- 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
-
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.
-
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
-
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.
-
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
- 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.
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.
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):
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.