acceptodds
Under review as a conference paper at ICLR 2027

Dual-Process Atomic Skill Learning: Decoupling Semantic Reasoning and Real-Time Control

Abstract

Language-conditioned imitation learning must discover reusable skills from demonstrations to generalize to multi-step instructions. Joint training of discrete skill hierarchies can produce concentrated codebook usage and unstable skill representations. We propose Dual-Process Atomic Skill Learning (DASL), which independently parameterizes the semantic update interval and low-level trajectory horizon, supervises each shared skill with actions across its interval, and regularizes skill-conditioned latent trajectories with a diffusion module used only during training. A slow Option Transformer predicts vector-quantized skills, while a Decision Transformer generates actions from full-rate observations without diffusion sampling at inference. Evaluations on LOReL, Franka Kitchen, a six-task CALVIN-D state-based protocol, BabyAI, and a real robot demonstrate improved skill acquisition and compositional performance in these settings. Direct codebook diagnostics and an isolated interval-supervision ablation support the proposed training mechanism, while the deterministic inference path preserves efficient control.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Language-conditioned Imitation Learnin, Skill Learning, Hierarchical Imitation Learning, Asynchronous Coordination.

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_yxBCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

DASL adds a denoising task during training to a LISA-style robot policy. It also changes how often skills are selected and how action errors train the skill selector. The denoising model is removed during inference. The policy has two networks: one selects a discrete skill from language and observations, and another predicts actions using that skill and current observations. DASL lets the skill-update interval and low-level trajectory horizon be configured separately. Each selected skill receives feedback from the actions that share it.

Critical strengths

The training-only diffusion objective is an interesting design choice. It provides an additional representation-learning signal without requiring iterative diffusion sampling during execution.

Critical weaknesses
  1. The comparison with LISA does not isolate the methodological contribution. Appendix F.7 uses (I=2,H=5) for DASL on LOReL state observations, but (I=H=10) for LISA. This changes the temporal configuration as well as the proposed learning components. A LISA baseline with matched skill duration and controller context is needed to establish how much improvement remains beyond these settings.

  2. The LISA reproduction also needs explanation. Table 1 reports 12.78% success on seen state-based LOReL instructions, whereas the original LISA paper reports 66.7% on its seen state-based evaluation.

Questions for authors
  1. DASL uses I=2I=2, while the reproduced LISA-state baseline uses I=H=10I=H=10. Can you compare them with the same skill update frequency, controller context length, and tuning budget?

  2. On seen LOReL-state instructions, you report 12.78%, whereas the original LISA paper reports 66.7%. Which differences in data, training, or evaluation explain this gap?

  3. The denoising target consists of trainable observation–skill embeddings, and clean skill embeddings are also supplied as conditioning. What shows that this learns better skills, rather than merely making the features easier to denoise? Can you compare against an otherwise identical denoiser without skill conditioning?

  4. Changing II changes both the observations entering the planner and how long each selected skill is held. Can you vary these separately to identify which change produces the benefit?

  5. The component ablations and codebook diagnostics use one fixed seed. Please report uncertainty for the diffusion and interval-supervision comparisons that support the central claims.

7 comments

We thank Reviewer yxBC for the time and effort devoted to reviewing our paper and for the valuable comments. Below, we address the weaknesses and questions raised in the review.

W1 & Q1. Fair Comparison between DASL and LISA

We thank the reviewer for highlighting the importance of temporal settings and tuning budgets. We have added a LISA comparison with the skill-update interval, controller inference context, and tuning budget matched to DASL.

First, we matched the skill-update interval and controller inference context. The LISA-state baseline in Table 1 uses I=H=10I=H=10, following the original paper's hyperparameter table. In the additional comparison, both methods use I=2I=2 and a DT inference context of five control steps.

Second, both methods use the same candidate hyperparameter configurations and tuning budget. The following results are macro-averages over the five language categories in Table 1, reported as mean ± SD over three training seeds. The DASL result is shared with Table 1.

MethodSkill-update interval IIDT inference contextState SR (%)
LISA2517.63 ± 3.87
DASL2554.22 ± 4.65

Under these matched settings, DASL's mean SR exceeds LISA's by 36.59 percentage points. This comparison provides evidence of gains beyond the original temporal-configuration differences.

LISA's skill selections remain concentrated on a few codebook entries after matching the settings. Together with task success, this observation motivates examining discrete skill utilization during joint training. Table 29 in Appendix F.8 reports this comparison and its settings.

W2 & Q2. Explaining the Lower LISA Reproduction Results

We thank the reviewer for raising this concern. Table 1 reports a LISA seen-state SR of 12.78 ± 9.02%, compared with 66.7 ± 5.2% in the original paper [1]. We clarify our reproduction settings and the training behavior observed in these runs.

First, we use the public LOReL Sawyer data and evaluate the original LISA paper's six manipulation tasks. These are opening and closing the drawer, turning the faucet left and right, moving the black mug right, and moving the white mug down. The dataset contains 50,000 trajectories of 20 steps each. The 12.78% result corresponds specifically to seen-state instructions; the other language categories are reported separately. It therefore differs in scope from the five-category averages in the matched comparison above. Our results are mean ± SD over three training seeds.

We use the official LISA implementation, starting from the original paper's hyperparameter table and performing hyperparameter search and tuning. Appendix C describes the baseline implementation and hyperparameter starting point, while Table 12 identifies the sources of the baseline results.

Second, our codebook diagnostics document skill underutilization and index collapse. Across three training seeds, the LISA baseline activates 4.33 ± 0.58 of the 20 codebook entries, with normalized usage entropy of 0.468 ± 0.036 (Table 5). These measurements quantify the concentration of skill assignments in our reproduction. Concentrated assignments also persist after matching the skill-update interval, controller context, and tuning budget.

These diagnostics identify a training issue associated with the low success rate in our runs. They do not by themselves establish which implementation or evaluation differences account for the full gap from the original result.

LCSD and LADS provide related evidence about LISA reproduction and codebook training. LCSD uses the official LISA code and reports 6.7 ± 3.3% success under its LOReL-state setting, associating the low performance with index collapse. Adding code reinitialization raises success to 43.4 ± 0.2% [2].

LADS also reports index collapse when reproducing LISA. Updating the vector-quantization implementation partially alleviates the issue, yielding 38.6 ± 3.9% on LOReL-state atomic-seen instructions [3]. This result is a LISA reproduction with a modified VQ implementation, as described in Appendix D of the LADS paper.

These studies use their own experimental settings. Their results show that VQ implementation and codebook training procedures can affect LISA reproduction, and provide context for the concentrated assignments observed in our runs. Appendix C documents these training observations alongside the baseline sources; Appendix F.8 reports the additional comparison under matched temporal settings.

Q3. Does the Diffusion Module Actually Promote Skill Learning?

We thank the reviewer for raising this question. We have added an otherwise identical denoiser without explicit skill conditioning and evaluated downstream task success alongside representation-scale and codebook diagnostics. We also compare against no diffusion and a stop-gradient variant.

The denoising target is sampled noise. In Equation (7), the prediction target is ϵ\epsilon. The trainable observation–skill embedding τ0\tau_0 is the clean latent representation used to construct the noisy input.

We isolate the explicit conditioning pathway by replacing the U-Net's skill condition with zeros. The fused-representation construction, denoising network architecture, and other training settings remain unchanged. The fused input itself still contains observation and skill information. In the stop-gradient variant, we detach τ0\tau_0 before adding noise while keeping the skill-conditioning pathway trainable; action and VQ supervision remain active.

Table 6 in Section 4.4 reports the following LOReL-state results as mean ± SD over three training seeds. Fused Embedding Norm is the mean L2L_2 norm over valid fused embeddings. Full DASL's success rate is shared with Table 1.

VariantDiffusionExplicit skill conditioningStop-gradient on τ0\tau_0Fused Embedding NormState SR (%)
B (w/o diffusion)×——8.9 ± 0.539.58 ± 2.53
Full DASL✓✓×8.8 ± 0.454.22 ± 4.65
w/o explicit skill conditioning✓××8.6 ± 0.547.07 ± 4.03
Stop-gradient on τ0\tau_0✓✓✓9.0 ± 0.451.83 ± 3.52

Explicit skill conditioning improves mean downstream task success in this comparison. With the denoiser architecture and other settings fixed, Full DASL achieves 54.22% mean SR, compared with 47.07% without explicit skill conditioning, a gain of 7.15 percentage points. This directly addresses the requested conditioning control using task execution as the outcome.

The variant without explicit conditioning also exceeds the no-diffusion Variant B by 7.49 percentage points. The stop-gradient variant reaches 51.83% mean SR, which is 12.25 percentage points above B and 2.39 points below Full DASL. Thus, downstream gains remain when diffusion gradients through the clean-input path are blocked while the skill-conditioning path remains trainable. All variants predict actions without diffusion denoising at inference.

Fused embedding norms characterize representation scale. Mean norms range from 8.6 to 9.0, with similar values for Full DASL and Variant B (8.8 and 8.9). These measurements describe the average scale of the fused representations.

Codebook diagnostics provide complementary evidence about skill use. Table 5 reports the following results over three training seeds:

VariantActive / 20HnormH_{\mathrm{norm}}Top-1 RatioSoft MI
B (w/o diffusion)19.00 ± 1.000.874 ± 0.0310.125 ± 0.0220.258 ± 0.011
Full DASL19.00 ± 0.000.936 ± 0.0200.096 ± 0.0250.273 ± 0.010

Both variants activate 19 codebook entries on average. Full DASL has higher mean usage entropy, a lower mean Top-1 usage ratio, and higher mean Soft MI, indicating more uniform skill use and stronger statistical association between language instructions and skill assignments.

Together, the task results and codebook diagnostics provide evidence that auxiliary diffusion benefits control performance and changes how skills are used. The conditioning and gradient-path controls are included in Section 4.4 and Table 6, and the multi-seed codebook diagnostics appear in Table 5.

Q4. Effects of the Skill-Update Interval

We thank the reviewer for pointing out that changing II affects both planner observations and skill duration. We configure the planner observation sampling interval SS and skill holding duration DD independently to examine each factor while holding the other fixed.

At each skill update, the planner reads causally available historical observations sampled at interval SS, and holds the selected skill for DD control steps. On LOReL-state, we fix H=5H=5, the DT inference context at five control steps, and the planner’s accessible history span. The low-level controller receives the current observation at every control step. Other training and evaluation settings remain unchanged. Results are mean ± SD over three training seeds.

Planner observation interval SSSkill duration DDState SR (%)
1139.35 ± 4.36
1247.04 ± 3.41
2254.22 ± 4.65

First, we fix observation sampling and vary skill duration. At S=1S=1, increasing DD from one to two steps raises mean SR from 39.35% to 47.04%, a gain of 7.69 percentage points. Holding skills for two steps improves mean performance while the planner history retains observations from every control step.

Second, we fix skill duration and vary observation sampling. At D=2D=2, increasing SS from one to two steps raises mean SR from 47.04% to 54.22%, a gain of 7.18 percentage points. Sparser planner observations improve mean performance at the same skill duration, while the low-level controller continues to receive observations at every step.

Both longer skill duration at S=1S=1 and sparser planner observations at D=2D=2 improve mean performance in the tested configurations. Table 21 in Appendix F.1 reports these experiments and specifies the fixed and varied factors.

Q5. Reliability of Component Ablations and Codebook Diagnostics

We thank the reviewer for recommending multi-seed evaluation. We have repeated the component ablations and codebook diagnostics using the same three training seeds under a common training and evaluation protocol.

The LOReL-state results report mean ± SD. Full DASL’s SR is from Table 1; codebook diagnostics are from the additional experiments.

MethodActive / 20HnormH_{\mathrm{norm}}Top-1 RatioSoft MISkills / InstructionState SR (%)
Variant A20.00 ± 0.000.936 ± 0.0230.104 ± 0.0160.246 ± 0.0138.74 ± 0.2829.14 ± 3.26
Variant B19.00 ± 1.000.874 ± 0.0310.125 ± 0.0220.258 ± 0.0115.34 ± 0.1939.58 ± 2.53
Variant C20.00 ± 0.000.932 ± 0.0270.105 ± 0.0180.252 ± 0.0128.53 ± 0.3133.46 ± 3.68
DASL w/o IS15.00 ± 0.000.742 ± 0.0500.312 ± 0.0600.247 ± 0.0126.17 ± 0.1345.92 ± 4.00
Full DASL19.00 ± 0.000.936 ± 0.0200.096 ± 0.0250.273 ± 0.0106.22 ± 0.1154.22 ± 4.65

Effect of diffusion. Full DASL exceeds Variant B by 14.64 percentage points in mean SR. Both activate 19 codes on average, while Full DASL has higher mean normalized usage entropy (0.874 → 0.936), lower mean Top-1 ratio (0.125 → 0.096), and higher mean Soft MI (0.258 → 0.273). Higher task success accompanies more uniform skill usage and stronger language–skill association.

Effect of interval-level supervision. Full DASL exceeds DASL w/o IS by 8.30 percentage points in mean SR. Mean active-code count increases from 15 to 19, entropy from 0.742 to 0.936, and Top-1 ratio decreases from 0.312 to 0.096. The w/o IS variant preserves forward skill sharing, the update interval, and all valid action losses, restricting only action-gradient paths into the skill branch. This comparison examines how action feedback to skills affects training outcomes.

The multi-seed results retain the main ablation trends, with standard deviations reporting variation across training seeds. We have consolidated the component ablations and diagnostics from the original Tables 5 and 16 into main-text Table 5, with analysis, protocol details, and result sources.

Conclusion

We have addressed the comparison settings, LISA reproduction, diffusion controls, temporal factors, and multi-seed reliability. The corresponding revisions are marked in blue in the manuscript.

Thank you again for your time and valuable feedback. We hope our response has addressed your concerns.

References

[1] Garg et al. LISA: Learning Interpretable Skill Abstractions from Language. NeurIPS, 2022.

[2] Ju et al. Rethinking Mutual Information for Language Conditioned Skill Discovery on Imitation Learning. ICAPS, 2024.

[3] Jiang et al. Discrete Latent Plans via Semantic Skill Abstractions. ICLR, 2025.

Reviewer_BKXhOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

DASL combines a slow discrete-skill planner with a fast action policy. It independently configures skill-update intervals and trajectory horizons, and uses diffusion only during training. The paper reports improved skill utilization and task performance across several simulated benchmarks and real-robot tasks.

Critical strengths
  • The design supports efficient inference. Training-only diffusion avoids iterative sampling during deployment, while the action policy retains observation feedback at every step.

  • Evaluation covers four simulated benchmarks and a physical robot. Component ablations and quantitative codebook diagnostics provide useful evidence beyond task success alone.

Critical weaknesses

Evidence for compositional generalization remains limited. Real-robot composite tasks have training demonstrations, and success on complete unseen four-subtask Kitchen sequences is 0%. The Kitchen gains mainly concern partial completion.

Questions for authors

see weaknesses

1 comment

We thank Reviewer BKXh for the time and effort devoted to reviewing our paper and for the valuable comments. Below, we address the concern about the experimental evidence for compositional generalization.

W1 & Q1. Experimental Validation of Compositional Generalization

We thank the reviewer for pointing out the limitations of our current compositional generalization experiments. We agree that generalization to unseen compositions should be distinguished from the ability to execute demonstrated compositional tasks. Our experiments cover compositional instruction execution on LOReL, evaluation on unseen compositions in Kitchen, and multi-stage task execution on a physical robot. We clarify the settings and results below.

First, the LOReL experiments provide additional evidence of compositional instruction execution. In the LOReL Sawyer compositional task evaluation, DASL achieves success rates of 23.61 ± 2.10% and 32.50 ± 3.99% with state and image observations, respectively (Table 4). With state observations, DASL’s mean success rate is close to that of LADS, which achieves 23.11 ± 2.76%. With image observations, DASL’s mean success rate exceeds SkillDiffuser’s 25.21 ± 2.70% by 7.29 percentage points.

Second, the Kitchen experiments evaluate execution on unseen compositions by holding out complete task combinations. We hold out three four-subtask compositional instructions for testing, while ensuring that every atomic subtask appears in the training data (Appendix B.2). Tables 16 and 17 report the corresponding results with state and image observations.

DASL’s current results on unseen combinations primarily reflect partial completion at K=1K=1 and K=2K=2. It has not completed an entire unseen four-subtask sequence with either state or image observations in this evaluation.

The baseline results further illustrate the difficulty of completing these unseen combinations. For full four-subtask completion (K=4K=4), LangDT, LISA, and LADS achieve 0%, 0%, and 0.9% with state observations, and 0%, 0.2%, and 0.4% with image observations, respectively. These results show that executing complete unseen four-subtask sequences remains challenging across the evaluated methods.

The three held-out combinations have different degrees of prefix overlap. The first two share prefixes of three and two subtasks with training trajectories, respectively, so experience with these subsequences may support completion of familiar prefixes. The third combination examines execution without a shared two-subtask prefix. We have clarified these overlaps and explicitly distinguished partial completion rates from full-combination success rates in the results analysis.

Finally, the physical robot experiments validate multi-stage execution with compositional demonstrations. We collected 20 training demonstrations for each of the two compositional tasks and conducted 20 test trials per task (Section 4.6 and Appendix G, Table 30). DASL succeeded in 15 and 16 trials, corresponding to success rates of 75% and 80%, respectively. LISA achieved 50% and 45%, giving differences of 25 and 35 percentage points on the two tasks. These results demonstrate sequential manipulation and task transitions in a physical environment under training that includes demonstrations of the corresponding compositional tasks.

Conclusion

Following your suggestion, we have clarified the capabilities evaluated by each experiment: LOReL provides additional results on compositional instruction execution, Kitchen demonstrates partial completion on unseen compositions, and the physical robot experiments validate multi-stage execution with compositional demonstrations. Executing complete unseen four-subtask sequences remains challenging. Longer task sequences and online adaptation are directions for future research. The corresponding revisions are marked in blue in the manuscript.

Thank you again for your time and valuable feedback. We hope our response has addressed your concerns.

Reviewer_X4WNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies discrete skill discovery for language-conditioned imitation learning without subtask annotations. Building on LISA, it observes that jointly trained discrete skill hierarchies can exhibit concentrated codebook usage and unstable skill representations. The proposed method, DASL, uses a slow Option Transformer that predicts a vector-quantized skill every I control steps from language and subsampled observations, and a Decision Transformer that adds the held skill embedding to full-rate observation embeddings and predicts an action at every step. Three components are emphasized: choosing the skill update interval independently of the low-level horizon, supervising each shared skill with the action losses of all steps in its interval, and an auxiliary skill-conditioned latent diffusion loss used only during training. The method is evaluated on LOReL, Franka Kitchen, a six-task CALVIN protocol, BabyAI, and a real robot. It reports higher average success than LISA, LADS, SkillDiffuser, and LCSD in most settings, with LADS remaining ahead on Kitchen with state inputs, and includes ablations and codebook diagnostics.

Critical strengths
  1. The problem is relevant. Codebook collapse and weak language–skill association are real issues in VQ-based skill learning. Studying them through direct diagnostics such as active codes and code-usage entropy, rather than success rate alone, is a good direction.

  2. The skill embedding observation is interesting. In the component ablation, adding the skill embedding to the observation stream raises the number of active codes from 4 to 20 and normalized entropy from 0.451 to 0.942, and it improves state success from 19.08% to 28.34%. The image gain is small and the ablation uses a single seed, but this is the largest single shift in the codebook diagnostics.

  3. The evaluation is broad and result provenance is transparent. The paper covers several simulated benchmarks and a real robot, and Table 11 clearly lists which baseline numbers are cited and which were rerun.

Critical weaknesses
  1. Two of the stated contributions are close to existing mechanisms. Interval-level supervision is not a new training signal. When a skill is held for several steps, backpropagating the action loss through the straight-through estimator already sends the gradients of all those steps to the shared skill, and LISA does the same over its horizon. Eq. 6 describes this gradient, up to normalization, rather than adding an objective, and it does not appear in the total loss or in Algorithm 1. The "w/o IS" ablation restricts the supervision reaching the slow branch to a single step per interval, so it removes part of a gradient that ordinary backpropagation already provides. Its drop in success rate and codebook usage shows that this gradient matters, but not that full-interval supervision is a new contribution. Separately, shaping a directly predicting policy with a diffusion objective used only during training has precedent, for example Diffusion Model-Augmented Behavioral Cloning (ICML 2024), which is not cited. The paper should discuss how its jointly trained latent-space variant differs.

  2. The decoupling of the skill interval and the horizon is not used in all settings and is not cleanly isolated. Kitchen uses I = H = 20, which is the shared-horizon setting the paper contrasts DASL with. In the method description and Algorithm 1, H enters only through the diffusion window. The Decision Transformer context length is not reported for any benchmark, and the interval used by Variant A is not stated. As a result, the step from Variant A to Variant B is hard to distinguish from simply using a shorter skill interval. The interval sensitivity study covers only LOReL, reports no variance or seed count, and shows large, irregular changes at the task level.

  3. The latent diffusion regularizer may admit degenerate solutions. The diffusion target is the learned fused embedding, which contains the skill embedding, and the denoiser is conditioned on the same skills with no stop-gradient indicated on the target. The loss could therefore in principle be lowered by shrinking the fused embeddings or making them predictable from the skill condition, rather than by modeling trajectories. The paper does not analyze this, and the loss weight differs by a factor of fifty between the state and image settings on LOReL. The appendix attributes the gain to the ability of diffusion to model multimodal trajectory distributions, but this is not tested: no samples are drawn from the denoiser, and the comparison only includes a VAE and an autoencoder regularizer.

  4. The LOReL baselines appear weaker than prior reports. The original LISA paper reports 66.7% on LOReL with state inputs, whereas this paper obtains 11% to 19%. LADS also reports being unable to reproduce 66.7%, but its own rerun of LISA reaches about 38%, still far above this paper's reproduction, while the LangDT reproduction here is close to its published value. The paper also reruns LADS on LOReL instead of citing its published results, and the rerun atomic results are clearly lower than those LADS reports. With state inputs, DASL's own seen and composite results are numerically below LADS's published numbers, although the averaging and possibly the splits differ; with image inputs DASL exceeds them. Since the LOReL ablation narrative and much of the reported gain are framed relative to these reproduced baselines, the paper should explain the discrepancies and compare against published numbers under a matched protocol.

  5. The ablations and cross-table numbers are not reliable enough to support the finer claims. All component ablations and codebook diagnostics use a single fixed seed, and the appendix sweeps report no seed count or variance, while the three-seed main results vary by several points. The default DASL setting on LOReL state appears as values between 53 and 61.9 across tables. Appendix B.1 attributes this to different protocols but does not say how they differ. Table 21 compares a baseline from one protocol with a diffusion result from another. Two claimed improvements are smaller than one standard deviation: the compositional state result against LADS and the Kitchen image result on the seen split. The synergy argument relative to an additive reference is built from four single runs, so it is suggestive rather than established.

  6. The dual-process framing is not reflected at inference. In the latency breakdown, the option selector is the largest single cost, and unlike the language encoder it is not marked as amortized, even though skills are updated only every few steps. It is unclear whether its output is cached between updates. DASL is also slightly slower than LISA under matched execution, and the further speedups in Appendix H come from model-agnostic engineering reported only for DASL. The paper itself notes that this execution pattern is already established in hierarchical policies, so in my view the contribution would be clearer if framed as the training scheme it actually proposes.

Questions for authors
  1. Can you explain the LOReL baseline discrepancies? LISA on LOReL state falls well below both its published result and LADS's rerun, while LangDT is close to its published number. Why is that, and why is LADS rerun instead of cited? How does DASL compare against the published numbers under a matched protocol?

  2. How is the w/o IS variant implemented? Is the gradient from the other steps simply blocked, which step is kept, and does rescaling the kept gradient change the conclusion?

  3. What prevents the diffusion loss from shrinking or collapsing the fused embeddings? Have you tried a stop-gradient on the diffusion target, or the same loss without skill conditioning? Reporting embedding norms during training would also help.

  4. Can you clarify the interval settings? Why does Kitchen use I = H? Could you show Kitchen with I different from H, an I = H control on CALVIN and BabyAI, the interval used by each ablation variant, and the Decision Transformer context length for every benchmark?

  5. Can the ablations and sweeps be rerun with multiple seeds under one protocol? Please also state how the configurations behind the different DASL numbers differ, and report Table 21 against the matched baseline.

  6. Is the option selector executed at every control step at inference, or is its output cached between updates? What is the latency when it runs only at update steps?

9 comments

We thank Reviewer X4WN for the time and effort devoted to reviewing our paper and for the valuable comments. Below, we address the weaknesses and questions raised in the review.

W1 & Q2. How Is the w/o IS Ablation Implemented?

Thank you for highlighting standard backpropagation and suggesting gradient rescaling. DASL jointly combines full-rate skill–observation fusion, independently configured temporal parameters, and skill-conditioned latent diffusion.

Interval-level supervision describes action feedback through shared skills. Standard backpropagation through the straight-through estimator aggregates action gradients into a skill shared across multiple timesteps. This also applies to skills held over multiple steps in LISA. Equation (6) describes gradients from the action-prediction loss, with no additional objective or separate IS loss in Algorithm 1. We have revised the contribution statement to frame IS as an analysis of action-gradient paths.

The w/o IS control preserves forward skill sharing, the update interval, model capacity, and every valid action loss. Only the first valid skill-input position per interval retains an action-gradient path to the slow branch; other repeated inputs are detached before skill projection. VQ and diffusion gradient paths remain unchanged. With causal attention, later action losses can still backpropagate through retained inputs (Appendix F.7).

The rescaling control uses the same mask and multiplies retained action-to-skill gradients by the interval's valid length: 2 for complete intervals with I=2I=2, and the actual valid length for truncated intervals.

Table 5 reports three-seed mean ± SD. State SR is 45.92 ± 4.00% for w/o IS, 51.67 ± 3.20% with rescaling, and 54.22 ± 4.65% for Full DASL. Rescaling raises mean active-code count from 15 to 17 and usage entropy from 0.742 to 0.828; Full DASL reaches 19 and 0.936. The multiplier is interval length; actual gradient norms depend on the backpropagated gradients.

Relationship to DBC. DBC [1] trains a diffusion model on expert state–action pairs, then fixes it to guide policy learning alongside behavioral cloning. DASL applies skill-conditioned denoising to learned observation–skill latent trajectories used for action prediction, jointly optimizing discrete skills, the denoiser, and the action policy. Both use diffusion during training and execute without iterative diffusion sampling. Section 2.2 now describes this relationship.

W1, W3 & Q3. How Does DASL Address Shrinkage or Collapse of Fused Embeddings under the Diffusion Loss?

Thank you for raising this issue. We examine representation scale and task performance through conditioning and gradient-path controls.

The denoising target is sampled noise. In Equation (7), the target is ϵ\epsilon, while τ0\tau_0 is the trainable clean fused representation used to construct the noisy input. Action prediction, VQ, and diffusion are jointly optimized (Section 3.5); representation norms are learned during training.

With architecture and other training settings fixed, we test two controls: replacing the U-Net's explicit skill condition with zeros, and detaching τ0\tau_0 before noising:

τn=αˉnsg⁡[τ0]+1−αˉnϵ.\tau_n=\sqrt{\bar{\alpha}_n}\operatorname{sg}[\tau_0]+\sqrt{1-\bar{\alpha}_n}\epsilon.

The zero-condition variant's fused input still contains observation and skill information. Detaching τ0\tau_0 blocks diffusion gradients through the clean-input path while retaining the trainable skill-conditioning path and action/VQ supervision.

Section 4.4 and Table 6 report three-seed mean ± SD. Norm is the mean L2L_2 norm over valid fused embeddings.

VariantDiffusionExplicit conditionStop-gradient on τ0\tau_0NormState SR (%)
B (w/o diffusion)×——8.9 ± 0.539.58 ± 2.53
Full DASL✓✓×8.8 ± 0.454.22 ± 4.65
w/o explicit conditioning✓××8.6 ± 0.547.07 ± 4.03
Stop-gradient on τ0\tau_0✓✓✓9.0 ± 0.451.83 ± 3.52

Full DASL exceeds the zero-condition variant by 7.15 percentage points in mean SR. The stop-gradient variant reaches 51.83% mean SR, which is 2.39 percentage points below Full DASL and 12.25 points above B, showing downstream gains with the skill-conditioning path retained.

Mean norms range from 8.6 to 9.0, with similar scales for Full DASL and B (8.8 vs. 8.9). These measurements characterize representation scale; Table 5 complements them with skill-use and language-association diagnostics. All variants execute without diffusion denoising. Diffusion weights are tuned separately for state and image inputs (0.5 and 0.01) and fixed across diffusion-enabled variants within each modality (Appendix D.2). We have revised the interpretation in Appendix F.3 to focus on the measured benefits of auxiliary latent regularization for task performance.

W2 & Q4. Can the Temporal Parameter Settings Be Clarified Further?

Thank you for your suggestions on temporal settings and experimental controls. We clarify the parameter roles and report cross-benchmark comparisons.

The temporal parameters have distinct roles. II is the skill-update interval, HH the auxiliary diffusion window, and LL the observation–action context length. DT trains on complete annotated trajectories with causal attention. LtrainL_{\mathrm{train}} includes batch padding; action losses use valid positions only. At inference, DT uses at most Linf=KL_{\mathrm{inf}}=K control steps and predicts an action every step. Diffusion uses the first min⁡(H,Ltrain)\min(H,L_{\mathrm{train}}) fused positions during training; DT uses the complete sequence.

KK and HH are configured independently and take equal values here: 5 for LOReL-state, 10 for LOReL-image, 20 for Kitchen, and 10 for CALVIN-D state and BabyAI GoToSeq.

LOReL settings appear in Appendix D.2, Table 14. A/C use I=HI=H; B, Full DASL, w/o IS, and gradient rescaling use I=2I=2, with H,KH,K fixed within each modality. A/B disable diffusion, so their effective difference is changing II from 5/10 steps for State/Image to 2, affecting planner observation sampling and skill sharing. With II and controller context fixed, HH has no independent computational role in A/B. With diffusion enabled, I,HI,H separately control skill updates and the diffusion window.

Cross-benchmark controls vary only II. We fix HH, DT context, and other settings within each benchmark. Results are three-seed mean ± SD (Appendix F.1, Tables 22–23).

Kitchen uses H=K=20H=K=20. N-rate is the mean number of completed subtasks per episode, up to four.

ObservationsIISeen N-rateUnseen N-rate
State201.66 ± 0.051.76 ± 0.14
State101.54 ± 0.081.62 ± 0.18
Image201.74 ± 0.091.71 ± 0.17
Image101.70 ± 0.121.66 ± 0.19

I=20I=20 gives higher mean N-rates across both modalities and splits, by 0.04–0.14. Independent configuration permits I=HI=H.

CALVIN and BabyAI use H=K=10H=K=10:

BenchmarkIISR (%)
CALVIN-D, six-task state548.05 ± 1.27
CALVIN-D, six-task state1052.13 ± 1.09
BabyAI GoToSeq, 1k trajectories263.93 ± 1.91
BabyAI GoToSeq, 1k trajectories1058.07 ± 2.58

Mean SR is 4.08 points higher with I=10I=10 than I=5I=5 on CALVIN, and 5.86 points higher with I=2I=2 than I=10I=10 on BabyAI. The preferred interval among these settings varies across tasks.

W4 & Q1. Differences among Baseline Results on LOReL

Thank you for raising these differences. We explain the reproduction settings and compare with the published results.

Sources and aggregation appear in Appendix C and Table 12. We use published results for corresponding settings and official implementations for additional modalities or instruction categories. Table 1 reports five language categories and their macro-average; Table 15 reports six tasks and their macro-average (Appendix B.1).

LISA reproduction and training behavior. We use official code, the original hyperparameters as the tuning starting point, public LOReL Sawyer data, and the original six tasks. LISA's seen-state SR is 12.78 ± 9.02%, versus the original 66.7 ± 5.2% [2].

We observed index collapse, including after matching temporal settings and tuning budgets. With I=2I=2 and a five-step DT context, LISA and DASL achieve five-category macro-average SRs of 17.63 ± 3.87% and 54.22 ± 4.65% (Appendix F.8, Table 29).

LCSD reports 6.7 ± 3.3% with official LISA code, rising to 43.4 ± 0.2% with code reinitialization [3]. LADS's LISA reproduction reaches 38.6 ± 3.9% on state atomic-seen instructions after updating VQ to partially mitigate collapse [4]. These studies use their own settings and show how codebook training affects reproduction. LangDT's six-task SR of 34 ± 7% (Table 15) is close to the original 33.3 ± 5.6% [2].

LADS re-evaluation and published results. LADS aggregates different rephrasings as Atomic Rephrased. We run its official implementation to obtain the separate seen, unseen noun, unseen verb, unseen noun & verb, and human-provided categories in Table 1.

For comparison with published LADS results [4], its image Atomic Seen and Composite SRs are 52.5 ± 2.1% and 24.0 ± 0.6%; DASL achieves 53.33 ± 5.36% and 32.50 ± 3.99%. With state inputs, published LADS achieves 68.3 ± 4.5% and 30.0 ± 11.2%, compared with DASL's 62.22 ± 3.85% and 23.61 ± 2.10%. These are cross-paper results from the respective evaluation settings. DASL has higher image-based means, while published LADS has higher state-based means on both tasks. This modality-dependent pattern also appears in Kitchen, where DASL has higher image N-rates and LADS higher state N-rates (Table 2).

W5 & Q5. Multi-Seed Ablations and Parameter Sweeps under a Common Protocol

Thank you for highlighting multi-seed evaluation. We repeated component ablations, codebook diagnostics, codebook-size sweeps, and regularization comparisons with the same three training seeds under a common protocol (mean ± SD).

Table 5 combines task success and codebook diagnostics. Emb. denotes skill embedding; Indep., independent I,HI,H; Diff., latent diffusion.

MethodEmb.I=HI=HIndep. I,HI,HDiff.Active / 20HnormH_{\mathrm{norm}}Top-1 RatioSoft MISkills / InstructionState SR (%)
LISA×✓××4.33 ± 0.580.468 ± 0.0360.338 ± 0.0350.018 ± 0.0063.72 ± 0.2111.08 ± 7.98
Variant A✓✓××20.00 ± 0.000.936 ± 0.0230.104 ± 0.0160.246 ± 0.0138.74 ± 0.2829.14 ± 3.26
Variant B✓×✓×19.00 ± 1.000.874 ± 0.0310.125 ± 0.0220.258 ± 0.0115.34 ± 0.1939.58 ± 2.53
Variant C✓✓×✓20.00 ± 0.000.932 ± 0.0270.105 ± 0.0180.252 ± 0.0128.53 ± 0.3133.46 ± 3.68
Full DASL✓×✓✓19.00 ± 0.000.936 ± 0.0200.096 ± 0.0250.273 ± 0.0106.22 ± 0.1154.22 ± 4.65
DASL w/o IS✓×✓✓15.00 ± 0.000.742 ± 0.0500.312 ± 0.0600.247 ± 0.0126.17 ± 0.1345.92 ± 4.00
w/o IS + gradient rescaling✓×✓✓17.00 ± 0.000.828 ± 0.0350.219 ± 0.0400.262 ± 0.0106.10 ± 0.1251.67 ± 3.20

Full DASL exceeds B and w/o IS by 14.64 and 8.30 percentage points in mean SR.

Codebook-size sweep (Table 27).

Codebook sizeState SR (%)Image SR (%)
548.36 ± 3.8237.20 ± 0.94
1046.12 ± 3.2932.70 ± 1.52
2054.22 ± 4.6550.13 ± 0.43
4037.68 ± 4.4140.30 ± 0.78

The original regularization Table 21 is now Table 25. Under the same three-seed protocol, SRs are 39.58 ± 2.53% for the unregularized Variant B, 44.27 ± 3.14% for VAE, 47.16 ± 3.37% for MLP autoencoder, and 54.22 ± 4.65% for latent diffusion, with parameter counts matched for the regularizers. The M=20M=20 and diffusion entries reuse Full DASL results.

Table 20 is a single-seed frequency-ratio sweep; Tables 21–23 report three-seed temporal controls.

W6 & Q6. Does the Option Selector Run at Every Control Step or Cache Its Output between Skill Updates?

Thank you for your suggestions on selector invocation and inference efficiency.

The original selector runs at every control step. II controls skill sampling and sharing. Full-policy latency is 73.48 ms (13.61 Hz; Table 7, Appendix H); the selector takes 31.817 ms, or 43.3% (Table 31).

With Eager-FP32 on LOReL-state and I=2I=2, caching runs the selector at update steps and reuses skills between updates. DT receives current observations and predicts actions every step.

ImplementationTimed stepsLatency (ms)Equivalent frequency (Hz)
Per-step recomputationAll steps73.4813.61
Skill caching, I=2I=2Cached steps only41.6624.00
Skill caching, I=2I=2All steps57.5717.37

The 41.66 ms includes observation processing, skill–observation fusion, DT prediction, and action postprocessing. Across all steps, latency falls by 21.7% to 57.57 ms. Frequency is 1000/mean latency (ms)1000/\text{mean latency (ms)}; 24.00 Hz applies only to cached steps (Table 32).

Table 7 compares methods under matched Eager-FP32 execution. LISA reaches 15.68/15.01 Hz with State/Image inputs; DASL reaches 13.61/13.97 Hz. Skill caching reduces selector calls; history preallocation and CUDA Graphs reduce allocation and execution overhead. Diffusion is used only during training. Section 4.5 and Appendix H describe these execution details.

Conclusion

We have clarified the training design, evaluation settings, and inference implementation, and added multi-seed controls. Revisions are marked in blue.

Thank you again for your time and valuable feedback. We hope our response has addressed your concerns.

References

[1] Chen et al. Diffusion Model-Augmented Behavioral Cloning. ICML, 2024.

[2] Garg et al. LISA: Learning Interpretable Skill Abstractions from Language. NeurIPS, 2022.

[3] Ju et al. Rethinking Mutual Information for Language Conditioned Skill Discovery on Imitation Learning. ICAPS, 2024.

[4] Jiang et al. Discrete Latent Plans via Semantic Skill Abstractions. ICLR, 2025.

Reviewer_3KptOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Dual-Process Atomic Skill Learning (DASL) to address concentrated codebook usage and unstable skill representations in language-conditioned hierarchical imitation learning. A slow Option Transformer selects discrete skills from a vector-quantized codebook, while a fast Decision Transformer predicts actions using continually updated observations. The framework independently configures the semantic update interval and low-level trajectory horizon. It combines interval-level action supervision of shared skills, skill–observation fusion, and an auxiliary skill-conditioned diffusion objective that regularizes fused latent trajectories during training and is removed at inference.

Extensive experiments across simulated benchmarks and real-world manipulation demonstrate gains in skill acquisition and compositional task performance.

Critical strengths

Comprehensive ablations and skill-learning diagnostics. Beyond reporting overall success rates, the paper evaluates skill–observation fusion, independent temporal parameters, and latent diffusion through component comparisons. It also examines the role of interval-level supervision by restricting the action feedback reaching the skill branch while preserving forward skill sharing. Codebook statistics—including active-code counts, usage entropy, and language–skill association—complement task success and help assess how these changes affect the learned representations. Additional comparisons with parameter-matched VAE and MLP autoencoder regularizers, together with sensitivity analyses of skill-update intervals and codebook sizes, provide useful evidence for evaluating the combined design.

Meaningful gains across simulation and physical deployment. DASL achieves improvements across several evaluated settings, including language variations and compositional instructions. On LOReL image-based compositional tasks, it reaches 32.50% success compared with SkillDiffuser’s 25.21%; under the six-task CALVIN-D state protocol, it achieves 48.05% compared with LCSD’s 35.60%. Real-robot evaluation further demonstrates an average success rate of 88% across five tasks, versus 56% for LISA. Inference measurements complement these performance results by showing the practical benefit of using diffusion only during training and retaining direct action prediction during execution.

Critical weaknesses

The broader lessons from the empirical findings could be articulated more clearly. The paper distinguishes its approach from prior methods such as LCSD and LADS and provides extensive evidence for the effectiveness of its design. However, the discussion could better explain what the ablations and codebook diagnostics reveal about the relationship between skill utilization, language–skill alignment, and control performance. This would make the contribution clearer beyond identifying an effective combination of components.

The relationship between the temporal parameters needs clarification. The paper emphasizes independently setting the skill-update interval II and trajectory horizon HH, while the low-level action policy uses a context length LL. The algorithm explicitly uses HH to determine the diffusion window, but does not clearly explain how LL is configured or whether it changes with HH. This makes it difficult to determine precisely what the proposed temporal decoupling changes, particularly in Variant B, which does not use diffusion.

Questions for authors

What do the ablations and codebook diagnostics reveal about the relationship between codebook utilization, language–skill alignment, and task success? In light of LCSD and LADS, could the authors clarify the additional insight these findings provide into learning useful discrete skills?

How is the low-level context length (L) configured, and what is its relationship to the trajectory horizon (H)? Please provide the values of (L), (I), and (H) for the component variants and clarify whether (H) affects action prediction or only the auxiliary diffusion window. In particular, what does (H) control in Variant B, where diffusion is disabled?

5 comments

We thank Reviewer 3Kpt for the time and effort devoted to reviewing our paper and for the valuable comments. Below, we address W1 & Q1 by discussing the implications of our ablation results and codebook diagnostics for discrete skill learning, and W2 & Q2 by clarifying the roles, configurations, and relationships of the temporal parameters.

W1 & Q1. Codebook Utilization, Language–Skill Alignment, and Task Success

We thank the reviewer for raising this question. Understanding the relationships among codebook utilization, language–skill association, and task success is important for learning effective discrete skills. The component ablation results below are reported as mean ± SD over three training seeds.

MethodActive / 20HnormH_{\mathrm{norm}}Soft MIState SR (%)
Variant A20.00 ± 0.000.936 ± 0.0230.246 ± 0.01329.14 ± 3.26
Variant B19.00 ± 1.000.874 ± 0.0310.258 ± 0.01139.58 ± 2.53
DASL w/o IS15.00 ± 0.000.742 ± 0.0500.247 ± 0.01245.92 ± 4.00
w/o IS + gradient rescaling17.00 ± 0.000.828 ± 0.0350.262 ± 0.01051.67 ± 3.20
Full DASL19.00 ± 0.000.936 ± 0.0200.273 ± 0.01054.22 ± 4.65

First, higher codebook utilization does not necessarily correspond to better task performance. Variant A activates all 20 entries, with mean usage entropy of 0.936. Variant B activates 19 entries on average, with entropy of 0.874, while mean SR increases from 29.14% to 39.58%. A and B use I=5I=5 and I=2I=2, respectively, with the controller context held fixed. Thus, changing the skill-update interval increases task success while active-code count and usage entropy decrease. Code count and usage uniformity alone are insufficient to measure the control effectiveness of learned skills.

Second, language–skill association complements codebook utilization. Usage entropy describes the overall skill-assignment distribution, whereas Soft MI measures the statistical association between language instructions and skill assignments. Full DASL has mean Soft MI of 0.273, compared with 0.246 for A and 0.258 for B. Across these variants, language–skill association and task performance do not vary monotonically: B has higher mean Soft MI than w/o IS (0.258 vs. 0.247), but lower mean SR (39.58% vs. 45.92%). Language-conditioned skill differentiation and control effectiveness therefore need to be considered together. We interpret these metrics alongside task success and the appendix visualizations to examine how skill assignments relate to task execution.

Finally, action feedback received by shared skills affects their learning and utilization. The w/o IS ablation preserves forward skill sharing, the update interval, model capacity, and all valid action-prediction losses, while restricting action-gradient paths into the high-level skill branch. Full DASL and w/o IS activate 19 and 15 codes on average, with usage entropies of 0.936 and 0.742, respectively, showing broader and more uniform codebook use with full interval supervision.

Under the same w/o IS mask, rescaling retained gradients by the interval’s valid length raises mean SR from 45.92% to 51.67%, a gain of 5.75 percentage points. Mean active-code count rises from 15 to 17, and entropy from 0.742 to 0.828. These results show that action-feedback paths and gradient scaling affect skill utilization and task performance. Interval supervision operates through standard backpropagation over shared skills, allowing skill representations to receive feedback from the actions they guide.

These findings clarify DASL’s focus relative to LCSD and LADS. LCSD promotes language–skill association through language reconstruction and improves utilization through code reinitialization. LADS separates learning the discrete latent-plan space and low-level policy from learning the language-conditioned high-level policy. Our additional finding is that action-feedback paths and their scaling affect both codebook use and task success within a jointly trained hierarchy, while usage entropy and language–skill association have different mean rankings from task success across the tested variants. These findings motivate evaluating discrete skills through utilization, language association, and control effectiveness together.

Table 5 reports these comparisons, complemented by the visualizations in Appendix E.5 (Figures 14–17) of language–skill associations, skill selection, projected representations, and atomic/compositional skill distributions.

W2 & Q2. Low-Level Context Length L and Trajectory Horizon H

We thank the reviewer for raising these questions. Clarifying the roles and relationships of the controller context length LL, skill-update interval II, and trajectory horizon HH is important for understanding temporal decoupling, particularly when diffusion is disabled. We denote the training input sequence length by LtrainL_{\mathrm{train}} and the maximum DT inference context by Linf=KL_{\mathrm{inf}}=K.

First, training uses complete annotated trajectories, while inference uses a sliding window with a fixed maximum length. In the DASL implementation, LtrainL_{\mathrm{train}} is the temporal dimension of the batched input sequence, including any padding. Trajectories of different lengths are padded within a batch, and action losses are computed only at valid positions. DT uses causal attention to predict actions from the current observation and preceding history. At inference, DT uses the most recent history of up to KK control steps and predicts the current action at every step.

Table 14 in Appendix D.2 lists the following LOReL component configurations. Each tuple denotes (I,H,Linf)(I,H,L_{\mathrm{inf}}). The third entry specifies the inference context length; all variants are trained on complete annotated trajectories.

VariantLOReL-stateLOReL-imageDiffusion
A(5, 5, 5)(10, 10, 10)×
B(2, 5, 5)(2, 10, 10)×
C(5, 5, 5)(10, 10, 10)✓
Full DASL(2, 5, 5)(2, 10, 10)✓
DASL w/o IS(2, 5, 5)(2, 10, 10)✓
w/o IS + gradient rescaling(2, 5, 5)(2, 10, 10)✓

KK and HH are configured separately and take equal values in these experiments. KK controls the DT inference history window, while HH controls the trajectory window used for diffusion training. Within each observation modality, all component variants retain the same HH and DT inference context.

Second, HH determines the auxiliary diffusion input length, while the controller predicts and executes the current action at each control step. During training, DT uses the complete observation–skill fusion sequence for action prediction. The diffusion branch separately takes the first HdiffH_{\mathrm{diff}} positions, where

Hdiff=min⁡(H,Ltrain).H_{\mathrm{diff}}=\min(H,L_{\mathrm{train}}).

DT is trained on the complete fused sequence, while diffusion uses the extracted trajectory window. With II and controller context fixed, changing HH adjusts the latent trajectory length modeled by auxiliary diffusion. The diffusion loss affects shared representations and the resulting policy through training gradients. At inference, the controller predicts actions step by step without diffusion denoising.

Finally, Variant B disables diffusion, so HH has no independent computational role once II and controller context are fixed. It remains in the configuration but does not participate in diffusion-loss or action-prediction computation. A and B both disable diffusion; their comparison changes the skill-update interval from five steps for State inputs and ten steps for Image inputs to two steps, while keeping controller context fixed. Changing II changes both the planner’s observation-sampling stride and the skill-sharing interval, while the controller context remains fixed.

A and C use the shared setting I=HI=H. B, Full DASL, and DASL w/o IS configure II independently while retaining the same HH and controller context. Temporal decoupling allows the skill-update interval II to be selected independently of the trajectory horizon HH. With diffusion enabled, skill-update frequency and diffusion trajectory window can be adjusted separately, while KK configures the controller’s inference context. Full DASL and DASL w/o IS retain the same forward temporal configuration and differ in the action-gradient paths into the skill branch.

Conclusion

We have clarified the relationships between codebook utilization, language–skill association, and task performance, and included the temporal configuration table. The manuscript now distinguishes full-trajectory training, DT inference context, and the diffusion trajectory window, and specifies the parameters changed by each ablation. The corresponding revisions are marked in blue in the manuscript.

Thank you again for your time and valuable feedback. We hope our response has addressed your concerns.

If the authors have addressed your concerns or provided clarifications, please join the discussion to indicate whether your concerns have been resolved or if you have any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for the time and care devoted to evaluating our work and for their detailed, constructive feedback. Your comments have helped us clarify the contribution, strengthen the experimental evidence, and describe the scope of our conclusions more precisely.

We have updated the manuscript PDF, with additions and revisions marked in blue. Several reviews raised related questions about the training mechanisms, temporal settings, and experimental comparisons. We summarize the main revisions below.

  1. Training mechanisms and evidence for skill learning. We clarified how interval-level supervision operates through standard backpropagation and revised the discussion of DASL’s relationship to prior work, including DBC. We added gradient-rescaling, conditioning, and stop-gradient controls, and expanded the analysis of how codebook utilization, language–skill association, and task success provide complementary evidence about learned skills (Section 2.2, Tables 5–6, and Appendix F.7).

  2. Multi-seed evaluation and baseline comparisons. We repeated the component ablations, codebook diagnostics, codebook-size sweeps, and regularization comparisons using the same three training seeds under a common protocol, reporting mean ± SD. We also added a LISA comparison with matched skill-update intervals, controller inference contexts, and tuning budgets, and clarified baseline sources, reproduction settings, and result aggregation (Appendix C and Tables 25, 27, and 29).

  3. Temporal parameters and controlled experiments. We clarified the roles of the skill-update interval, diffusion horizon, and controller context, distinguishing full-trajectory training from inference context and the auxiliary diffusion window. We provided the component configurations and added controls that separately vary planner observation sampling and skill duration, together with interval comparisons on Kitchen, CALVIN, and BabyAI (Table 14 and Tables 21–23).

  4. Scope of compositional evaluation. We revised the analysis to distinguish compositional instruction execution on LOReL, partial completion of unseen combinations in Kitchen, and multi-stage robot execution with compositional demonstrations. We also clarified the overlap between held-out Kitchen combinations and training subsequences, and distinguished partial completion from full-sequence success (Section 4.2, Appendix B.2, and Appendix G).

  5. Inference implementation and efficiency. We clarified when the option selector is invoked and added latency measurements for skill caching between updates. The revised presentation distinguishes matched execution comparisons from implementation optimizations and reports their respective effects on inference cost (Section 4.5 and Appendix H).

Detailed explanations and results are provided in our responses to the individual reviews. We appreciate the reviewers’ thoughtful suggestions and welcome further questions or clarification requests.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.