acceptodds
Under review as a conference paper at ICLR 2027

Atomic-to-Composite Distillation via Task-Wise Progress Potentials and Hidden-State Alignment in Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models achieve strong performance on individual manipulation skills, yet their success drops substantially when these skills must be executed sequentially as composite tasks. We identify two key challenges underlying this atomic-to-composite gap. First, composite behaviors are difficult to learn: collecting long sequence demonstrations is costly, while reinforcement learning (RL) with sparse binary rewards provides little learning signal for trajectories that achieve only partial progress. Second, composite instructions do not explicitly indicate which atomic task should be active at each stage, creating ambiguity at subtask transitions. Moreover, the same skill succeeds far less often under a composite instruction than under its own atomic instruction, suggesting that composite conditioning fails to reliably invoke competence the policy already possesses. We propose ATLAS, an end-to-end RL framework that leverages the policy's own atomic-task competence to address both challenges. To address sparse credit assignment, task-wise progress potential shaping learns reusable progress potentials from successful and failed atomic rollouts and composes them across stages to provide dense, phase-aware rewards while preserving the optimal policy. To bridge the representation gap, phase-conditioned atomic-to-composite hidden-state alignment aligns composite-conditioned representations with their corresponding atomic-conditioned representations at each stage, allowing the composite-conditioned policy to directly exploit skills already encoded by the atomic policy without inference-time planners or prompt switching. These complementary mechanisms bridge atomic and composite learning within a unified training framework. Across simulation and real-robot experiments, ATLAS consistently outperforms zero-shot VLA baselines, standard PPO, and existing progress-reward approaches, achieving gains of up to 33pp on RoboLab and 40pp in real-world settings.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action model, Compositional Control

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_ceJWOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ATLAS, a RL framework for VLAs. The goal is to improve VLA performance on composite tasks, where VLAs tend to struggle. IE a VLA may be able to complete tasks A and B, but fail to complete A then B. They make two major contributions as far as I can tell. The first is a dense potential-based reward signal that is based on the probablity of completing a subtask. Then these are chained togethor so that they act basically as an indicator of if the agent will complete the task or not (w/ discounting). Then, by adding reward that subtracts the current score from the score at the next step, this encourages the agent to make progress on this potential, ie increase its probability of completing the task. They additionally create a regularizar that encourages the composite latent representations to match the subtask latent representations. They run experiments in both sim and real life and show good performance.

Critical strengths

This paper addresses a critical problem, long-horizon behaviors for VLAs. Their solution is intuitive and seems effective. Presentation is clear and easy to follow. Experiments look good and illustrate both the problem and their solution.

Critical weaknesses

Nothing major.

Questions for authors
  1. Does ATLAS stand for something?
  2. What is 33pp and 40pp in the introduction? Is this percent?
  3. For the alignment part, I get that you want to align the latents. However, it seems like you want to align the latent for the composite task A->B with the latent of task A only when task A is active in the composite task. Otherwise, you would just learn the mean of the latents A,B for A->B. Is this what you do? I do not see where you account for the active task in section 3.4. The text around line 90 seems to imply something like this is happening though.
  4. Why is there a significant difference in A,B performance in Table 4? Shouldnt this ablation only affect the composite tasks?
2 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer ceJW

We thank the reviewer for the positive evaluation and the accurate summary of our method, and for recognizing that ATLAS addresses a critical problem for VLAs, that the approach is intuitive and effective, and that the experiments clearly illustrate both the problem and our solution in simulation and on a real robot. We address each question below.

[Q1] Does ATLAS stand for something?

ATLAS stands for Atomic-to-composite Transfer via Latent Alignment and potential Shaping, referring to our task-wise progress-potential shaping and phase-conditioned atomic-to-composite hidden-state alignment. We have added this explanation at the first mention of ATLAS in both the abstract and the introduction.

[Q2] What is 33pp and 40pp in the introduction? Is this percent?

"pp" denotes percentage points, meaning absolute differences in success rates rather than relative percent increases. We have replaced "pp" with "percentage points" in the revision.

[Q3] For the alignment part, … it seems like you want to align the latent for the composite task A→B with the latent of task A only when task A is active in the composite task. Otherwise, you would just learn the mean of the latents A,B for A→B. Is this what you do? I do not see where you account for the active task in section 3.4.

Yes. During training, the teacher for each composite observation is the atomic instruction of the currently active subtask, selected using ntn_t, the number of completed phases. For the task A→B, the target is task A until task A is completed and task B afterward, so each sample has a phase-specific target rather than an averaged one. We clarified this in Sec.3.4.

[Q4] Why is there a significant difference in A,B performance in Table 4? Shouldn't this ablation only affect the composite tasks?

ATLAS jointly trains atomic and composite tasks, and task-wise potential shaping is applied to atomic rollouts as well. Therefore, removing potential shaping affects atomic-task performance, while changes to alignment can indirectly affect atomic performance through shared policy updates.

Thanks for the clarifications.

Reviewer_atwLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the “Atomic-to-Composite gap” in VLA models, where models suffer a significant performance drop when tasks are executed sequentially as composite tasks. The authors propose "ATLAS," an end-to-end reinforcement learning method to bridge this gap. They propose hidden-state alignment between the representations during composite task execution and the corresponding atomic task representations. The proposed method outperforms zero-shot VLA baselines and standard PPO using simulation and real-robot experiments.

Critical strengths
  • The paper's structure is clear, and the figures explain the problem setting and the overall framework.
  • By combining progress-based reward shaping and hidden-state alignment, it simultaneously addresses the sparse reward problem and the ambiguity at subtask transitions.
Critical weaknesses
  • The evaluation is primarily limited to composite tasks consisting of two subtasks. It remains unclear how the alignment scales for longer sequences.
  • The paper lacks details of the experimental settings, and the justification of the comparison to zero-shot VLA models is lacking.
  • Detailed analysis regarding the sample efficiency and robustness of the task-wise progress model is lacking.
Questions for authors
  • If extended to three or more tasks, is there a risk that the errors in the Task-wise progress might accumulate or that representation alignment might be unstable during training?
  • The real-robot experiments evaluate only 10 episodes per task. This may be statistically insufficient to prove the robustness and real-world effectiveness of the method.
  • How does the method accurately detect phase completion and switch the potential without the privileged information of the simulator during training?
4 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer atwL [1/3]

We are glad the reviewer found the paper’s structure clear and the figures helpful in explaining the problem setting and overall framework. We also appreciate the recognition that combining progress-based reward shaping and hidden-state alignment addresses both the sparse-reward problem and ambiguity at subtask transitions. We have carefully addressed the reviewer's comments below.

[W1-1] The evaluation is primarily limited to composite tasks consisting of two subtasks. It remains unclear how the alignment scales for longer sequences. [Q1] If extended to three or more tasks, is there a risk that the errors in the Task-wise progress might accumulate or that representation alignment might be unstable during training?

At the formulation level, ATLAS extends directly to tasks with three or more subtasks: progress potentials are composed across phases, while hidden-state alignment is applied only to the atomic instruction corresponding to the currently active phase. Thus, the method requires no sequence-length-specific changes.

To verify this at the experimental level, we extend Suite 1 with a third subtask, jointly training on three atomic tasks and two three-step compositions and evaluating on trained and four unseen compositions, as well as atomic tasks. As shown in Tables R1a–b, ATLAS improves trained and unseen composition success under both criteria (Order: 6.8% to 14.6% and 6.8% to 12.5%; Strict: 5.7% to 13.0% and 6.5% to 10.4%) and atomic-task success (Order: 76.7% to 87.2%; Strict: 74.7% to 84.4%). PPO underperforms the π₀.₅ baseline on all aggregated metrics (Table R1a) and on most individual compositions, consistent with the challenges of sparse rewards over longer horizons. We clarified this in Appendix B.4.

Table R1a. Overall success rates (%) on atomic and three-stage composite tasks. Mean±SD across four repeated evaluations of the same 24 scenes/task (96 episodes/task).

CriterionMethod3-subtask Trained3-subtask UnseenAll 3-subtaskAtomic Overall
Orderπ₀.₅6.8±2.66.8±1.86.8±1.876.7±5.0
Order+PPO1.6±1.03.4±1.02.8±0.847.2±4.1
Order+ATLAS(Ours)14.6±6.112.5±2.313.2±2.087.2±3.1
Strictπ₀.₅5.7±2.66.5±1.86.2±1.974.7±3.3
Strict+PPO1.0±1.23.4±1.02.6±1.046.2±4.4
Strict+ATLAS(Ours)13.0±4.310.4±0.911.3±1.384.4±4.0

Table R1b. Success rates (%) by three-stage composition and atomic task. Mean±SD across four repeated evaluations of the same 24 scenes/task (96 episodes/task). Trained: A→B→C,A→C→B. Unseen: remaining four compositions.

CriterionMethodA→B→CA→C→BB→A→CB→C→AC→A→BC→B→AAtomic AAtomic BAtomic C
Orderπ₀.₅8.3±5.95.2±2.16.2±5.414.6±2.46.2±4.20.0±0.062.5±10.283.3±5.984.4±12.0
Order+PPO2.1±2.41.0±2.111.5±2.12.1±2.40.0±0.00.0±0.043.8±2.444.8±7.153.1±6.2
Order+ATLAS(Ours)19.8±15.09.4±7.933.3±7.64.2±0.07.3±4.05.2±2.187.5±3.486.5±4.087.5±6.8
Strictπ₀.₅6.2±5.45.2±2.16.2±5.414.6±2.45.2±5.20.0±0.061.5±10.479.2±5.983.3±11.3
Strict+PPO2.1±2.40.0±0.011.5±2.12.1±2.40.0±0.00.0±0.042.7±4.044.8±7.151.0±8.6
Strict+ATLAS(Ours)17.7±11.08.3±5.928.1±4.04.2±0.05.2±2.14.2±0.084.4±4.085.4±2.483.3±10.8
AuthorsOpenReview ↗

Rebuttal for Reviewer atwL [2/3]

[W2] The paper lacks details of the experimental settings, and the justification of the comparison to zero-shot VLA models is lacking.

(1) Details of the experimental settings: We have provided the full experimental settings in Appendix A, including task suites, simulation and real-robot setups, training hyperparameters, evaluation protocols, and baseline configurations. We would appreciate it if the reviewer could specify which experimental details are missing or insufficient, and we will promptly address them in the revised manuscript.

(2) Justification of the comparison to zero-shot VLA models: We include zero-shot generalist VLAs not as directly comparable baselines but as references showing that the atomic-to-composite gap is not specific to π0.5π_{0.5}. All three use DROID checkpoints sharing the Franka joint-position interface, and even the strongest (Cosmos 3 Edge Policy) falls below the independent-execution expectation A×B in every suite (e.g., Suite 3: 64.9 expected vs. 37.5/29.2 observed in Table 2).

[W3] Detailed analysis regarding the sample efficiency and robustness of the task-wise progress model is lacking.

We conduct additional experiments to evaluate the sample efficiency and robustness of our task-wise progress model. We collected 384 episodes over 16 rollout rounds, randomly sampled training episodes from this pool, and repeated task-wise progress model training 20 times. We report AUROC, which measures how well the model discriminates eventual skill success using time-matched observations within each phase.

As shown in Table R2, our potential outperforms Temporal Progress and Phase Count at every evaluated fitting-data budget, achieving AUROCs of 0.614 with 24 episodes per task, 0.643 with 96, and 0.666 with 192. This allows the fitting budget to be chosen based on the desired balance between performance and data-collection cost. Moreover, across 20 random fitting subsets, AUROC variability is modest and generally decreases as the fitting budget grows (SD = 0.041 with 3 episodes per task vs. 0.006 with 192), suggesting that the learned potential becomes less sensitive to the particular episodes sampled. We added these details in Appendix B.2.

Table R2. Time-matched AUROC across different potential-fitting data budgets. AUROC evaluates discrimination of eventual skill success using time-matched, within-phase observations. Results are reported as mean ± SD over 20 random fitting-data subsets. Phase Count achieves 0.500 by construction because it is constant within each phase.

Episodes / TaskLearned (ATLAS)Temporal ProgressPhase Count
30.585 ± 0.0410.514 ± 0.0250.500
120.593 ± 0.0180.492 ± 0.0190.500
240.614 ± 0.0190.485 ± 0.0230.500
96 (4 rounds)0.643 ± 0.0120.490 ± 0.0170.500
1920.666 ± 0.0060.492 ± 0.0080.500
AuthorsOpenReview ↗

Rebuttal for Reviewer atwL [3/3]

[Q2] The real-robot experiments evaluate only 10 episodes per task. This may be statistically insufficient to prove the robustness and real-world effectiveness of the method.

To address this concern, we expand the real-robot evaluation from 10 to 30 paired layouts per task, evaluating ATLAS and the base π₀.₅ under identical initial conditions. As shown in Table R3, ATLAS achieves higher success rates across both tasks and success criteria, with the largest improvement in Suite 1 Strict (83.3% vs. 43.3%, +40.0 percentage points, p = 0.004). Pooled exact McNemar tests across 60 paired trials further show improvements in Overall Order (pp = 0.039) and Overall Strict (p < 0.001), providing stronger evidence of real-world effectiveness. We include videos of all evaluation episodes in the supplementary material. They show that the base π₀.₅ policy sometimes fails to follow the instructed order, picking up and placing objects without regard to the requested sequence. We clarified this in Appendix B.5.

Table R3. Real-world success rates (%) and paired exact McNemar tests across two manipulation suites. Overall: 60 paired trials (30/suite). Parentheses show McNemar p-values comparing ATLAS with π₀.₅.

MethodOverall OrderOverall StrictSuite 1 OrderSuite 1 StrictSuite 2 OrderSuite 2 Strict
π₀.₅85.045.086.743.383.346.7
ATLAS (Ours)96.7 (p=0.039)76.7 (p<0.001)100.0 (p=0.125)83.3 (p=0.004)93.3 (p=0.375)70.0 (p=0.065)

[Q3] How does the method accurately detect phase completion and switch the potential without the privileged information of the simulator during training?

We do use the privileged information of the simulator during training to determine phase completion, as described in Sec. 3.3 (Eq. 2, Lines 239–240). This information is used only for training-time shaping and phase-conditioned alignment. For real-world training, phase completion can be determined using an external subtask-completion detector. At inference time, no privileged simulator information or explicit phase switching is required, as ATLAS operates as a single policy conditioned on observations and the composite instruction.

Dear Reviewer atwL,

Thank you again for your detailed review. As the discussion period is closing, we would be grateful for any reaction to our response.

To address your concerns about longer task sequences, we extended our evaluation to three-subtask compositions, including unseen task orders, where ATLAS improves overall success over the baselines. We also added a sample-efficiency and robustness analysis of the task-wise progress model across different data budgets, expanded the real-robot evaluation to 30 paired trials per task with statistical significance tests, and clarified the rationale for including zero-shot VLA baselines and the use of simulator information during training. Details are provided in our rebuttal above.

We understand your concerns were substantial, and we would greatly appreciate knowing whether these additional experiments and clarifications address them, or which concerns remain.

Best regards,
The Authors

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

This paper studies the gap between atomic-skill competence and sequential composite-task execution in vision-language-action models. It proposes ATLAS, which jointly fine-tunes atomic and composite tasks using reinforcement learning with two additional mechanisms. First, task-specific potential functions are learned from successful and failed atomic rollouts to predict discounted skill completion. These potentials are composed using simulator-provided phase information to shape rewards. Second, composite-conditioned hidden representations are aligned with representations induced by the atomic instruction corresponding to the active phase. Both mechanisms operate during training, preserving the original single-policy inference procedure. Experiments cover four RoboLab suites, each containing two atomic tasks and their two composite orderings, alongside two physical-robot composite tasks. The reported results show improvements over the initial π₀.₅ policy, standard PPO, and the implemented GVL-style reward baseline.

Critical strengths
  1. Potential-based shaping and phase-conditioned representation alignment address complementary problems: sparse feedback and difficulty invoking the appropriate skill under composite instructions. The resulting policy requires neither an additional planner nor inference-time prompt switching, which is a meaningful deployment advantage. See Sections 3.2–3.4.

  2. The reported improvements are substantial and include preliminary physical-robot validation. Averaging the eight composite tasks in Table 2, order-preserving success increases from approximately 17.2% to 35.4%, while strict-selection success increases from 11.5% to 27.6% relative to π₀.₅. Reporting strict-selection success is valuable because it distinguishes faithful instruction following from eventual completion involving irrelevant object placements. The physical-robot results provide additional, although limited, evidence of transfer beyond simulation.

Critical weaknesses
  1. ATLAS trains on both atomic and composite tasks, whereas the principal PPO baselines train only on composite tasks. Although Table 4 removes individual components, it does not report an atomic-plus-composite PPO baseline that removes both shaping and alignment while retaining the same sampling and optimization settings. Consequently, the incremental contribution of the two central mechanisms beyond joint training remains unclear. A simple phase-count potential, such as Φq(s)=nt/Nq\Phi_q(s)=n_t/N_q, would also help establish how much benefit comes from the learned within-phase potential rather than access to subtask-completion information. Also, appendix A.7 additionally reports 4–16 rollout rounds for fitting the potentials; these interactions and their costs should be included explicitly in budget comparisons. Reliability is also difficult to assess: only one training seed is reported, simulation uses 24 episodes per task, and physical evaluation uses 10 episodes per task, without uncertainty estimates for the main success-rate tables. Finally, all evaluated compositions contain two pick-and-place subtasks and are included during training, limiting conclusions about reuse across unseen compositions or longer sequences.

  2. quation (2) assigns potential 1 after task completion, while the evaluation protocol terminates episodes upon success or timeout. Summing the shaping rewards in Equation (3) gives \sum_{t=0}^{T-1}\gamma^t(r't-r_t) = \lambda{\mathrm{prog}} \left[-\Phi_q(s_0)+\gamma^T\Phi_q(s_T)\right]. Without appropriate boundary handling, the final term can depend on the policy, termination time, and completed phases. The potential-based form alone therefore does not establish the claimed policy invariance in this episodic setting. The paper should specify terminal potentials, termination masks, and bootstrapping for genuine termination versus time-limit truncation. This issue is addressed in existing analyses of reward shaping in episodic reinforcement learning. Equation (4) also leaves unclear how language-token representations from atomic and composite instructions are matched despite their different lengths and token positions. Whether the method uses subclause extraction, pooling, positional matching, or another correspondence is essential to defining and reproducing the objective.

Questions for authors
  1. How does ATLAS compare with atomic-plus-composite joint PPO under the same task sampling, phase weighting, and total interaction budget? How does it compare with shaping based only on the completed-phase count? These comparisons would materially affect my assessment of the proposed mechanisms.

  2. How are success termination and time-limit truncation handled when computing shaping rewards and critic targets? Specifically, what terminal potential, termination mask, and bootstrap convention are used, and how do they remove the policy-dependent boundary term above?

  3. Which language tokens are compared in Equation (4), and how are different instruction lengths handled? Does the teacher use the current policy or a frozen copy? Are the potential functions evaluated on features from a frozen model or from the backbone being updated through LoRA?

  4. Can the authors provide results across independent training seeds and episode-level evaluation outcomes? For Figure A4, are fitting and evaluation sets separated by complete trajectories, are potential-fitting trajectories excluded from evaluation, and what do the error bars represent?

8 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer NYit [1/7]

We thank the reviewer for noting that potential-based shaping addresses sparse feedback and phase-conditioned alignment supports skill selection under composite instructions, as well as for recognizing the planner-free inference, substantial gains, strict-selection evaluation, and preliminary physical-robot validation. We address the reviewer's comments below.

[W1-1] ATLAS trains on both atomic and composite tasks, whereas the principal PPO baselines train only on composite tasks. ... the incremental contribution of the two central mechanisms beyond joint training remains unclear. [Q1-1] How does ATLAS compare with atomic-plus-composite joint PPO under the same task sampling, phase weighting, and total interaction budget?

We added an atomic–composite joint PPO baseline that removes both shaping and alignment while keeping the same task sampling, phase weighting, RL interaction budget, and optimization settings as ATLAS (3rd row of Tables R1–R4). For more reliable estimates, we repeated the evaluation on the same 24 scenes four times per task.

Tables R1–R4 show that joint training alone (3rd row) gives only modest, inconsistent gains over composite-only PPO (composite: 18.9% → 21.6% Order, 15.4% → 17.4% Strict), with no Order gain in Suite 1 and atomic drops on some tasks (e.g., Suite 1 atomic B: 69.8% → 45.8%). Adding shaping and alignment on the same setup makes the gains consistent: ATLAS outperforms joint PPO in composite success across all four suites and achieves the highest atomic success averaged over the eight atomic tasks (68.1% vs. 60.7%). Two-sided paired Wilcoxon signed-rank tests over 32 paired evaluation groups (4 suites × 2 compositions × 4 repetitions) show significant gains under both Order (34.0% vs. 21.6%, p = 8.3 × 10⁻⁵) and Strict (27.5% vs. 17.4%, p = 8.7 × 10⁻⁴) (Table R5). We added this baseline to Tables A4–A7 in the revision.

Table R1. Success rates (%) on RoboLab Suite 1. Mean ± SD across four repeated evaluations of the same 24 scenes per task (96 episodes/task). All methods use identical scenes.

MethodOverall OrderOverall StrictA→B OrderA→B StrictB→A OrderB→A StrictAtomic AAtomic B
π₀.₅30.7±2.029.7±1.029.2±3.428.1±4.032.3±6.231.2±5.459.4±6.367.7±12.0
+PPO22.9±5.121.4±5.525.0±4.824.0±4.020.8±10.218.8±9.950.0±3.469.8±9.2
+Joint PPO22.9±8.522.9±8.518.8±9.918.8±9.927.1±9.927.1±9.951.0±15.045.8±5.9
+PPO w/GVL18.8±1.716.1±3.113.5±5.211.5±2.124.0±5.220.8±4.839.6±10.555.2±12.0
+ATLAS(Ours)47.9±8.245.8±6.154.2±7.651.0±6.241.7±12.340.6±11.076.0±4.087.5±10.8

Table R2. Success rates (%) on RoboLab Suite 2. Mean ± SD across four repeated evaluations of the same 24 scenes per task (96 episodes/task). All methods use identical scenes.

MethodOverall OrderOverall StrictA→B OrderA→B StrictB→A OrderB→A StrictAtomic AAtomic B
π₀.₅7.3±4.33.1±4.010.4±5.43.1±4.04.2±5.93.1±4.062.5±10.220.8±9.0
+PPO15.6±1.213.5±1.215.6±6.211.5±4.015.6±5.215.6±5.270.8±14.029.2±5.9
+Joint PPO23.4±3.113.0±5.225.0±6.814.6±8.021.9±4.011.5±4.089.6±4.250.0±7.6
+PPO w/GVL7.8±2.04.7±2.611.5±4.07.3±6.24.2±0.02.1±2.452.1±9.924.0±5.2
+ATLAS(Ours)29.7±4.619.8±2.730.2±7.114.6±7.229.2±5.925.0±5.980.2±7.141.7±12.3

Table R3. Success rates (%) on RoboLab Suite 3. Mean ± SD across four repeated evaluations of the same 24 scenes per task (96 episodes/task). All methods use identical scenes.

MethodOverall OrderOverall StrictA→B OrderA→B StrictB→A OrderB→A StrictAtomic AAtomic B
π₀.₅16.7±7.48.9±4.36.2±5.42.1±4.227.1±11.015.6±7.162.5±9.049.0±12.0
+PPO17.2±3.112.5±3.84.2±0.02.1±2.430.2±6.222.9±8.758.3±3.450.0±9.0
+Joint PPO19.8±6.016.1±5.27.3±8.66.2±8.032.3±8.626.0±9.277.1±11.059.4±11.0
+PPO w/GVL17.7±3.612.5±3.88.3±6.85.2±7.927.1±2.419.8±4.057.3±13.355.2±11.5
+ATLAS(Ours)27.6±4.622.4±2.018.8±8.715.6±9.236.5±10.429.2±5.972.9±13.053.1±7.9

Table R4. Success rates (%) on RoboLab Suite 4. Mean ± SD across four repeated evaluations of the same 24 scenes per task (96 episodes/task). All methods use identical scenes.

MethodOverall OrderOverall StrictA→B OrderA→B StrictB→A OrderB→A StrictAtomic AAtomic B
π₀.₅8.3±3.82.1±1.710.4±5.41.0±2.16.2±4.23.1±2.161.5±9.841.7±12.3
+PPO19.8±3.614.1±4.919.8±4.011.5±2.119.8±9.216.7±9.079.2±5.952.1±8.7
+Joint PPO20.3±6.917.7±6.521.9±6.217.7±7.118.8±8.017.7±7.171.9±7.140.6±9.2
+PPO w/GVL16.1±5.711.5±1.217.7±5.210.4±2.414.6±7.212.5±4.856.2±14.233.3±7.6
+ATLAS(Ours)30.7±6.221.9±5.029.2±5.916.7±3.432.3±7.927.1±8.075.0±10.258.3±9.0
AuthorsOpenReview ↗

Rebuttal for Reviewer NYit [2/7]

Table R5. Paired Wilcoxon signed-rank tests across four RoboLab suites. Each comparison uses 32 paired evaluation groups (4 suites × 2 compositions × 4 repeats on the same 24 scenes). W+ / W−: sums of positive / negative ranks (n: nonzero differences). Two-sided p.

BaselineOrder: ATLAS vs. Baseline (%)W+ / W− (n)pStrict: ATLAS vs. Baseline (%)W+ / W− (n)p
π₀.₅34.0 vs. 15.8496 / 0 (31)1.2 × 10⁻⁶27.5 vs. 10.9528 / 0 (32)7.5 × 10⁻⁷
+ PPO34.0 vs. 18.9462 / 3 (30)2.3 × 10⁻⁶27.5 vs. 15.4393 / 13 (28)1.5 × 10⁻⁵
+ Joint PPO34.0 vs. 21.6352 / 26 (27)8.3 × 10⁻⁵27.5 vs. 17.4394 / 71 (30)8.7 × 10⁻⁴
+ PPO w/ GVL34.0 vs. 15.1496 / 0 (31)1.2 × 10⁻⁶27.5 vs. 11.2487 / 9 (31)2.8 × 10⁻⁶
AuthorsOpenReview ↗

Rebuttal for Reviewer NYit [3/7]

[W1-2] A simple phase-count potential, such as Φ_q(s) = n_t/N_q, would also help establish how much benefit comes from the learned within-phase potential rather than access to subtask-completion information. [Q1-2] How does it compare with shaping based only on the completed-phase count?

To assess the contribution of learned within-phase progress, we additionally evaluate a phase-count potential baseline, Φq(s)=ntNq\Phi_q(s)=\frac{n_t}{N_q}, while retaining the same hidden-state alignment. All methods are evaluated on the same 24 scenes per tasks in Tables 2 and 4. As shown in Table R6, our learned potential achieves higher success than phase-count shaping on every metric, improving overall composite success by +12.5 percentage points (Order) and +20.8 percentage points (Strict).

We attribute this gain to within-phase progress, which phase-count shaping lacks: being constant within each phase, it cannot distinguish progressing trajectories from stalled or failing ones. Accordingly, phase count achieves a within-phase AUROC of 0.500 by construction, whereas our learned potential achieves 0.643 (Table R7, 4 rounds), providing finer-grained credit assignment. We added these additional experiments in Appendix B.2 (L. 1212-1221).

Table R6. Comparison of learned task-wise and phase-count potentials on RoboLab Suite 1. Success rates (%) over 24 episodes/task, using the same scenes as Tables 2 and 4.

MethodAlignmentOverall OrderOverall StrictA→B OrderA→B StrictB→A OrderB→A StrictAtomic AAtomic B
None (Joint PPO)✗12.512.54.24.220.820.829.241.7
Phase count✓33.325.037.529.229.220.866.779.2
Learned (ATLAS)✓45.845.858.358.333.333.379.2100.0

Table R7. Time-matched AUROC across potential-fitting data budgets. AUROC measures eventual skill-success discrimination using time-matched, within-phase observations. Phase Count is 0.500 by construction.

Episodes/TaskLearned (ATLAS)Temporal ProgressPhase Count
30.5850.5140.500
120.5930.4920.500
240.6140.4850.500
96 (4 rounds)0.6430.4900.500
1920.6660.4920.500
384 (16 rounds)0.6840.4920.500

[W1-3] Appendix A.7 additionally reports 4–16 rollout rounds for fitting the potentials; these interactions and their costs should be included explicitly in budget comparisons.

In our experiments, all four suites use four rollout rounds (96 episodes per atomic task; 192 episodes / 86,400 environment steps per suite) for potential fitting, compared with 1,280 episodes / 576,000 environment steps for RL training. Potential fitting requires approximately 2 GPU-hours (20% additional GPU compute), while fitting the ridge-regression heads takes only 9 seconds on CPU. We explicitly reported these additional costs in Appendix A.7 (L. 1015–1017).

We adopt four rounds as a practical trade-off: AUROC rises from 0.643 at four rounds to 0.684 at 16 (Table R7), and already exceeds Temporal Progress (0.490) and Phase Count (0.500) at four rounds. Since potentials learned from atomic skills are reused across compositions, no additional fitting is needed per composite task. The fitting cost can be reduced further: with a single round (24 episodes per task, ~3.75% of RL environment steps), the learned potential still reaches 0.614, and even 3 episodes per task (<0.5%) yield 0.585, both above Temporal Progress and Phase Count. We added these details in Appendix B.2 (L. 1222–1234).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the Area Chair and all reviewers for their constructive feedback. Before the end of the author-reviewer discussion, we briefly summarize the reviewers' noted strengths, their concerns, and how we addressed them, with pointers to the threads.

Strengths noted by the reviewers

  1. Addresses a critical problem, long-horizon composite execution for VLAs, with an intuitive and effective solution [ceJW].
  2. Shaping and alignment address complementary problems: sparse feedback and invoking the right skill under composite instructions [NYit, atwL, ceJW].
  3. Planner-free, single-policy inference, a meaningful deployment advantage [NYit].
  4. Substantial gains with real-robot validation, and a strict-selection metric that captures faithful instruction following [NYit, ceJW].
  5. Clear structure and presentation [atwL, ceJW].

Concerns and our responses

  1. Longer sequences and other task types [NYit W1-5; atwL W1-1, Q1]. On three-subtask compositions, ATLAS roughly doubles π₀.₅ on trained orders (6.8% → 14.6%) and improves unseen orders (6.8% → 12.5%), while PPO falls below the base policy. On a new suite with stacking, ATLAS also achieves the highest composite success, without any sequence-length- or task-specific changes.

  2. Expanded evaluation and training seeds [NYit W1-4, Q4-1; atwL Q2]. We expanded simulation evaluation to 96 episodes per task (ATLAS beats all baselines, p < 10⁻³) and added three training seeds, where ATLAS shows the highest success with lower variability than the PPO baselines in both overall metrics. On the real robot, with 30 paired trials per task, ATLAS improves Overall Strict from 45.0% to 76.7% (p < 0.001), including 43.3% → 83.3% on Suite 1.

  3. Mechanisms vs. joint atomic–composite training [NYit W1-1, Q1-1]. We added a Joint atomic-composite PPO baseline with the same task sampling, phase weighting, and budget as ATLAS. Joint training alone gives only modest gains, while ATLAS outperforms it in all four suites (Order 34.0% vs. 21.6%; Strict 27.5% vs. 17.4%; both p < 10⁻³).

  4. Learned potential vs. phase count [NYit W1-2, Q1-2]. Under the same alignment, our learned potential outperforms a phase-count potential (Φ = n_t/N_q) by +12.5 (Order) and +20.8 (Strict) percentage points. It captures within-phase progress (AUROC 0.643) that phase count cannot (0.500 by construction).

  5. Cost and sample efficiency of potential fitting [NYit W1-3; atwL W3]. Fitting adds about 20% compute and the learned potentials are reused across compositions without refitting. The learned potential outperforms Temporal Progress and Phase Count at every fitting budget, even with 3 episodes per task, while exhibiting low variability across subsets, demonstrating its robustness.

  6. Method details and clarifications [NYit W2; atwL W2; ceJW]. We clarified terminal and truncation handling, corrected a truncation implementation issue and verified its limited empirical impact, and reframed potential-based shaping as a dense credit-assignment mechanism rather than a policy-invariant one. We also clarified that alignment targets the active phase's atomic instruction, and that phase information is used only during training. We also clarified the rationale for including zero-shot VLA baselines.

We sincerely thank the AC and the reviewers for their time and efforts throughout the review period. We hope that our rebuttal and revisions have relieved the major concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.