acceptodds
Under review as a conference paper at ICLR 2027

Ordered Action Transport for Generative Robot Policies

Abstract

Generative robot policies, including vision-language-action models (VLAs) and world-action models (WAMs), generate an action chunk by jointly refining a sequence of action tokens, one per future action. These tokens inherit rotary position encoding (RoPE), which relates two tokens by how many steps apart they are. However, existing positional encodings represent action-token relations primarily by their positions in the sequence, overlooking the motion induced by the actions are taken and in what order. We propose Action-Transport RoPE (AT-RoPE), which advances the encoding at each step by a rotation determined by the action rather than by a fixed one. Each current noisy action parameterizes an orthogonal operator, and these operators are composed in execution order, so that the prefix transformations of two tokens differ by the composition of the actions between them. AT-RoPE reduces to RoPE-style rotation when the step is action-independent and to an order-insensitive summed prefix when the operators commute. For backbones with rotary attention, a complementary pathway further conditions query-key matching on the actions being compared. Across four VLA and WAM backbones, AT-RoPE improves success on LIBERO and RoboTwin 2.0 for every backbone, with consistent gains on long-horizon tasks, and improves zero-shot robustness on LIBERO-Pro and LIBERO-Plus. Ablations confirm that ordered composition outperforms prefix-free and summed-prefix alternatives, and rollouts with AT-RoPE are smoother.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, World-Action Models, Rotary Position Embedding, Robotics

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_f6aSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AT-RoPE, an action-conditioned encoding method that augments the original positional encoding. It composes learned orthogonal transformations in action-sequence order to explicitly incorporate information from preceding candidate actions. For WAM backbones, a complementary pathway also conditions query–key matching on the actions being compared. The authors evaluate the method on four backbones and multiple benchmarks, reporting improvements in average success rates.

Critical strengths
  1. The paper studies an interesting topic in embodied AI: how to encode action sequences. Standard RoPE provides positional information, while robot control also depends on the content and execution order of actions. Exploring ways to encode this structure more explicitly is worthwhile.

  2. The experiments are extensive, covering four backbones, multiple simulation benchmarks, and real-robot tasks.

  3. The paper is generally well written, and the main idea is easy to follow.

Critical weaknesses
  1. My main concern is sensitivity to errors in earlier candidate actions. Although the prefixes are recomputed at each refinement step and incorporated through gated residual updates, they are still built from potentially incorrect candidate actions. Could this additional dependence make an incorrect action chunk harder to revise? Targeted tests of plan correction and recovery from execution failures would help clarify this concern.

  2. I am also concerned about task-irrelevant variation in demonstration actions. The paper already includes noisy-action training and robustness evaluations, but these do not directly isolate whether the added prefix pathway learns unnecessary dependencies on incidental movements in the demonstrations. A controlled comparison with and without such variation would help assess this possibility.

  3. The settings of the optional VLA components are not fully clear. Algorithm 1 and Appendix B.1 describe observation-derived coefficient conditioning and prefix-derived query tokens, but I could not find which options were enabled for each reported VLA configuration. The paper states that paired comparisons retain the native backbone settings; please also specify the settings of these added components.

  4. Some main-benchmark gains are small, and their uncertainty is not reported. For example, Table 1 shows LIBERO average success increasing from 99.3% to 99.4% for OpenWAM-α and from 96.9% to 97.4% for π0.5. Figure 4 provides useful five-run means and standard deviations for the separate prefix ablation, but these do not establish the variability of the Table 1 results. Reporting training-run variability and evaluation confidence intervals for the main comparisons would help assess whether these small improvements are reliable.

Questions for authors

See weakness.

7 comments
AuthorsOpenReview ↗

Response W1. Correction of corrupted candidate actions

Response W1. Correction of corrupted candidate actions

1) Controlled prefix-perturbation experiment

Evaluation scope. We compare Fast-WAM with and without AT-RoPE on RoboTwin, using 900 cases per model.

Targeted intervention. At the fifth step of ten denoising steps, we perturb the first 25% of the candidate chunk with Gaussian noise on non-gripper dimensions, scaled by one per-dimension action scale, then continue joint refinement normally.

Matched conditions. Within each model, perturbed and unperturbed runs share the same initial noise, and both models receive matched perturbations in physical action units.

2) Correction metrics and results

Metrics. We measure normalized prefix recovery RR and the perturbation-induced change in demonstration-referenced error over the segment selected for execution:

R &= 1 - \frac{\operatorname{RMS}_{P}(a_{\mathrm{pert}}-a_{\mathrm{clean}})}{\operatorname{RMS}_{P}(\delta_{\mathrm{eff}})}, \\\\ \Delta\mathrm{MSE}_{X} &= \mathrm{MSE}_{X}(a_{\mathrm{pert}},a_{\mathrm{demo}}) - \mathrm{MSE}_{X}(a_{\mathrm{clean}},a_{\mathrm{demo}}). \end{aligned}$$ where $P$ is the perturbed prefix, $X$ is the execution segment, and $\delta_{\mathrm{eff}}$ is the applied perturbation; $a_{\mathrm{pert}}$ and $a_{\mathrm{clean}}$ are the final predictions with and without intervention. Thus, $R=1$ denotes complete removal of the injected deviation relative to the unperturbed prediction. **Comparable perturbation tolerance.** AT-RoPE and the baseline achieve closely matched normalized prefix recovery of **99.642% and 99.630%**, respectively. The added prefix pathway therefore preserves the baseline's ability to correct corrupted candidates, with **no observed increase in sensitivity or fragility under the tested prefix perturbations**. **Execution-segment error.** The paired difference in perturbation-induced $\Delta\mathrm{MSE}_{X}$ is close to zero, with its 95% confidence interval spanning zero. **Table J. Recovery from intermediate prefix corruption on RoboTwin using Fast-WAM (900 cases per model).** Paired effects are AT-RoPE minus baseline; CI denotes the 95% confidence interval. **(a) Normalized prefix recovery** | Metric | Baseline | AT-RoPE | |---|---:|---:| | Normalized recovery, $100R$ (%) | 99.630 | **99.642** | **(b) Paired effects** | Paired effect | Estimate | 95% CI | |---|---:|---:| | Recovery difference $\Delta R$ (×10⁻⁴) | +1.12 | [+0.47, +1.84] | | Difference in $\Delta\mathrm{MSE}_{X}$ (×10⁻⁶) | +2.00 | [−6.20, +10.00] | **Connection to the design.** The measured recovery is consistent with reconstructing prefixes from updated candidates at every refinement step, rather than retaining corrupted prefixes as fixed commitments. > ### 3) Complementary task-level evidence **Evaluation scope.** The intervention isolates intermediate candidate correction, while benchmark and real-robot evaluations provide complementary evidence from complete task execution. **Benchmark performance.** AT-RoPE improves benchmark-average success for **all four backbones on RoboTwin 2.0** (manuscript Table 3). **Real-world performance.** Success increases from **65% to 80%** for object packing, **20% to 90%** for cylinder insertion, and **20% to 75%** for dynamic object manipulation (Figure 5). **These results show that preceding-action conditioning improves task performance while retaining effective candidate correction under the tested prefix perturbation.**
AuthorsOpenReview ↗

Response W2. Sensitivity to task-irrelevant demonstration variation

Response W2. Sensitivity to task-irrelevant demonstration variation

1) Controlled training-data comparison

Training conditions. We train Fast-WAM with and without AT-RoPE using reference demonstrations and demonstrations containing task-irrelevant movement variation, yielding four model–training-condition combinations.

Evaluation scope. Each configuration is evaluated over 500 trials, directly testing sensitivity to training-demonstration variation rather than inference-time candidate noise.

2) Performance and relative sensitivity

Retained performance advantage. AT-RoPE outperforms the baseline under both conditions, with its advantage increasing from 2.2 points on reference demonstrations to 2.8 points on varied demonstrations.

Smaller observed decrease. Introducing the variation reduces success by 0.8 points with AT-RoPE versus 1.4 points for the baseline, a 0.6-point smaller decline.

Table K. Task success under reference and varied-demonstration training. Counts are successful trials/total trials; changes and advantages are in percentage points.

PolicyReference demonstrationsVaried demonstrationsChange
Fast-WAM baseline78.8% (394/500)77.4% (387/500)−1.4
Fast-WAM + AT-RoPE81.0% (405/500)80.2% (401/500)−0.8
AT-RoPE advantage+2.2+2.8+0.6

3) Connection to the design and correction study

Adjustable prefix dependence. The learned gated residual update adjusts the contribution of prefix-conditioned features while retaining the native token representation and prediction objective (manuscript Eq. (5)).

Complementary diagnostics. W1 tests correction of corrupted candidates during inference, whereas this experiment shows that AT-RoPE retains its task-performance advantage under the tested variation in training demonstrations.

AuthorsOpenReview ↗

Response W3. Settings of the optional VLA components

Response W3. Settings of the optional VLA components

1) Enabled and disabled options

Observation conditioning: enabled (b_c = 1). Both π₀.₅ and StarVLA-π use observation context from native prefill together with current candidate actions to condition the operator coefficients.

Prefix-derived query tokens: disabled (b_f = 0). The optional query sequence is empty, so the decoder suffix contains exactly HH transported action tokens; ordered token transport remains enabled.

2) Token transport and gating

Operator settings. Both VLA integrations use 512 transported channels, p = 4, and K = 4 generators shared across blocks.

Token gate. The gate uses only token features, βi=σ(bh(hi))\beta_i=\sigma(b_h(h_i)), with s_h = 1 and the explicit action/time gate terms set to zero. Action conditioning remains in the operator coefficients, and native decoder time conditioning is retained.

Local Q/K modulation. This pathway is used only in WAM integrations, not in either reported VLA model.

Table L. Fixed AT-RoPE settings for the reported VLA configurations, shared across each backbone's evaluations. Generators are shared across feature blocks.

Added-component settingπ₀.₅StarVLA-π
Observation-derived coefficient conditioning, b_cEnabled (1)Enabled (1)
Prefix-derived query tokens, b_fDisabled (0)Disabled (0)
Additional query-token count00
Ordered token transportEnabledEnabled
Transported feature channels512512
Block width, p44
Number of generators, K4 (shared)4 (shared)
Token-gate conditioningToken features; s_h = 1Token features; s_h = 1
Local post-RoPE Q/K modulationNot usedNot used

3) Consistency across evaluations

Fixed added-component configuration. All reported π₀.₅ and StarVLA-π evaluations use the settings above, including π₀.₅'s RoboTwin and real-robot experiments, without benchmark-specific tuning of these components.

Checkpoint reuse. For π₀.₅, the same checkpoint trained on the four LIBERO suites is reused for LIBERO, LIBERO-Pro, and LIBERO-Plus.

Native versus added settings. “Retaining native backbone settings” refers to the data-processing pipeline, optimizer, learning-rate schedule, training updates, and native inference configuration; Table L explicitly specifies the added components described in Algorithm 1 and Appendix B.

Reviewer_e93fOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces AT-RoPE, which incorporates ordered action-conditioned feature transformations into generative robot policies. The method maps current noisy actions to orthogonal operators, accumulates them in execution order, and updates action tokens through gated residual connections. For WAM backbones, it also introduces local action-conditioned Q/K modulation. AT-RoPE preserves each backbone’s native training objective. Experiments on four VLA and WAM backbones demonstrate improvements in simulation success rates, robustness under distribution shifts, and performance on three real-world tasks.

Critical strengths
  1. The method has a clear design supported by targeted structural ablations. Ordered products of noncommuting operators can distinguish different orderings of the same actions, providing a reasonable inductive bias for action sequences. Table 4 compares Local, Sum-prefix, and ordered-prefix composition, and separately evaluates the token and Q/K pathways. Figure 4 further reports results from five independent retraining runs for each prefix variant. These experiments provide empirical support for the role of ordered composition.
  2. Evaluation covers multiple backbones, with substantial gains in several settings. All four backbones achieve higher average success rates on both LIBERO and RoboTwin 2.0. Beyond standard benchmarks, Fast-WAM improves by 20.7 percentage points on LIBERO-Plus. Real-world success rates increase by 70 percentage points for cylinder insertion and 55 percentage points for dynamic object manipulation.
Critical weaknesses
  1. Implementation settings and computational overhead are not fully reported. Appendix B does not specify settings such as the number of generators and block width. The paper also does not quantify the additional cost of recomputing operators and prefix products at each refinement step.
  2. The smoothness analysis has limited coverage. Figures 1 and 3 analyze the curvature and velocity spectra of selected trajectories. However, aggregate statistics across tasks and multiple rollouts are missing, making it difficult to determine whether these improvements hold more broadly.
Questions for authors
  1. I still observe noticeable jitter in the real-world videos and find it difficult to assess the smoothness advantage of AT-RoPE over the baseline. Could the authors provide real-world comparisons under comparable conditions, supported by trajectory metrics, to clarify whether the improvements shown in Figures 1 and 3 also occur during real-world execution?
  2. Section 4.5 reports higher real-world success rates and discusses alignment during insertion and action coordination in dynamic object manipulation. Could the authors provide comparative examples of the baseline and AT-RoPE to explain how ordered action transport changes execution behavior at key stages and which types of failures it reduces?
6 comments
AuthorsOpenReview ↗

Response W1. Implementation settings and computational cost

Response W1. Implementation settings and computational cost

Appendices B.2–B.4 now specify the added-module settings and report measured costs, including operator and prefix recomputation throughout the complete action query.

1) Explicit implementation settings

Operator configuration. Token transport uses 512 channels, p = 4, and K = 4 in both VLA integrations and the profiled Fast-WAM; Fast-WAM Q/K modulation uses 128 channels per head, p = 4, and K = 8.

VLA options. Both VLA backbones enable observation conditioning (b_c = 1), disable additional query tokens (b_f = 0), and use a token-feature gate with s_h = 1, without local Q/K modulation.

Fast-WAM gates. The token gate uses token, action, and refinement-time inputs with s_max = 1, while Q/K gate scales are bounded by λ_max = 1.

Table F. Numerical settings of the reported VLA integrations and the profiled Fast-WAM configuration. Width denotes transported channels or channels per attention head; token generators are shared across blocks.

IntegrationPathwayWidthpK
π₀.₅ / StarVLA-πToken transport51244
Fast-WAMToken transport51244
Fast-WAMAction Q/K128/head48
Fast-WAMVideo Q/K128/head48

2) Complete-query and training costs

Measurement protocol. Profiling uses one A100-SXM4-80GB, batch size one, ten denoising steps, BF16 backbone and FP32 operator computation, with H = 16 for Fast-WAM and H = 10 for π₀.₅. We report the median of ten queries after warm-up, with observation input and initial noise fixed.

Recomputation included. Complete queries take 216 ms and 69.9 ms, including every refinement step's operator and prefix computation and, for Fast-WAM, action-side Q/K modulation; video-side Q/K is computed once per query and reused.

Parameter and training overhead. Both models add less than 0.001% parameters, without new attention or feed-forward layers; matched Fast-WAM training adds 33 minutes and 1.80 GiB of peak reserved memory per device, with optimizer, learning-rate schedule, and update count unchanged.

Table G. Additional parameters, complete ten-step query latency, and measured Fast-WAM training costs.

(a) Parameters and inference latency

BackboneBackbone params.Added parametersBase → Ours (ms)Diff. (ms)
Fast-WAM6.02B47,378 (≈0.00079%)190 → 21626
π₀.₅3.35B11,375 (≈0.00034%)66.5 → 69.93.4

(b) Fast-WAM training costs

MetricBaselineAT-RoPEIncrease
Training wall-clock time8 h 45 min9 h 18 min33 min
Peak reserved GPU memory per device75.87 GiB77.67 GiB1.80 GiB

3) Component costs and prefix-computation efficiency

Separate component measurements. Table H separates token transport and Q/K modulation at matched H = 16, p = 4, batch size one, hardware, and precision. These standalone timings are not additive components of compiled-query latency.

Table H. Additional parameters and isolated component latency. Q/K timings include shared preparation amortized over 30 layers; parameter counts are branch totals.

ComponentAdded parametersIsolated latency
Fast-WAM token transport4,2153.105 ms/call
Fast-WAM action Q/K5,2360.504 ms/layer
Fast-WAM video Q/K37,9270.530 ms/layer
π₀.₅ token transport11,3750.548 ms/call

Measurement scope. The Fast-WAM token microbenchmark uses sequential scanning, whereas its reported complete query uses parallel scanning; the π₀.₅ input matched to H = 16 is constructed from recorded inputs, with a cost of 0.514 ms/call at native H = 10.

Operator-time attribution. Identifiable π₀.₅ matrix-exponential and midpoint/prefix-labelled kernels account for 1.269 ms (1.82%) and 0.088 ms (0.13%) per query, respectively, with compiler-fused operations remaining grouped.

Parallel-prefix acceleration. Fast-WAM's order-preserving Hillis–Steele scan reduces complete-query latency from 223.54 to 216.33 ms and accelerates its isolated CUDA Graph prefix scan by 2.23× at H = 16. The reported π₀.₅ implementation uses jax.lax.scan; local Q/K modulation adds no prefix scan.

Action-chunk scaling. Increasing Fast-WAM's chunk length eightfold (H = 16 → 128) raises parallel-prefix latency only from 41 to 54 μs, compared with 91 to 549 μs sequentially.

Block-width scaling. Fast-WAM token-module timings vary by less than 1.2% across p ∈ {2, 4, 8, 16}, using trained weights and recorded inputs at p = 4 and norm-calibrated synthetic generators at the other widths (Appendix B.4).

AuthorsOpenReview ↗

Response W2. Aggregate smoothness across tasks and rollouts

Response W2. Aggregate smoothness across tasks and rollouts

1) Expanded coverage and reporting

Trajectory coverage. We extend the selected examples in Figures 1 and 3 to 400 trajectories in total across all four LIBERO suites: Object, Spatial, Goal, and Long.

Metrics and aggregation. We report suite-level mean curvature and high-frequency (HF) energy ratio, with overall improvements summarized as the equal-weight mean of the four within-suite relative reductions.

2) Aggregate results

Consistent curvature reduction. AT-RoPE lowers mean curvature in all four suites, with relative reductions of 6.98–57.63% and an equal-suite mean reduction of 23.16%.

Lower high-frequency energy share. The HF energy ratio decreases in three of four suites, with an equal-suite mean relative reduction of 8.85%.

Long-horizon behavior. On Long, mean curvature decreases by 18.01%, while the HF energy ratio remains nearly unchanged (0.8572% → 0.8612%).

Table I. Aggregate smoothness over 400 trajectories in total across the four LIBERO suites. Mean curvature is in m⁻¹ and HF energy ratio is in %. Lower curvature and HF energy ratio are better; Rel. red. denotes the within-suite relative reduction (%).

SuiteCurvature: BaselineCurvature: AT-RoPERel. red. (%)HF ratio: BaselineHF ratio: AT-RoPERel. red. (%)
Object68.87929.18157.630.28970.229820.68
Spatial123.900115.2576.982.01761.91275.20
Goal96.38986.75010.001.15371.038310.00
Long50.35341.28418.010.85720.8612−0.47
AuthorsOpenReview ↗

Response Q1: Real-world smoothness under comparable execution conditions

Response 1: Real-world smoothness under comparable execution conditions

AT-RoPE improves measured real-world smoothness under matched execution conditions. We use cylinder insertion as a precision-sensitive test case. Compared with π0.5\pi_{0.5}, AT-RoPE reduces mean high-frequency end-effector velocity RMS by 11.9% over full executions and by 17.5% in an equal-duration window. The full-execution high-frequency energy fraction decreases by 44.0%. These results provide quantitative real-world evidence for the motion-quality benefits illustrated in Figures 1 and 3.

Controlled acquisition and common processing. Both methods use the same Franka arm, XHand dexterous hand, objects, and cameras, with matched initial cylinder arrangements. Execution settings are fixed: a 20-Hz control rate, action horizon of 30, 16 executed steps per prediction, and action-smoothing coefficient of 0.5. We record measured arm and hand feedback, issued joint targets, and external and wrist-camera videos. Frame-receipt timestamps and robot-state read midpoints share a host clock; nearest-time matching gives an approximately 24-ms per-recording 95th-percentile frame-to-arm offset. Recorded timestamps determine physical elapsed time and video alignment.

The same processing pipeline is applied to both methods. Measured Cartesian end-effector positions are linearly interpolated onto a 20-Hz grid and differentiated using forward differences. We apply an orthonormal DCT-II to each signed velocity component and sum squared coefficients across XYZ axes, without additional low-pass smoothing. HF velocity RMS measures the absolute magnitude of velocity fluctuations above 2 Hz; HF energy fraction measures their share of non-DC velocity energy. Each execution receives equal statistical weight. Failures, pauses, and recovery attempts are retained, and Figure 7 identifies successful and unsuccessful executions.

Quantitative comparison. The table below summarizes the key results from revised Table 6. Values are mean ± sample standard deviation; all metrics use measured end-effector motion, and HF denotes frequencies above 2 Hz.

Execution windowMetricπ0.5\pi_{0.5}π0.5\pi_{0.5} + AT-RoPE
Full executionHF velocity RMS (mm/s)8.01 ± 1.447.06 ± 0.48
Full executionHF energy fraction (%)2.46 ± 0.801.37 ± 0.29
Initial 20 sHF velocity RMS (mm/s)8.52 ± 1.137.03 ± 0.46
Initial 20 sHF energy fraction (%)1.69 ± 0.441.21 ± 0.20
Initial 20 sMean speed (mm/s)50.27 ± 6.3249.05 ± 3.72
Initial 20 sPath length (m)1.005 ± 0.1260.981 ± 0.074

The initial 20-s window starts at the first issued command for every execution, providing a common observation duration. In this window, HF velocity RMS decreases by 17.5% while mean speeds differ by only 2.4% and path lengths remain similar. This demonstrates lower high-frequency fluctuations at comparable movement levels. AT-RoPE also shows smaller standard deviations for both HF metrics in both windows, indicating more consistent motion quality across initial arrangements. Figure 7 complements the group summaries with paired comparisons under matched initial arrangements.

We have added the acquisition and processing protocol, metric definitions, and results to Section 4.5 and Appendix D.3, including Equation 8, Table 6, and Figure 7. The full-execution and equal-duration results jointly support a real-world smoothness advantage in the evaluated frequency band.

Reviewer_16Q9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Action-Transport RoPE (AT-RoPE), a module that incorporates action-conditioned, order-sensitive feature transformations into generative robot policies. At each denoising step, the current candidate actions parameterize skew-symmetric matrices whose exponentials define orthogonal operators. These operators are composed in temporal order to construct prefix-dependent transformations, which update action tokens through gated residual connections. For WAM backbones, the method additionally applies local action-conditioned modulation to queries and keys after native RoPE. Experiments across four VLA/WAM backbones report improvements on LIBERO, RoboTwin 2.0, robustness benchmarks, and real-world manipulation tasks. Ablations examine prefix composition and the contributions of the token and Q/K pathways.

Critical strengths
  1. Consistent improvements across multiple policy backbones. The method is evaluated on four backbones spanning VLA and WAM policies, with improvements in average success rates in the paired LIBERO and RoboTwin 2.0 comparisons. The substantial gains reported for real-world cylinder insertion and dynamic object manipulation are also encouraging. These results suggest that explicit action-conditioned structure can be a useful addition to generative robot policies.
  2. Ablations address the principal architectural choices.

Comparisons among Local, Sum-prefix, and ordered-prefix variants help assess the contribution of the composition rule. The separate evaluation of token transport and Q/K modulation further clarifies their respective effects. Five independent retraining runs for the prefix variants strengthen the evidence beyond a single training run.

Critical weaknesses
  1. The methodological novelty appears limited, and the incremental contribution over related approaches is not sufficiently established. Content-dependent positional encoding has been explored in CoPE and CARoPE; learned skew-symmetric generators and matrix-exponential rotations are closely related to LieRE; and ordered products of matrix exponentials are a central construction in Path Development Networks. Accordingly, the contribution appears to lie primarily in adapting these ideas to generative action decoding, rather than introducing a new principle of order-sensitive composition. The Local and Sum-prefix ablations support the chosen design relative to these internal variants, but do not sufficiently establish its advantage over suitable adaptations of related methods. A clearer delineation of inherited components and new design choices, supported by more direct comparisons, is needed to substantiate the methodological contribution.

  2. The computational overhead of the additional components is not reported, preventing an assessment of the performance–cost trade-off. The method recomputes action-conditioned matrices, blockwise matrix exponentials, and prefix products at every denoising step. The WAM implementation additionally introduces layerwise Q/K modulation. Although blockwise computation reduces the cost, it does not establish that the overhead is negligible. The paper does not report additional parameter counts, measured inference latency, training overhead, or peak memory usage. This omission matters because some benchmark improvements are modest—for example, the increase from 99.3% to 99.4% for OpenWAM-α on LIBERO and approximately 0.9–1.8 percentage-point improvements in average success on RoboTwin 2.0. Without computational measurements, it is difficult to judge whether these gains justify the added complexity in practical robot deployment.

Questions for authors
  1. Could the authors clarify the relationship between AT-RoPE and CARoPE, LieRE, and Path Development Networks? Beyond conditioning the operators on current candidate actions, which design choices constitute the main methodological contribution, and which experiments isolate their benefits?
  2. Could the authors provide a controlled comparison against a relevant content-conditioned encoding or a prefix-processing module with a comparable parameter budget, using the same backbone, data, and training configuration?
  3. What are the additional parameter counts, end-to-end inference latency per action query, training time, and peak memory usage for each backbone? Please also separate the costs of token transport and Q/K modulation under matched hardware, precision, action-chunk length, and denoising steps.
  4. What fraction of inference time is spent computing matrix exponentials and prefix products? Does the implementation use a sequential or parallel scan, and how does the overhead scale with action-chunk length and block width?
5 comments
AuthorsOpenReview ↗

Response W1 & Q1–Q2. Methodological novelty and controlled comparisons

Response W1 & Q1–Q2. Methodological novelty and controlled comparisons

AT-RoPE uses candidate-action content and execution order to organize the representations used for joint action refinement.

1) Differences from related work

Compared with CoPE/CARoPE. These methods condition positional structure on context or token content; AT-RoPE uses current candidate actions and their execution order to construct transformations applied directly to native action tokens.

Compared with LieRE. LieRE learns positional rotations through skew-symmetric generators; AT-RoPE parameterizes operators with evolving candidate actions and composes their prefixes to transform the representations used for joint refinement.

Compared with Path Development Networks. These networks construct path representations through ordered matrix-exponential products; AT-RoPE rebuilds candidate-action prefixes at each refinement step and uses them to update the representations of the actions being revised.

2) Core methodological contributions

First action-driven operator transport for joint refinement. AT-RoPE is the first to introduce execution-ordered, action-conditioned operator transport into joint action-chunk refinement in generative robot policies. At every refinement step, candidate actions reconstruct the ordered prefixes that directly transform their own token representations through gated residual updates, while the entire chunk remains jointly decoded.

Complementary token and attention interfaces. Token transport supplies preceding-action context, while local post-RoPE Q/K modulation in WAMs conditions attention matching on the actions being compared without another prefix scan. Both interfaces retain the native positional encoding, attention mask, and training objective.

3) Experimental evidence

Related encodings. Under the same Fast-WAM backbone, data, and training configuration, AT-RoPE exceeds CoPE, CARoPE, LieRE, and SoPE on every LIBERO suite, with 9.45–23.20-point average gains (Table A).

Ordered composition. AT-RoPE exceeds Local and Sum-prefix by 1.2 and 0.8 average-success points, respectively, with the same ranking across five runs (manuscript Table 4(a) and Figure 4).

General prefix processing. AT-RoPE exceeds Prefix-GRU with learned readout by 1.4/0.5 points in Clean/Randomized success (Table B).

Direct token transport. AT-RoPE exceeds learned readout using the same operator construction and ordered prefixes by 0.6/0.2 points in Clean/Randomized success, isolating the update interface (Table B).

Composition direction. Reversing only the prefix-product direction, while preserving prefix membership, the half-step factor, gating, action positions, supervision, and execution order, lowers Clean/Randomized success by 0.5/0.4 points (Table B).

Complementary pathways. Combining token transport and Q/K modulation improves average success by 0.7 and 1.2 points over token-only and Q/K-only variants, respectively (manuscript Table 4(b)).

Table A. Related encoding comparisons. Related encoding adaptations within Fast-WAM under the same backbone, data, preprocessing, optimizer, learning-rate schedule, and training-update count. SoPE adds a geometry-aware spherical-coordinate comparison. Success rates (%); Δ Avg. is variant minus AT-RoPE in percentage points.

ConfigurationSpatialObjectGoalLongAvg.Δ Avg.
Fast-WAM98.2100.097.095.297.60−1.00
+CoPE83.695.690.661.682.85−15.75
+CARoPE72.489.880.658.875.40−23.20
+LieRE84.897.088.464.683.70−14.90
+SoPE88.698.291.078.889.15−9.45
+AT-RoPE (ours)98.2100.099.097.298.60—

Table B. Controlled prefix-processing comparisons on Fast-WAM / RoboTwin 2.0. All variants retain the same local Q/K architecture and training/evaluation settings and are trained separately. Success rates (%); parentheses show variant minus AT-RoPE in percentage points.

Token-side variantClean ↑Rand. ↑
Prefix-GRU + learned readout78.7 (−1.4)2.8 (−0.5)
Shared-prefix learned readout79.5 (−0.6)3.1 (−0.2)
Reverse-prefix transport79.6 (−0.5)2.9 (−0.4)
AT-RoPE: forward-prefix transport80.13.3
AuthorsOpenReview ↗

Response W2 & Q3–Q4. Computational cost, scaling, and performance–cost trade-off

Response W2 & Q3–Q4. Computational cost, scaling, and performance–cost trade-off

AT-RoPE achieves substantial robustness and manipulation gains with less than 0.001% additional parameters on both profiled backbones, without adding attention or feed-forward layers.

1) Parameters, inference latency, and training costs

Profiling setup. We use one A100-SXM4-80GB, batch size one, ten denoising steps, BF16 backbone computation, and FP32 operator computation, with native chunk lengths H = 16 for Fast-WAM and H = 10 for π₀.₅.

Additional parameters. AT-RoPE adds only 47,378 parameters to Fast-WAM and 11,375 to π₀.₅, corresponding to approximately 0.00079% and 0.00034% of their backbone sizes.

Complete-query latency. The full ten-step query, including repeated operator and prefix computation, takes 216 ms for Fast-WAM and 69.9 ms for π₀.₅, compared with the measured Fast-WAM baseline of 190 ms and the measured π₀.₅ baseline of 66.5 ms.

Training time and memory. The matched Fast-WAM training comparison adds 33 minutes and 1.80 GiB of peak reserved GPU memory per device, with the optimizer, learning-rate schedule, and training-update count unchanged.

Execution-schedule context. Fast-WAM's 216 ms query occupies 43.2% of the 500 ms interval corresponding to ten executed control steps at 20 Hz.

Table C. Parameter additions, complete ten-step query latency, and measured Fast-WAM training costs.

(a) Parameters and inference latency

BackboneBackbone params.AT-RoPE additionBase → Ours (ms)Diff. (ms)
Fast-WAM6.02B47,378 (≈0.00079%)190 → 21626
π₀.₅3.35B11,375 (≈0.00034%)66.5 → 69.93.4

(b) Fast-WAM training costs

MetricBaselineAT-RoPEIncrease
Training wall-clock time8 h 45 min9 h 18 min33 min
Peak reserved GPU memory per device75.87 GiB77.67 GiB1.80 GiB

2) Component costs and scaling

Separate module profiling. Table D separates token transport and Q/K modulation at matched H = 16, p = 4, batch size one, hardware, and precision; these standalone timings are distinct from the integrated compiled-query latency.

Shared computation. Action Q/K runs at every denoising step, whereas video Q/K is computed once per query and reused; per-layer timings include shared preparation amortized over 30 layers.

Table D. Additional parameters and isolated component latency. Parameter counts are branch totals; Q/K timings include shared preparation amortized over 30 layers.

ComponentAdded parametersIsolated latency
Fast-WAM token transport4,2153.105 ms/call
Fast-WAM action Q/K5,2360.504 ms/layer
Fast-WAM video Q/K37,9270.530 ms/layer
π₀.₅ token transport11,3750.548 ms/call

Module measurement scope. The Fast-WAM token microbenchmark uses the sequential standalone module, and the π₀.₅ input matched to H = 16 is constructed from recorded inputs; at native H = 10, its token-transport call takes 0.514 ms.

Matrix-exponential and prefix costs. For π₀.₅, identifiable matrix-exponential and midpoint/prefix-labelled kernels account for 1.269 ms (1.82%) and 0.088 ms (0.13%) per query, respectively. Fused operations remain grouped; Fast-WAM's compiled trace does not separately resolve these categories, so its prefix cost is assessed through the scan measurements below.

Reported scan implementations. Fast-WAM uses an order-preserving Hillis–Steele parallel scan, while π₀.₅ uses sequential jax.lax.scan; local Q/K modulation requires no additional prefix scan.

End-to-end acceleration. Replacing Fast-WAM's sequential scan with the parallel scan reduces complete-query latency from 223.54 to 216.33 ms, saving 7.21 ms per query in the fixed-input comparison.

AuthorsOpenReview ↗

Response W2 & Q3–Q4. Computational cost, scaling, and performance–cost trade-off

Isolated scan acceleration. At H = 16, parallel scans provide 2.23–2.90-fold speedups (Table E); the JAX parallel variant is benchmarked separately from the sequential implementation used for the reported 69.9 ms query.

Table E. Isolated prefix-scan latency at H = 16, p = 4, and batch size one.

Implementation / modeSequential (μs)Parallel (μs)Speedup
Fast-WAM / CUDA Graph91.1440.962.23×
Fast-WAM / eager989.18424.962.33×
π₀.₅ / JAX315.55108.862.90×

Action-chunk scaling. Increasing Fast-WAM's chunk length eightfold (H = 16 → 128) raises parallel-prefix latency only from 41 to 54 μs, versus 91 to 549 μs sequentially, with prefix dependency depth reduced from O(H) to O(log H).

Complete token-module scaling. In separate sequential-module tests, H = 8 → 128 increases token-module latency from 2.66 to 9.35 ms for Fast-WAM and from 0.47 to 1.74 ms for π₀.₅.

Block-width scaling. Across p ∈ {2, 4, 8, 16}, Fast-WAM token latency varies by less than 1.2%, Q/K timings are approximately flat, and π₀.₅ shows small, non-monotonic changes. The p = 4 test uses trained weights and recorded inputs; other widths use norm-calibrated synthetic generators.

Numerical agreement. Parallel scans outperform sequential scans throughout H ∈ {8, 16, 32, 64, 128, 256}, with a maximum absolute deviation of approximately 6.9 × 10⁻⁶ from the FP64 sequential reference.

3) Performance gains and repeatability

Robustness and manipulation gains. Fast-WAM improves LIBERO-Plus success from 51.5% to 72.2% (+20.7 points), while π₀.₅ improves real-world cylinder insertion from 20% to 90% (+70 points) (the manuscript's LIBERO-Plus table and Figure 5).

Main-benchmark repeatability. Five-run LIBERO results are 97.51 ± 0.11% for π₀.₅, 98.65 ± 0.15% for Fast-WAM, and 99.37 ± 0.06% for OpenWAM-α.

Benefits beyond near-ceiling averages. OpenWAM-α reaches 98.64 ± 0.17% on LIBERO-Long versus its reported 98.2% baseline; on RoboTwin, Fast-WAM and OpenWAM-α reach 41.79 ± 0.22% and 69.94 ± 0.30%, versus the reported 39.9% and 69.0% references.

These measurements establish a favorable performance–cost trade-off: a compact parameter addition, quantified training and inference costs, efficient parallel prefix computation, and repeatable task-performance benefits.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

We thank Reviewers 16Q9, e93f, and f6aS for their constructive feedback. We have clarified the methodological contribution and expanded the evidence through direct comparisons, computational measurements, broader smoothness analysis, targeted robustness tests, and repeated main-benchmark evaluations. We summarize the principal updates below; the individual responses provide detailed protocols and results.

1. Methodological novelty and direct comparisons — 16Q9

Action-driven joint refinement. AT-RoPE is the first to introduce execution-ordered, action-conditioned operator transport into joint action-chunk refinement in generative robot policies. Candidate actions determine the transformations used to refine their own representations, distinguishing this decoding mechanism from contextual positional encoding, learned positional rotations, and path representations.

Controlled evidence. Under the same Fast-WAM backbone, data, and training configuration, AT-RoPE exceeds CoPE, CARoPE, LieRE, and SoPE adaptations on every LIBERO suite, with 9.45–23.20 pp higher average success. Additional Prefix-GRU, shared-prefix learned-readout, and reverse-prefix comparisons support the benefits of direct token transport and the specified composition direction beyond the original Local/Sum-prefix ablations.

2. Measured computational costs and parallel execution — 16Q9 and e93f

Complete-query profiling. We report additional parameters and complete ten-step query latencies on one A100-SXM4-80GB, with batch size one, BF16 backbone computation, and FP32 operators; native chunk lengths are 16 for Fast-WAM and 10 for π₀.₅.

BackboneAdded parameters (backbone fraction)Baseline reference → AT-RoPE query latency
Fast-WAM47,378 (0.00079%)190 (reported) → 216 ms
π₀.₅11,375 (0.00034%)66.5 → 69.9 ms

Training and prefix computation. Matched Fast-WAM training adds 33 minutes and 1.80 GiB of peak reserved GPU memory per device, with the update count unchanged. Order-preserving parallel scanning reduces its complete-query latency from 223.54 to 216.33 ms; increasing chunk length from 16 to 128 raises isolated parallel-prefix latency only from 41 to 54 μs.

3. Broader smoothness and targeted robustness tests — e93f and f6aS

Aggregate smoothness — e93f. The analysis now covers 400 trajectories in total across all four LIBERO suites: mean curvature decreases in every suite, with an equal-suite mean relative reduction of 23.16%. The HF energy ratio decreases in three suites, with an equal-suite mean relative reduction of 8.85%; on Long, it remains nearly unchanged alongside an 18.01% curvature reduction.

Candidate correction — f6aS. A controlled intermediate-prefix perturbation study on RoboTwin uses 900 cases per model and yields 99.642% normalized recovery with AT-RoPE versus 99.630% for the baseline. The paired recovery difference is small and positive, with its 95% confidence interval above zero.

Demonstration variation — f6aS. With 500 trials per configuration, AT-RoPE retains an advantage after training with task-irrelevant movement variation: 80.2% versus 77.4% success, with a smaller observed decrease of 0.8 versus 1.4 pp.

4. Five-run main-benchmark results — 16Q9 and f6aS

Coverage beyond the prefix ablation. We report five AT-RoPE runs for each of four backbones on both LIBERO and RoboTwin 2.0. Values below are benchmark-average success rates (%) reported as mean ± sample standard deviation.

AT-RoPE backboneLIBERORoboTwin 2.0
π₀.₅97.51 ± 0.1159.34 ± 0.15
StarVLA-π / StarVLA97.37 ± 0.1625.82 ± 0.17
Fast-WAM98.65 ± 0.1541.79 ± 0.22
OpenWAM-α99.37 ± 0.0669.94 ± 0.30

5. Matched real-world comparisons and execution behavior — e93f

Matched real-world smoothness — e93f Q1. Using the same robot platform, matched initial arrangements, and identical control and smoothing settings, AT-RoPE reduces high-frequency end-effector velocity RMS by 11.9% over complete cylinder-insertion executions and 17.5% over the common initial 20-second window. Mean speeds in the latter remain comparable (50.27 versus 49.05 mm/s), providing measured evidence of smoother execution at similar movement speeds (Section 4.5; Appendix D.3).

Execution behavior and failure reduction — e93f Q2. In the matched cylinder-insertion analysis, executions with tipping followed by incomplete recovery or insertion decrease from 70% to 10%. Timestamped baseline–AT-RoPE comparisons show more reliable progression through grasp, transfer, insertion, and release, including earlier release and withdrawal in a pair where both methods succeed (Appendix D.4; Figure 8).

AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank Reviewers 16Q9, e93f, and f6aS for their careful reading and constructive feedback. We have revised the manuscript accordingly and provide point-by-point responses below, addressing related weaknesses and questions together where appropriate.

For ease of reference, the corresponding revisions in the main text and appendices are highlighted in blue for Reviewer 16Q9, green for Reviewer e93f, and purple for Reviewer f6aS. Revisions that jointly address comments from two or more reviewers are highlighted in orange.

AuthorsOpenReview ↗

A Kind Follow-up on Rebuttal

Dear Reviewers,

Thank you again for the time and care you have devoted to reviewing our paper. We sincerely appreciate your thoughtful comments, which have helped us recognize where our work needs clearer explanations and stronger support.

We have carefully considered your feedback and done our best to address your concerns in our rebuttal. We understand that reviewing requires substantial effort alongside many other commitments. If your schedule permits, we would be very grateful if you could take another look at our responses and share any remaining concerns.

We recognize that our responses may not yet fully resolve every point you raised, and we would greatly appreciate the opportunity to clarify anything that remains unclear. Your further guidance would be invaluable in helping us improve the work.

Thank you very much for your time, patience, and consideration.

With sincere appreciation, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.