TurnOnStove36.7 / 35.040.0 / 36.7
Equal-task mean47.2 / 48.948.3 / 51.1

Within tasks, we match N, action/value denoising steps (5/1), horizon, termination, candidate seeds and initial simulator/camera/model settings. Calibrated readouts are frozen. Selectors score identical candidates at shared observations within each generator; candidates differ across generators and trajectories diverge after selection. Episodes are paired by initial state.

Equal-task-weighted paired effects (pp; 95% scene-cluster bootstrap CIs):

ContrastEffect [95% CI]
ANCHOR+CF − Base-FT, Value+1.11 [−1.67, +3.33]
ANCHOR+CF − Base-FT, Anchor+2.22 [−0.56, +5.00]
Anchor − Value, Base-FT+1.67 [−2.22, +5.56]
Anchor − Value, ANCHOR+CF+2.78 [0.00, +5.56]
Generator × selector+1.11 [−2.22, +5.00]

Cost and coverage. Candidate budgets match, latency is reported, not equalized. Generation plus communication takes 3.31–3.33 s/decision (two GPUs, PnP/Stove) and 1.71–1.72 s (six GPUs, Drawer); geometric scoring adds ~102 ms. Loading, simulation and video writing are excluded. All candidates tie at the no-keypoint-coverage floor in 36.4%/57.1%/11.2% of PnP/Drawer/Stove decisions, ties select candidate 0, with no fallback tested. The original RoboCasa 74.5% Anchor versus 74.2% Value comparison lacks episode-level matched pairs.

LingBot/RoboTwin: missing Value comparison. We did not identify a ready-to-use task-value head in the released configuration/weight index of robbyant/lingbot-va-posttrain-robotwin. An additional learned critic was not evaluated, this is a coverage limitation, not an inherent barrier to value-based selection. Zhao et al. [1] similarly marked LingBot Value as N/A. Submitted Table 3 reports N=8 SRs of 94.0/94.6/95.2% for consensus/Anchor/combined using the ANCHOR+CF generator (N=1: 92.4%). These point estimates demonstrate geometric BoN on a second backbone, without establishing superiority over consensus or Value. The matched factorial and timings above are Cosmos. We appreciate this suggestion and view a matched LingBot evaluation with a learned critic as an important direction for future work.

W1: Generalization and adaptation

New embodiments require geometry, calibration, keypoint definitions and head/adapter adaptation (§4.4). Detachment blocks agreement-loss gradients to the head; readout error can persist. Synthetic-render median errors are 2.74 px unoccluded and 28.38 px with one occluded keypoint; unseen left-to-right/right-to-left views give 15.35/89.88 px. Across 20 target-scene subsets, eight-scene (192-image) head-only adaptation gives 5.86/11.49 px in the two target settings and takes ~2.6 s on one A100. This excludes data preparation and policy/adapter tuning. These render tests do not establish zero-shot reliability on generated futures.

Table 5 reports 69.4% SR with ±2 cm/±3° extrinsic noise (single seed). Table 7 gives command–realization errors of 1.9/12.7/4.1 px before/during/after contact. Locality can penalize valid object motion outside swept support and assumes action-independent camera motion.

Coupling evidence. Joint sampling without teacher forcing improves matching accuracy from 0.54 to 0.86 and margin from 1.4 to 14.2 px (Table 6). This suggests that the generated futures are more closely associated with their corresponding actions, rather than being interchangeable across sampled candidates. Importantly, this evaluation does not rely on teacher-forced actions and therefore provides evidence of action–future correspondence under joint generation. Together with policy gains of +3.4/+2.3/+6.6 pp on RoboCasa/RoboTwin/LIBERO-Plus, these results support both improved correspondence and downstream training utility across different benchmarks. The results are consistent with improved coupling contributing to policy performance, but do not isolate it as the sole cause of the observed gains.

[1] Zesen Zhao et al. Test-Time Scaling for World Action Models via Zero-Shot Geometric Verification. In CVPR 2026 Workshop on EAI.

Reviewer_Z3vuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the mismatch between predicted observations and actions in world action models. It anchors the robot state in the imagined future to the forward kinematics of the predicted action, enabling perturbation of actions for counterfactual training without ground-truth futures, and reuses the same residual to rank candidate actions before execution. Experiments are done on two WAM backbones and simulation environments plus a held-out simulation environment, showing improved action sensitivity after training with the proposed method.

Critical strengths
  • The idea is straight forward and well motivated. When a WAM fits on the action per state in training data, it undermines using its imagined rollouts for planning and verification. Forward kinematics is a natural source of counterfactual supervision to evaluate and improve the action awareness.

  • The evaluation is thorough. The proposed CF-Bench attributes observed dynamics to robot and environment separately, and has held-out perturbation families. The gains are consistent across multiple setups and in the case of a few-shot embodiment transfer.

Critical weaknesses
  • W1: The method has only been tested on simulations. In real world, robot's forward kinematics can be much harder to approximate, so the method might enforce incorrect supervision. There is no systematic discussion about this limitation, or how to resolve this.
Questions for authors
  • Q1: With regard to W1, do you anticipate the problem to happen in the real world (or test by real world experiments if possible), and can you address this potential problem?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Z3vu

We thank the reviewer for the constructive evaluation and for recognizing the contributions of ANCHOR and CF-Bench.

W1/Q1. Robustness of FK-based supervision and real-world applicability.

1. Evidence across diverse settings. Our experiments cover different environments. Under matched training and evaluation settings, ANCHOR+CF improves task success over base fine-tuning by +3.4, +2.3, and +6.6 percentage points, respectively. LIBERO-Plus further evaluates transfer beyond the original training configurations, with particularly encouraging improvements under layout and camera shifts.

CF-Bench additionally examines diverse action perturbations, including families held out during training. On RoboCasa, the independently measured command-normalized keypoint response improves from 0.26/0.15 (base-FT) to 0.79/0.71 (ANCHOR+CF) for in-distribution/held-out perturbations, while normalized environment counterfactual error decreases from 1.00 to 0.81. These complementary experiments provide evidence of effectiveness across different WAM architectures, simulated environments, and perturbation settings, rather than a single favorable task. Although they cannot establish physical-robot robustness, they support the generality of the proposed mechanism within the evaluated conditions.

2. FK supervision and commanded-versus-realized motion. ANCHOR exploits a specific source of geometric information: given the robot model, controller interface, and camera calibration, the commanded robot geometry can be projected into the image, even for counterfactual actions that were never executed. This provides partial, action-conditioned supervision without requiring ground-truth counterfactual futures or a complete physics simulator.

The distinction between commanded and realized motion is important. Contact, compliance, controller tracking error, inaccurate robot parameters, and camera miscalibration can cause the physically observed robot pose to differ from its commanded geometric target. ANCHOR therefore does not claim to model contact-mediated object dynamics or guarantee physically realized future poses.

Our original submission already investigates this issue through a contact-phase analysis in RoboCasa simulation (Appendix Table 5):

Contact phaseReprojection discrepancy (px)Realized-motion referenceANCHOR+CF response
Pre-contact1.90.970.88
In-contact12.70.740.61
Post-contact4.10.910.79

The last two columns are command-normalized motion-response measurements, not pose-accuracy scores or universal upper bounds. ANCHOR+CF reaches approximately 91%/82%/87% of the corresponding empirical references. The larger in-contact reprojection discrepancy identifies a regime in which commanded geometry becomes a less reliable proxy for realized motion. Thus, the evaluation explicitly characterizes this limitation rather than assuming perfect execution.

3. Additional calibration and visual-readout diagnostics. We further examine two sources of geometric uncertainty relevant to physical deployment.

First, in a geometry-only simulation diagnostic at 224×224 resolution, independent camera-extrinsic perturbations of ±2 cm per translation component and ±3° per rotation-vector component shift FK projections by 9.37 px median and 17.35 px P95. This directly quantifies sensitivity to imperfect camera calibration, although it does not measure real-camera or end-to-end policy robustness.

Second, the geometric constraint depends on reliably identifying the robot body from the generated future representation. We therefore evaluated the separately calibrated keypoint readout on held-out synthetic renders:

Readout conditionMedian error (px)P95 error (px)
No occluder2.7421.74
One keypoint occluded28.38106.67
Unseen left-to-right viewpoint15.3594.01
Unseen right-to-left viewpoint89.88159.88

These results identify significant degradation under occlusion and unseen viewpoints. In a repeated head-only adaptation study with 20 target-scene subsets per setting, eight target-view scenes reduce median held-out errors to 5.86/11.49 px in the two viewpoint directions.

4. Practical mitigation and planned real-robot validation. Simulation plays an important role in our central research question: CF-Bench can reset to matched initial states, execute alternative actions, and obtain ground-truth counterfactual outcomes, whereas ANCHOR's geometric training targets do not require these rollouts. This controlled evaluation makes it possible to investigate action–future consistency systematically. We regard physical-robot validation as an important next step. We plan to evaluate ANCHOR on real robots. If the paper is accepted, and hardware access, we will make our best effort to include real-robot results in the camera-ready version.

Reviewer_N6UdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the misalignment issue for world action models between the generated future frame and the actions. To combat this issue, this paper converts the predicted actions into coordinates through forward kinematics and camera calibration and penalizes the discrepancy between the resultant coordinates and those of the model-imagined frame.

Critical strengths
  1. The paper tackles a critical question in world action models: the potential inconsistency between the imagined future and the predicted action. This question is important and timely.

  2. It devises a simple, mechanical correction mechanism to enforce alignment between the two outputs from world action models, for both demonstrations and counterfactuals. The solution is robust in spirit (grounded by the mechanic projections), and the paper demonstrates respectful improvements in its experiments.

Critical weaknesses
  1. The correction mechanism might be limited - the pose estimation and projection components may not provide sufficiently accurate targets. The projected pose can differ from the actual future pose. For instnace, their method does not directly account for interactions with other objects. So, one may suspect that its effectiveness may be limited in more complex scenarios. Moreover, the locality constraint may discourage legitimate object motion outside the robot’s swept region.

  2. The method relies on additional components and information, including a robot kinematic model and a keypoint estimation head. Errors in these components could signifcantly undermine the method.

  3. Experiments are mostly simulation. Although I understand the difficulty, real-robot evaluation may provide invaluable support to justify their capability of handling complex real-world tasks.

Questions for authors

Please see the weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer N6Ud

We thank the reviewer for the constructive feedback and valuable suggestions.

W1. Commanded geometry, contact, and locality. ANCHOR does not use forward kinematics (FK) to predict the complete physical future. Rather, FK and camera calibration provide partial, action-conditioned supervision of commanded robot geometry, even for unexecuted counterfactual actions without ground-truth future observations.

Our RoboCasa simulation analysis quantifies commanded-versus-realized discrepancies.

PhaseReprojection error (px)Realized-motion referenceANCHOR+CF response
Pre-contact1.90.970.88
In-contact12.70.740.61
Post-contact4.10.910.79

The last two columns measure command-normalized responses, not pose accuracy. ANCHOR+CF reaches approximately 91%/82%/87% of the phase-specific empirical references. The larger contact-phase error highlights the limits of commanded geometry as a proxy for physical realization. ANCHOR does not claim to model contact-mediated object dynamics.

Locality is a soft inductive bias rather than a physical law. Sliding, rolling, articulated motion, and indirect contact may produce legitimate changes outside the robot's swept region. The submitted ablation reports 70.3% task success with locality versus 68.5% without it, alongside a separately reported 32% increase in environment counterfactual error without locality. Since the full-model result averages three seeds while the ablation is single-seed, these are descriptive comparisons.

W2. Readout reliability, calibration, and adaptation. The separately calibrated, stop-gradient keypoint head prevents the anchor loss from directly modifying the measurement head, but accuracy still depends on robot geometry, camera calibration, and visual readout.

We conducted held-out synthetic-render diagnostics to quantify sensitivity to occlusion and viewpoint changes:

ConditionMedian error (px)P95 (px)
No occluder2.74 [2.48, 3.12]21.74
One keypoint occluded28.38 [24.11, 43.65]106.67
Left-to-right cross-view15.35 [8.71, 24.51]94.01
Right-to-left cross-view89.88 [76.69, 113.66]159.88

Brackets denote 95% scene-level confidence intervals for the median. These results quantify substantial degradation under occlusion and unseen viewpoints.

To assess adaptation feasibility, we additionally evaluated head-only cross-view calibration over 20 independently sampled target-scene subsets per setting, using held-out synthetic validation scenes:

Scenes (images)ViewMedian error (px)P95 (px)Fit time (s)
1 (24)L-to-R20.51 [9.09, 24.27]91.12 [75.75, 110.93]0.42
4 (96)L-to-R7.87 [5.36, 21.37]76.02 [63.40, 93.58]1.29
8 (192)L-to-R5.86 [4.62, 8.62]70.35 [63.95, 85.00]2.60
1 (24)R-to-L40.96 [22.29, 54.85]124.69 [90.23, 143.64]0.42
4 (96)R-to-L14.95 [11.13, 23.41]76.85 [47.85, 88.49]1.30
8 (192)R-to-L11.49 [8.94, 16.14]63.88 [47.65, 80.14]2.61

Increasing target scenes from 1 to 8 reduces median held-out error by approximately 71–72% across both directions, demonstrating the benefit of target-view adaptation, although P95 errors remain substantial. Brackets here are min–max ranges across sampled subsets, not confidence intervals. Fitting times measure head-only optimization on one A100, excluding data acquisition, annotation, and adapter/policy tuning.

A separate synthetic camera-extrinsic perturbation (±2 cm, ±3°) produces a median FK projection shift of 9.37 px (P95: 17.35 px). These analyses characterize measurement and head-adaptation requirements, not robustness of generated videos, full embodiment transfer, or real-camera perception.

W3. Real-robot validation. Simulation is particularly suitable for studying ANCHOR's central question: whether imagined futures respond consistently to counterfactual actions. It enables controlled action perturbations, reproducible initial states, and precise analysis of robot geometry, contact phases, and measurement errors. These controlled conditions allow us to isolate the effect of action–future consistency without confounding factors from physical execution. We are preparing an evaluation protocol for future real-robot experiments, focusing on contact-rich manipulation, commanded-versus-realized motion, geometric readout reliability, and task success.

Overall, ANCHOR+CF improves simulated task success over base-FT by 3.4/2.3/6.6 percentage points on RoboCasa/RoboTwin/LIBERO-Plus and achieves command-normalized keypoint responses of 0.79 in-distribution and 0.71 on held-out perturbations. These results support mechanically grounded counterfactual supervision without requiring ground-truth counterfactual rollouts or a full physics simulator.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Reviewers

Summary of the Revision

We thank all reviewers for their constructive feedback and valuable suggestions. We have revised our manuscript with additional controlled experiments, expanded robustness analyses, and clearer Best-of-N (BoN) comparisons.

Our central contribution remains unchanged. ANCHOR addresses the mismatch between predicted actions and imagined futures in World Action Models. By grounding the predicted robot geometry in forward kinematics (FK), ANCHOR provides action-conditioned supervision for both demonstrated and unexecuted counterfactual actions without requiring ground-truth counterfactual future rollouts for training. The original evaluations show task-success improvements of +3.4/+2.3/+6.6 percentage points on RoboCasa, RoboTwin 2.0, and LIBERO-Plus, respectively.

The main revisions are summarized below.

  1. Elevated BoN Evaluation and Clarified the Decision-Making Contribution. Following the reviewer's suggestion, we have moved the BoN results into the main manuscript and more clearly distinguished improvements in WAM generation from improvements attributable to candidate selection. Anchor Score offers an execution-free geometric consistency signal without reward labels or a separately trained value head. We now explicitly distinguish this capability from claims of superior task-success ranking.

  2. Added a Matched Generator-by-Selector BoN Evaluation (Additional Cosmos/RoboCasa Experiments). We conduct a controlled 2×2 comparison crossing Base-FT and ANCHOR+CF generators with Value and Anchor selectors at N=8. The study covers three transfer tasks, with 60 closed-loop episodes per condition per task and matched candidate-generation settings. For example, ANCHOR+CF achieves 30.0% versus 25.0% success with Anchor versus Value selection on PnPCounterToMicrowave, and 86.7% versus 80.0% on OpenDrawer; however, Value performs better on TurnOnStove (40.0% versus 36.7%).

  3. Expanded Statistical Uncertainty, Inference-Cost Reporting, and Baseline Clarifications (BoN Analysis). We report equal-task-weighted generator, selector, and interaction effects with 95% scene-cluster bootstrap confidence intervals, all of which touch or include zero. We also clarify that the submitted RoboCasa BoN@8 comparison (74.5% Anchor versus 74.2% Value) is descriptive because the original paired episode records are unavailable. The new experiments report matches candidate budgets, measures generation latency, and approximately 102 ms additional geometric scoring time.

  4. Clarified Evidence for Action–Future Coupling Beyond Teacher-Forced Evaluation (Counterfactual Evaluation). Alongside CF-Bench's held-out perturbation tests, we emphasize the joint-sampling analysis without teacher forcing. The reported action–future matching accuracy improves from 0.54 to 0.86, with matching margin increasing from 1.4 to 14.2 pixels. On RoboCasa, command-normalized keypoint response improves from 0.26/0.15 to 0.79/0.71 on in-distribution/held-out perturbations, while normalized environment counterfactual error decreases from 1.00 to 0.81. These measurements provide complementary evidence for action–future correspondence, independently of the exploratory BoN ranking results.

  5. Added Readout Robustness Diagnostics for Occlusion and Unseen Viewpoints (Keypoint Readout Analysis). We evaluated the separately calibrated keypoint head on held-out synthetic renders. Median error increases from 2.74 pixels without occlusion to 28.38 pixels with one occluded keypoint, while unseen left-to-right/right-to-left viewpoints produce 15.35/89.88 pixels median error. We also report upper-tail errors and uncertainty estimates, explicitly identifying situations in which the geometric readout becomes unreliable.

  6. Expanded Target-View Adaptation and Calibration Sensitivity Analysis (Adaptation and Geometry Diagnostics). Across 20 sampled target-scene subsets per setting, head-only adaptation using eight target-view scenes (192 images) reduces median cross-view errors to 5.86/11.49 pixels. Median fitting time is approximately 2.6 seconds on one A100, excluding data preparation and policy adaptation. Separately, synthetic camera-extrinsic perturbations of ±2 cm and ±3° produce a median FK projection shift of 9.37 pixels. These diagnostics quantify adaptation requirements and calibration sensitivity rather than establishing full real-world robustness.

AuthorsOpenReview ↗

Summary of Our Responses to Reviewers

Dear Area Chair and Reviewers,

Thank you again for your thoughtful reviews and helpful suggestions. As the author-response period draws to a close, we would like to briefly summarize the additional evidence and clarifications provided in response to the main questions.

1. Controlled BoN evaluation (Reviewer MfcB). Following the reviewer’s suggestion, we report an additional generator-by-selector comparison on Cosmos/RoboCasa, crossing Base-FT and ANCHOR+CF with Value and Anchor selection at N=8. Our response includes matched candidate budgets, uncertainty estimates, and inference-cost measurements, which help distinguish generator improvements from selection benefits. We interpret these follow-up ranking results as exploratory.

2. Geometric reliability and adaptation (Reviewers MfcB, N6Ud, and Z3vu). To better characterize the dependencies highlighted in the reviews, we provide additional synthetic-render diagnostics for occlusion and unseen viewpoints, together with camera-calibration sensitivity and head-only target-view adaptation analyses. These measurements help clarify the readout’s limitations and the benefit of target-view calibration.

3. Commanded versus realized motion and applicability (Reviewers N6Ud and Z3vu). Our responses clarify the role of FK as partial supervision of commanded robot geometry and discuss the contact-phase discrepancy measurements. We also explicitly discuss the locality assumption.

Alongside these clarifications, the joint-sampling evaluation without teacher forcing shows matching accuracy improving from 0.54 to 0.86, which provides complementary evidence for improved action–future correspondence.

We hope this summary helps the AC and reviewers navigate our detailed responses and consider the additional evidence.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.