- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes ScopeVLA, a method for task-grounded spatial conditioning in multi-view VLA models. It adds learnable Global and Wrist queries to the VLM to read out task-relevant spatial distributions over the third-person and wrist views from intermediate layers. These distributions are supervised by targets built from the end-effector trajectory, which is turned into a 3D workspace field and projected onto each view. The Global distribution is used as an attention bias on third-person visual tokens, and the Wrist distribution modulates wrist-view geometric features from a frozen DA3 model, which are then added as extra keys and values in selected layers of a VLA-Adapter action head. Experiments on CALVIN ABC→D, LIBERO-Plus, and four real-world tasks show improvements over VLA-Adapter, from 4.50 to 4.62 average length on CALVIN, from 59.1 to 67.4 on LIBERO-Plus, and from 61.25% to 80.0% on the real-world tasks.
Critical strengths
- The problem is relevant. Geometric features contain both task-relevant and irrelevant structure, and selecting the task-relevant part while treating the third-person and wrist views differently is a reasonable direction for geometry-enhanced VLAs.
- The design is practical. The supervision is built automatically from trajectories and depth without manual labels, and depth, calibration, and trajectories are only needed during training. The NULL option and the zero-initialized bias keep the original model unchanged at initialization, and the inference cost is reported.
- The evaluation covers both simulation and real-world tasks, and the gains over VLA-Adapter are consistent across them. The ablations cover the components, view allocation, and injection depth, and the appendix shows how the routes respond to instruction and scene changes.
Critical weaknesses
- The novelty is limited, and the relation to prior work on task-relevant grounding is not clear. The trajectory-based supervision is essentially a heatmap of the future end-effector path on the image, and using trajectories or paths as intermediate representations or supervision has been studied in TraceVLA, HAMSTER, and PEEK, where PEEK also derives task-relevant masks automatically from trajectories. Supervising policy attention toward task-relevant regions (GuidedVLA, Gaze-Regularized VLA), instruction-aware visual feature selection (OTTER), and masking irrelevant regions (ARRO, BYOVLA) are also closely related, but none of them is discussed, and ReconVLA appears in Table 1 without any discussion. The architecture itself combines existing components such as query readout, attention bias, multiplicative gain, and extra keys and values, so the design contribution is somewhat limited.
- The baseline set is selective, and the main claims depend on it. The OpenVLA, UniVLA, and π0 numbers in Table 2 are taken from the LIBERO-Plus paper, but stronger methods from the same table are left out, including OpenVLA-OFT (69.6), RIPT-VLA (68.4), OpenVLA-OFT_m (67.9), and π0-Fast (61.6). ScopeVLA (67.4) is below the first three, and with these baselines it is no longer first on Camera, Light, and Background, so the claims of the highest overall score and first place in five of seven categories only hold for the selected baselines. The paper refers to parameter-comparable baselines, yet 7B models such as OpenVLA, UniVLA, and Spatial Forcing are kept in the table. In addition, the real-world experiments only compare with VLA-Adapter, and no variance is reported for the CALVIN gain from 4.50 to 4.62.
- The ablations do not test the main claim, and the analysis is not very convincing. The main claim is task grounding from trajectory supervision, but the ablations only add or remove modules and never change the supervision, for example by removing the routing loss or replacing the trajectory field with object masks or a uniform prior. In Table 4(a), the largest single gain comes from adding DA3 wrist features (46.3 to 56.5), while G-Condition adds only 4.8 points, so much of the improvement may come from the extra geometry model. The intervention study in Appendix A only evaluates the Stage I readout and does not test whether the policy actually uses the routes, and under wrong-task instructions the Wrist route changes little (KL from 0.102 to 0.146). The motivation in Section 3.2.1 also relies on visualizations of two selected attention heads.
- Some choices in the experimental setup raise concerns. The readout layer is selected on CALVIN and the injection layers are selected with Table 4(c) on LIBERO-Plus Object, and both are evaluation benchmarks. All ablations use only the LIBERO-Plus Object subset without variance, and although the model is trained on LIBERO, no results on standard LIBERO are reported. The real-world evaluation has only 4 tasks with 20 trials per method. Training also requires calibrated depth for both views and full trajectories, which may limit the method on large datasets without calibration.
Questions for authors
- What is the main novelty compared with methods that use trajectories or paths, such as TraceVLA, HAMSTER, and PEEK, and with methods that supervise attention directly, such as GuidedVLA and Gaze-Regularized VLA? Can the authors compare with at least one of them under the same backbone?
- Why are OpenVLA-OFT, RIPT-VLA, OpenVLA-OFT_m, and π0-Fast from the LIBERO-Plus paper not included in Table 2, and what is the criterion for parameter-comparable baselines? Also, the UniVLA total in Table 2 (43.9) differs from the original paper (42.9).
- Can the authors ablate the supervision source, for example by removing the routing loss or using object masks instead of the trajectory field? How much does trajectory supervision add on top of the DA3 features alone?
- If the routes are replaced with a uniform distribution or with routes from a wrong task at inference time, how do the success rate and behavior of the policy change? Under a wrong instruction, do the routes move toward the region of the new task?
- Were the readout layer and injection layers selected on a separate validation split? Can the authors report variance on CALVIN and LIBERO-Plus, as well as results on standard LIBERO?
- How would the method work on data without calibrated depth? Can the real-world experiments include geometry-enhanced baselines such as FALCON or Spatial Forcing, with more tasks and trials?