- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
Main contribution: SMA-VLA extends VLA-Adapter (a 0.5B Qwen2.5 backbone with a Policy Bridge) using three add-on modules. The central claim is that conditioning task semantics on proprioception, and aligning them with the coarse direction of demonstrated motion, improves generalization when the robot’s initial state, the instruction wording, or the visual scene shifts.
Key innovations:
- MCE (Multi-Level Context Enhancement): a softmax-weighted mixture over VLM layers. Its difference from the final layer is injected into every layer as a gated, zero-initialized residual.
- PC-SMA (Proprioception-Conditioned Semantic–Motion Alignment):
- A task query built from pooled language features plus gated proprioception.
- The query applies a bounded multiplicative modulation to visual tokens.
- A training-only auxiliary loss pushes a direction head toward the net translational direction of the expert chunk. The loss is a hinge outside a cone whose width adapts to how coherent the motion is.
- PFB (Patch-Focused Bridge): action-to-patch attention inside Bridge blocks. It adds a gated residual of focused context minus the uniform patch average.
Critical strengths
- Well-motivated problem. The paper argues that VLA policies under-use proprioception when mapping semantics to motion. This is a reasonable hypothesis, and the robot-state column of LIBERO-Plus is a natural testbed for it.
- Careful, conservative module design. All three modules use bounded, zero-initialized gates. The policy therefore starts as the original VLA-Adapter computation, which makes the additions low-risk and easy to adopt. The direction loss is confined to coarse translation, and the cone width scales with motion coherence. These are thoughtful details that avoid over-constraining multimodal or curved demonstrations.
- No inference-time cost from the auxiliary branch. The direction head and loss are training-only.
- Targeted out-of-distribution gains. The +12.3 pp on robot-state shifts and +6.7 pp on language shifts over the direct baseline are sizable. In the ablation (Table 4), they line up with the addition of PC-SMA, which supports the paper’s narrative.
- Honest reporting of failures. The paper does not hide that it underperforms on camera and noise shifts or on LIBERO-Pro positional swaps.
- Clear writing. The method section is concise and the equations are easy to follow.
Critical weaknesses
W1. Statistical significance of the main results.
- No seeds, standard deviations, or confidence intervals are reported anywhere.
- With 500 rollouts per suite at around 96% success, the binomial standard error per suite is about 0.9 pp. Most per-suite differences in -Table 1 are therefore within about one standard error, before even counting seed-to-seed variance.
- SMA-VLA is slightly worse than VLA-Adapter on LIBERO-Object.
- On LIBERO-Plus, the 0.46 pp overall margin over Anchor-Align is not meaningful without variance estimates.
- The claim of “the highest overall success rate” needs to be softened or supported with multiple seeds.
W2. Provenance and comparability of baseline numbers.
- Please state which baseline numbers were reproduced and which were copied from prior papers, and under what settings.
- The VLA-Adapter and OpenVLA-OFT LIBERO numbers in Table 1 appear lower than the figures I recall from the original papers. (Reviewer note: verify this against the original papers before including it.) If the authors re-trained these baselines, the training budget and configuration should be specified.
- In Table 2, baselines are reported to one decimal place, while SMA-VLA is reported to two. This suggests the baseline numbers were taken from other sources. It is then unclear whether they follow the same “train on LIBERO-Spatial only” protocol stated in Section 4.2.
W3. Compute fairness.
- SMA-VLA is trained for 60K steps and then a further 20K steps of PFB refinement.
- The ablation states that all variants share the same total step count. The main comparison against VLA-Adapter does not say the same.
- Please report VLA-Adapter trained for the same 80K steps, as well as parameter counts and inference latency for each added module.
W4. The ablations do not isolate the claimed mechanism.
The paper’s central claim is that semantic–motion alignment drives generalization. However, PC-SMA bundles three separate things: proprioceptive conditioning of the query, token modulation, and the auxiliary direction loss. The following ablations are needed:
- PC-SMA without the alignment loss (λ_SMA = 0).
- PC-SMA without proprioception in the query.
- A fixed cone in place of the coherence-adaptive cone, and no motion-coherence weighting.
- Sensitivity to λ_SMA.
- PC-SMA and PFB each without MCE.
Without the first two, the robot-state gain could come simply from feeding proprioception into visual conditioning rather than from alignment. A related question: in Table 4, MCE+PC-SMA and the full model obtain identical robot-state scores (64.86%). Is this a coincidence, or a copy error?
W5. The mechanism is under-supported.
- The gradients of L_SMA are stopped at the VLM, so the loss trains only the query and proprioception projections and the direction head. - Its effect on actions therefore runs entirely through the bounded token modulation in Eq. 5.
- The paper gives no evidence that this pathway does what is claimed. Useful evidence would include:
- Direction-prediction accuracy in- and out-of-distribution.
- Visualizations of the relevance distribution r.
- Evidence that the query changes appropriately when the initial state is perturbed.
- It is also not obvious why aligning with demonstrated directions, which come from in-distribution initial states, should help when starting states are perturbed. A short analysis would strengthen the paper.
- The ablation shows mixed component effects:
- PC-SMA reduces camera robustness, from 94.68% to 89.36%.
- PFB reduces language robustness, from 84.10% to 81.54%.
- MCE alone slightly reduces language robustness.
These patterns deserve discussion rather than the blanket statement that the modules are “complementary.”
W6. Novelty relative to Anchor-Align.
- The related work notes that Anchor-Align already “promotes language-action consistency using motion-direction supervision.” This is close to the core of PC-SMA.
- Anchor-Align appears to be contemporaneous. Even so, since the authors compare against it, they should state precisely what is new: proprioceptive conditioning, the adaptive cone, MCE/PFB, or some combination.
- More broadly, MCE and PFB resemble existing ideas: layer mixing (M2-VLA, LayerRoute) and focused patch attention (FocusVLA, AVA-VLA). The overall contribution reads as a careful but incremental combination on top of VLA-Adapter.
W7. Limited evaluation scope.
- All results are in simulation on the LIBERO family.
- All out-of-distribution evaluation uses a single training suite (Spatial).
- Generality would be much better supported by any of the following:
- Out-of-distribution evaluation from other training suites.
- A second benchmark, such as SimplerEnv, CALVIN, or RoboTwin.
- A small real-robot study, especially for robot-state shifts, where real proprioceptive noise matters.
- The 2.6% result on positional swaps suggests the policy still largely memorizes target locations. This sits uneasily with the claim of state-dependent motion reasoning and should be discussed in more depth.
W8. Reproducibility details. The reproducibility statement says Section 3 fully specifies the method, but several values are missing:
- τ, ρ_s, g_M, g_T, and the bounds on g_P. λ_SMA.
- How cone angles θ_t and weights w_t are computed from coherence.
- The static-chunk threshold.
- Which Bridge blocks are “eligible” for PFB.
- The learning rate for the second stage.
W9. Qualitative results. Figure 3 shows single successful trials of SMA-VLA alongside failures of VLA-Adapter. Without per-condition statistics or representative failure cases for SMA-VLA, it adds little evidence.
Questions for authors
- Were the baselines in Tables 1–3 re-run under your protocol and compute budget? If not, what were their sources and training setups?
- Can you report mean ± std over at least three seeds for Tables 1, 2, and 4?
- What happens with λ_SMA = 0, and with proprioception removed from the query?
- How accurate is the direction head out-of-distribution, and does it correlate with rollout success?
- Why does PC-SMA hurt camera-shift performance?
- What is the inference latency and parameter overhead relative to VLA-Adapter?
- How exactly does PC-SMA differ from Anchor-Align’s motion-direction supervision?
Response to Authors
Thank the authors for their response and efforts. While some of my concerns have been addressed, I remain some concerned about the paper’s novelty and experimental evaluation. I will raise my rating to 2, but I believe the paper needs further improvement to meet the bar for acceptance at ICLR.