acceptodds
Under review as a conference paper at ICLR 2027

PrimitiveVLA: Learning Reusable Motion Primitives for Efficient and Generalizable Robotic Manipulation

Abstract

Vision-Language-Action (VLA) models offer a promising route to general-purpose robotic policies, yet downstream adaptation remains data-intensive, with limited transfer to new task compositions. Direct instruction-to-control fine-tuning treats multi-stage demonstrations as complete task trajectories, offering little structure for reusing recurring interaction patterns. We propose PrimitiveVLA, a primitive-centric Disassembly-and-Assembly framework. During fine-tuning, a vision-language model (VLM) infers primitive sequences, while a large language model (LLM) localizes their temporal boundaries to construct primitive-aligned samples. During inference, the VLM generates a primitive flow; the VLA executes it in a closed loop with an LLM-generated causal switching program. Unified Multimodal Representation (UMR) maps instance-specific descriptions to unified primitive instructions while grounding objects and targets through mask-augmented observations. Across Libero, RLBench, and real-robot experiments, PrimitiveVLA improves data efficiency and zero-shot transfer to unseen task compositions and long-horizon tasks while maintaining or improving in-distribution performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, motion primitives, compositional generalization, data efficiency

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

PrimitiveVLA splits demonstrations into reusable motion primitives, trains a VLA on them, and assembles them for new tasks. Experiments cover LIBERO, RLBench, and a real robot.

Critical strengths
  1. The gains on new task compositions are large across three backbones. For example, OpenVLA improves from 9.25% to 45.50% on LIBERO-90-Novel (Table 2).

  2. Real-robot results and component ablations support the approach. And the approach is quite reasonable to me.

  3. the writing is clear and easy to follow.

Critical weaknesses
  1. The long-horizon tests reset the arm between subtasks, so they do not establish perforemance on uninterrupted execution.

  2. The data-efficiency claim is less convincing: the half-data OpenVLA gain is only 0.57 percentage points, without training-seed uncertainty. Changed sampling also makes the source of the gains unclear.

  3. There are so many model used in the pipeline: Qwen3-VL,DeepSeek-V3,Qwen2.5-VL-72B + SAM + Cutie... Is this necessary? Why choose a different model?

Questions for authors
  1. How does the method perform without arm reset between subtasks?
  2. So Do the data-efficiency gains hold across seeds and with matched primitive/frame sampling?
  3. Why use Qwen3.5-Omni-Plus for direct segmentation rather than the same Qwen3-VL used in your pipeline(Do you use audio input)? A comparison using the same model and equivalent inputs would better isolate the benefit of program-based segmentation.
7 comments
AuthorsOpenReview ↗

Response to W1 and Q1 (1/2)

We appreciate the reviewer's comments on compositional generalization, the real-robot results, and the ablation experiments. We address the concerns about uninterrupted execution, data efficiency, and model choice in separate responses.

W1 and Q1. Performance without external arm resets

The submitted LIBERO-Long results with external resets do not establish uninterrupted execution with the original checkpoints. Without those resets, the original OpenVLA checkpoints achieve 0.00% and π0.5 + PrimitiveVLA achieves 20.50%. In a separate evaluation using reset-augmented training data, π0.5 + PrimitiveVLA reaches 78.50% without external resets.

The LIBERO-Long results in the paper use a protocol that resets the arm between subtasks while preserving object states. The baselines use the primitive completion checks in the causal switching program to detect when an arm reset is needed and apply the same reset conditions. The program only triggers arm resets for the baselines; their action policies receive the full long-horizon task instruction.

Using the existing checkpoints fine-tuned for the paper, we completed the evaluation without external arm resets on the LIBERO-Long tasks reported in the paper. For each task, we used the same 50 initial states, success criteria, and total execution budget as in the paper. The with-reset results in the table are from Table 2 of the submitted paper.

BackboneMethodLIBERO-Long (external reset)LIBERO-Long (no external reset)
OpenVLAStandard fine-tuning4.50%0.00%
OpenVLAPrimitiveVLA38.50%0.00%
π0.5Standard fine-tuning30.50%1.25%
π0.5PrimitiveVLA80.25%20.50%

Without external arm resets, success falls sharply in all four settings. With π0.5, PrimitiveVLA remains higher than standard fine-tuning (20.50% vs. 1.25%), though absolute success is low; both OpenVLA methods score 0.00%.

AuthorsOpenReview ↗

Response to W1 and Q1 (2/2)

W1 and Q1 (continued). Learned reset for uninterrupted execution

To test whether reset supervision reduces the arm-state mismatch between subtasks, we keep all original demonstrations and randomly select roughly one quarter of demonstrations per task for attempted augmentation, retaining only validated arm-return segments planned offline by a motion planner from task success to the initial arm pose. In total, 818 demonstrations received a validated return segment, each yielding one appended reset primitive episode. Both methods train on the same original trajectories and added reset frames: standard fine-tuning uses complete augmented trajectories with the full-task instruction, while PrimitiveVLA segments them and labels each return as reset. This primitive is exclusive to the supplementary experiment; the original method retains its fixed set of 11.

The reset-augmented comparison uses π0.5 with the paper's fine-tuning step count and total execution budget. Both data settings are evaluated without external arm resets on the same tasks and initial states. After each nonfinal subtask, PrimitiveVLA switches to the learned reset primitive, and the VLA generates actions that return the arm toward its initial pose. The causal switching program advances to the next subtask when changes in the observed end-effector pose remain below the stopping thresholds for several consecutive steps, even if the arm has not fully returned to its initial pose. Every arm action during task execution comes from the VLA, and reset actions count toward the same budget.

MethodLIBERO-Long (original data)LIBERO-Long (reset-augmented data)
Standard fine-tuning1.25%14.25%
PrimitiveVLA20.50%78.50%

With π0.5 and no external arm resets, adding the same reset trajectories raises success from 1.25% to 14.25% for standard fine-tuning and from 20.50% to 78.50% for PrimitiveVLA, with the same number of fine-tuning steps. PrimitiveVLA generates the return actions through a learned reset primitive, reaching 78.50%, only 1.75 percentage points below the original checkpoint's 80.25% under external resets. The drop in success without external resets and the recovery after reset training indicate that the mismatch between arm states at subtask transitions and the starting states in the short-horizon demonstrations is a major contributor to failures in uninterrupted execution in this π0.5 evaluation; external resets or training with the learned reset primitive can alleviate this mismatch.

External resets align the arm with the next subtask's starting state, which the original short-horizon demonstrations do not teach the policy to reach. Removing this alignment lowers success; reset-augmented training lets π0.5 PrimitiveVLA generate the return actions itself. In the revision, we will label the execution protocol for each set of LIBERO-Long results and report the additional evaluations of execution without external arm resets.

AuthorsOpenReview ↗

Response to W2 and Q2

W2 and Q2. Data-efficiency gains across seeds and matched sampling

Updated to include the third OpenVLA training seed; the results below now summarize three seeds.

In the repeated OpenVLA comparison, PrimitiveVLA with 50% of the demonstrations remains comparable to standard fine-tuning with 100% across three training seeds. Both use the unchanged official frame-level sampler, and we add no primitive-level sampling weights.

For OpenVLA, Table 3 compares PrimitiveVLA trained on 50% of the demonstrations with standard fine-tuning on 100%. The original single-run difference is small, and the paper did not quantify variation across training seeds. We reassess this comparison with a three-seed result.

In this three-seed OpenVLA data-efficiency comparison, standard fine-tuning and PrimitiveVLA use the official frame-level RLDS sampler without modifying its sampling logic or adding primitive-level weights. OpenVLA flattens trajectories into single-step samples before shuffling and predicts one action per frame. For a fixed set of demonstrations, primitive segmentation does not introduce episode-level reweighting or change one-step action targets. The three-seed results therefore compare the methods under the same sampling rule while retaining the intended 100% versus 50% data difference.

We repeated the 50%-data PrimitiveVLA and 100%-data standard fine-tuning runs with the original seed and two additional seeds fixed before fine-tuning, independently of the results. Both methods followed the corresponding data selection, training settings, and evaluation protocol described in the paper. The table reports the mean and sample standard deviation across the three seeds.

MethodData ScaleLIBERO-ObjectLIBERO-SpatialLIBERO-GoalLIBERO-90Mean
Standard fine-tuning100%87.20% ± 0.35%84.00% ± 3.47%75.20% ± 2.80%69.66% ± 1.15%79.02% ± 0.44%
PrimitiveVLA50%84.93% ± 2.32%85.80% ± 1.31%74.53% ± 1.17%71.27% ± 1.43%79.13% ± 0.84%

For each training seed, we first average the success rates across the four suites and then calculate the mean and standard deviation across seeds. Across the three training seeds, PrimitiveVLA with 50% of the data achieved a four-suite mean success rate of 79.13%, comparable to 79.02% for standard fine-tuning with 100% of the data.

Appendix D.7 attributes more sampling opportunities for short primitives to splitting demonstrations into episodes. This explanation is imprecise for the official samplers used in our experiments. We use the official training samplers for OpenVLA, OpenVLA-OFT, and π0.5 without modifying their sampling logic or adding primitive-level weights. All three use frame-indexed sampling, so increasing the episode count alone does not increase the sampling weight of short primitives. At segment boundaries, OpenVLA-OFT can lose valid action-window starts, while π0.5 retains frame-indexed starts but can change future-action targets. OpenVLA predicts one action per frame, so these action-window boundary effects do not apply to it. We will revise Appendix D.7 to describe these backbone-specific sampling effects.

Reviewer_BR24OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is motivated by the idea that robot trajectories can be decomposed into stages that share similar motions across tasks. For example, pick and place can be decomposed into grasp, move, and place. The proposed method, PrimitiveVLA, uses a VLM to infer the primitive sequence and an LLM-generated program to locate temporal boundaries. It then combines standardized primitive instructions with object masks for VLA fine-tuning. At inference, these primitives are assembled into a sequence to complete the task. This approach allows demonstrations from different tasks to contribute to learning shared behaviors. Experiments on LIBERO, RLBench, and a UR5e robot show improved data efficiency and performance on unseen task compositions and long-horizon tasks, while in-distribution performance generally remains comparable or improves.

Critical strengths

The paper addresses a clear limitation of task-level fine-tuning. Similar motions appear across many tasks, but conventional supervision does not explicitly align them. Organizing training around shared primitives provides a concrete way to reuse demonstrations, making the approach well motivated.

Critical weaknesses

None

Questions for authors
  1. How much manual work was needed to get the segmentation and switching programs working? For example, did you revise prompts, select examples, tune thresholds, or fix generated code? How much of this would need to be repeated for a new dataset?

  2. In LIBERO-Long, how do you decide when each subtask ends and when to reset the arm? Do the baselines use the same information and reset rules? Have you tested whether the improvements hold without resets between subtasks?

4 comments
AuthorsOpenReview ↗

Response to Q1

We appreciate the reviewer's assessment that organizing training around shared primitives is well motivated, and the questions about the manual work involved in program construction and the arm-reset protocol for long tasks.

Q1. Manual preparation and program reuse across datasets

The original experiments used the 11-primitive vocabulary (Section 3.2). For each data setting, we specify the state and action fields, select three reference segmentations for each primitive represented in its training data (Appendix A.2), and revise primitive definitions or completion rules when needed. We test the LLM-generated offline segmentation program on all reference trajectories. In previous settings, 1–3 generation-and-check rounds were needed before the generated program and thresholds passed all examples; we corrected logic or runtime errors but did not manually adjust numerical thresholds. We then fix the program and thresholds across tasks. For a new dataset, we reuse them if its interface and required motions are compatible, or repeat the relevant preparation.

For online causal switching, the LLM generates the causal switching program before execution from the primitive definitions, the calibrated offline segmentation program, and the length of the causal history window (Section 4.3.2 and Appendix A.4). We generate it without using evaluation trajectories or manually annotated evaluation boundaries. We reuse the LIBERO-90 causal switching program for both LIBERO-90-Novel and LIBERO-Long (Appendix B.1). If a new dataset requires recalibrating the offline segmentation program, we use the updated offline segmentation program to regenerate the causal switching program.

Reuse depends on the data interface and motion conditions. Tasks in the same environment with compatible state and action conventions can reuse the offline segmentation program and thresholds. We also observed threshold reuse across robot platforms. For the same set of real-world tasks, the numerical thresholds calibrated on UR5e data can be applied to Franka data without adjustment. We specify the state and action fields according to each platform's data interface.

We will document these program-reuse conditions and the required dataset-level preparation in the revision.

AuthorsOpenReview ↗

Response to Q2 (1/2)

Q2. Subtask boundaries and arm-reset protocol

In the submitted LIBERO-Long evaluation, the final primitive of each nonfinal subtask triggers an external arm reset; the baselines use the same completion checks and reset protocol. Removing external resets sharply lowers success with the original checkpoints, so the submitted results do not establish uninterrupted execution with those checkpoints. In a separate π0.5 experiment, a learned reset primitive raises PrimitiveVLA to 78.50% without external resets.

The LIBERO-Long tasks evaluated in the paper consist of short-horizon LIBERO-90 tasks, which define the subtask boundaries. At each nonfinal boundary, the causal switching program uses the final primitive's completion check to trigger an external reset that returns the arm to its initial pose while preserving object states. For the baselines, this check only triggers the reset; their policies continue to receive the full long-horizon instruction. The simulator determines full-task success. We will clarify these boundaries, reset triggers, baseline inputs, and success criteria in Section 5.3 and Appendix B.1, which describe the shared protocol.

Using the existing checkpoints fine-tuned for the paper, we completed the evaluation without external arm resets on the same LIBERO-Long tasks. For each task, we used the same 50 initial states, success criterion, and total execution budget as in the paper. The with-reset results in the table are from Table 2 of the submitted paper.

BackboneMethodLIBERO-Long (external reset)LIBERO-Long (no external reset)
OpenVLAStandard fine-tuning4.50%0.00%
OpenVLAPrimitiveVLA38.50%0.00%
π0.5Standard fine-tuning30.50%1.25%
π0.5PrimitiveVLA80.25%20.50%

Without external arm resets, success falls sharply in all four settings. With π0.5, PrimitiveVLA remains higher than standard fine-tuning (20.50% vs. 1.25%), though absolute success is low; both OpenVLA methods score 0.00%.

AuthorsOpenReview ↗

Response to Q2 (2/2)

Q2 (continued). Reset-augmented execution without external arm resets

To test whether reset supervision reduces the arm-state mismatch between subtasks, we keep all original demonstrations and randomly select roughly one quarter of demonstrations per task for attempted augmentation, retaining only validated arm-return segments planned offline by a motion planner from task success to the initial arm pose. In total, 818 demonstrations received a validated return segment, each yielding one appended reset primitive episode. Both methods train on the same original trajectories and added reset frames: standard fine-tuning uses complete augmented trajectories with the full-task instruction, while PrimitiveVLA segments them and labels each return as reset. This primitive is exclusive to the supplementary experiment; the original method retains its fixed set of 11.

The reset-augmented comparison uses π0.5 with the paper's fine-tuning step count and total execution budget. Both data settings are evaluated without external arm resets on the same tasks and initial states. After each nonfinal subtask, PrimitiveVLA switches to the learned reset primitive, and the VLA generates actions that return the arm toward its initial pose. The causal switching program advances to the next subtask when changes in the observed end-effector pose remain below the stopping thresholds for several consecutive steps, even if the arm has not fully returned to its initial pose. Every arm action during task execution comes from the VLA, and reset actions count toward the same budget.

MethodLIBERO-Long (original data)LIBERO-Long (reset-augmented data)
Standard fine-tuning1.25%14.25%
PrimitiveVLA20.50%78.50%

With π0.5 and no external arm resets, adding the same reset trajectories raises success from 1.25% to 14.25% for standard fine-tuning and from 20.50% to 78.50% for PrimitiveVLA, with the same number of fine-tuning steps. PrimitiveVLA generates the return actions through a learned reset primitive, reaching 78.50%, only 1.75 percentage points below the original checkpoint's 80.25% under external resets. Together, these results indicate that the mismatch between arm states at subtask transitions and the starting states in the short-horizon demonstrations is a major contributor to failures in uninterrupted execution in this π0.5 evaluation; external resets or training with the learned reset primitive can alleviate this mismatch. We report the augmented-data experiment separately from the original-checkpoint evaluation.

External resets align the arm with the next subtask's starting state, which the original short-horizon demonstrations do not teach the policy to reach. Removing this alignment lowers success; reset-augmented training lets π0.5 PrimitiveVLA generate the return actions itself. In the revision, we will label the execution protocol for each set of LIBERO-Long results and report the additional evaluations of execution without external arm resets.

AuthorsOpenReview ↗

Follow-up to Reviewer BR24

Thank you for taking the time to review our work. We have provided detailed responses to your questions about manual program preparation, program reuse, and the LIBERO-Long arm-reset protocol, including evaluations without external resets. We have also outlined the corresponding changes we plan to make in the manuscript.

We would appreciate your feedback on whether we have adequately answered your questions. If anything remains unclear, we would be happy to provide further clarification during the discussion period.

Thank you again for your consideration.

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

PrimitiveVLA fine-tunes VLAs on primitive-aligned segments instead of full task trajectories. A VLM infers each demo's primitive sequence from a fixed set of 11 hand-defined primitives, and an LLM-generated state/action program localizes the boundaries. A "Unified Multimodal Representation" (UMR) then replaces instance-specific language with templated primitive instructions plus SAM/Cutie object masks. At inference, a VLM (retrieval-augmented) produces a primitive flow, and an LLM-written causal program decides when to switch primitives. On Libero, RLBench and a UR5e, with OpenVLA, OpenVLA-OFT and π0.5, the paper reports gains in zero-shot compositional generalization (e.g., Libero-90-Novel for OpenVLA-OFT from 13.5% to 71.0%), long-horizon tasks, and data efficiency. In-distribution performance is roughly maintained.

Critical strengths
  1. The method is practical and cleanly motivated: the primitive disassembly is automated, and the boundary segmentation is validated against manual annotation.

  2. The gains on compositional and long-horizon settings are large and consistent across three backbones and in real-robot experiments (π0.5 long-horizon from 10% to 65%). The ablations and failure-mode analysis help separate the contributions of UMR and disassembly.

Critical weaknesses
  1. The OOD comparison is not apples-to-apples: the method adds object masks, a VLM planner with retrieval over the training library, and a code-based switching controller, none of which the baselines receive. The "w/o Disass." ablation (full-task instruction + mask) already recovers most of the Novel gain for OpenVLA-OFT (60.0% versus 71.0%), so a stronger baseline with a planner and masks would better show what primitive decomposition itself contributes. This makes the attribution of gains to the primitive paradigm hard to verify.

  2. Statistical rigor and generality are limited: the authors state they do not quantify variance across training seeds, yet many claims rest on differences of a few points on 50-trial or 20-trial evaluations (e.g., "matching full-data performance" is 80.10% vs. 79.53%). The primitive set is designed around Libero-style tasks, and the Libero-Long evaluation excludes two of the ten tasks with unseen subtasks, with several RLBench results being very low (OpenVLA-OFT at 4.5%), so the claims of broad generalization are weaker than presented.

Questions for authors
  1. Can you provide baselines that get the same inputs as PrimitiveVLA (a VLM-generated subtask plan, e.g., π0.7-style subtask language, plus object masks) but without primitive-level segmentation and unified instructions? This would be the most decisive evidence for the primitive paradigm.

  2. Could you report results across multiple training seeds, at least for the main Libero-90-Novel and data-efficiency comparisons?

  3. Why were the two Libero-10 tasks with unseen subtasks excluded, and how does the method behave on them? Relatedly, how does it fail when a required primitive is missing from the fixed set of 11?

  4. How sensitive is performance to errors in the VLM-generated primitive flow at test time, given that the retrieval library comes from the same tasks used for training?

8 comments
AuthorsOpenReview ↗

Response to W1.

We thank the reviewer for recognizing the automated primitive disassembly, the segmentation validation, and the results across backbones and on a real robot. We agree that a matched subtask-level baseline is needed to more directly assess the contribution of primitive-level organization. We address the reviewer's remaining concerns in separate responses.

W1. Fairness of the OOD comparison

Even when a baseline also uses object masks, a VLM planner with retrieval over the training library, and a code-based switching controller, PrimitiveVLA improves success on LIBERO-90-Novel and LIBERO-Long. These added components therefore do not fully explain the gains of the primitive paradigm.

The Full-task + Mask ablation already tests masked observations, but its policy follows one full-task instruction throughout execution. Retrieval-based subtask planning and instruction switching by the causal switching program therefore have no role in that baseline. To compare methods while these components are used, we add a VLM-generated subtask + mask baseline. Like PrimitiveVLA, it uses object-mask generation and tracking, VLM planning with retrieval from its subtask-level library, and a causal switching program based on the same primitive completion checks.

For the VLM-generated subtask + mask baseline, we use the same task instruction and initial RGB image as PrimitiveVLA. The VLM retrieves top-3 examples from a subtask-level library built from decompositions of the same training data and generates semantic subtask instructions. Each subtask aligns with a contiguous primitive subsequence, allowing the same state/action-based offline segmentation program to locate its endpoint. We train this baseline on continuous subtask trajectories with those instructions and masked observations, using the same source demonstrations, training steps, and batch size. During evaluation, the causal switching program tracks primitive completion and updates the instruction and mask targets after each subtask's final primitive; single-stage tasks need no switch. We match the paper's initial states, success criteria, execution budget, and LIBERO-Long arm-reset protocol. The w/o Disass. (Full-task + Mask) and PrimitiveVLA results come from Table 5.

BackboneMethodLIBERO-90LIBERO-90-NovelLIBERO-Long
OpenVLAw/o Disass. (Full-task + Mask)79.60%33.75%4.25%
OpenVLAVLM-generated subtask + mask76.14%41.50%25.00%
OpenVLAPrimitiveVLA79.88%45.50%38.50%
OpenVLA-OFTw/o Disass. (Full-task + Mask)94.30%60.00%39.75%
OpenVLA-OFTVLM-generated subtask + mask85.86%55.50%55.00%
OpenVLA-OFTPrimitiveVLA94.70%71.00%66.50%

The subtask baseline improves LIBERO-Long over Full-task + Mask with both backbones, although it falls below that ablation on LIBERO-90-Novel with OpenVLA-OFT (55.50% vs. 60.00%). PrimitiveVLA exceeds the subtask baseline by 13.50 and 11.50 points on LIBERO-Long, 4.00 and 15.50 on LIBERO-90-Novel, and 3.74 and 8.84 on LIBERO-90 (OpenVLA and OpenVLA-OFT, respectively).

The subtask baseline and PrimitiveVLA both use VLM planning with top-3 retrieval, object masks, and a causal switching program based on the same completion checks. These results show that, with these components shared, the primitive paradigm still helps novel-task generalization and long-horizon execution. In the revision, we will present this matched comparison alongside the existing full-system and mask ablations.

AuthorsOpenReview ↗

Response to Q1

Q1. Component-matched subtask baseline

The VLM-generated subtask + mask baseline described in our response to W1 uses semantic subtask instructions, as in the reviewer's π0.7-style subtask-language example, and masked observations; it does not use primitive-level training episodes or unified primitive instructions. The table in our W1 response shows PrimitiveVLA outperforming it on LIBERO-90, LIBERO-90-Novel, and LIBERO-Long with both OpenVLA and OpenVLA-OFT.

With OpenVLA-OFT, the subtask baseline improves LIBERO-Long success from 39.75% for Full-task + Mask to 55.00%, while PrimitiveVLA reaches 66.50%. On LIBERO-90-Novel, however, the subtask baseline reaches 55.50%, compared with 60.00% for Full-task + Mask and 71.00% for PrimitiveVLA.

Semantic subtasks help organize the successive stages of a long task. However, some LIBERO-90-Novel tasks contain only one semantic subtask, which may still involve several distinct motions. In these tasks, subtask-level training and execution keep those motions together under one instruction. PrimitiveVLA separates recurring motions and aligns them across tasks under shared primitive instructions, so the policy can reuse and compose them even within a single semantic stage (Sections 1 and 3.1). Appendix D.1 supports this mechanism: same-primitive segments are more similar across tasks than different-primitive segments. This offers a possible explanation for PrimitiveVLA's advantage on LIBERO-90-Novel.

The matched results therefore support a benefit from the primitive paradigm beyond semantic staging. We will include the baseline and this analysis in the revision.

AuthorsOpenReview ↗

Response to Q2

Q2. Training-seed variability in Novel and data efficiency

Updated to include the third OpenVLA training seed; the OpenVLA results below now summarize three seeds.

Across the completed training-seed repeats, PrimitiveVLA improves LIBERO-90-Novel and LIBERO-Long success at every evaluated OpenVLA and π0.5 seed. With OpenVLA, its four-suite mean using 50% of the demonstrations remains comparable to standard fine-tuning using 100%.

The original experiments followed the backbone-specific settings in Appendix B.3 without searching over training seeds, and the paper did not quantify training-seed variability through repeated fine-tuning. To assess that variability, we evaluated OpenVLA and π0.5 with three training seeds each: each backbone's original seed and the same two additional seeds, fixed before fine-tuning independently of the results. The OpenVLA data-efficiency comparison uses the same three OpenVLA seeds. All repeats follow the paper's data selection, fine-tuning settings, and evaluation protocol; the tables report mean ± sample standard deviation across training seeds. The number of repeats was limited by the time and compute available during the discussion period.

OpenVLA data efficiency.

MethodData ScaleLIBERO-ObjectLIBERO-SpatialLIBERO-GoalLIBERO-90Mean
Standard fine-tuning100%87.20% ± 0.35%84.00% ± 3.47%75.20% ± 2.80%69.66% ± 1.15%79.02% ± 0.44%
PrimitiveVLA50%84.93% ± 2.32%85.80% ± 1.31%74.53% ± 1.17%71.27% ± 1.43%79.13% ± 0.84%

For the four-suite aggregate, we first average the success rates across suites within each seed, then compute the mean and standard deviation across seeds. Across the three training seeds, PrimitiveVLA with 50% of the data achieved a four-suite mean success rate of 79.13%, comparable to 79.02% for standard fine-tuning with 100% of the data.

OpenVLA and π0.5 generalization.

Both backbones use the original LIBERO-90 training set and their respective fine-tuning configurations. We follow the evaluation settings in the paper for LIBERO-90, LIBERO-90-Novel, and LIBERO-Long.

BackboneMethodLIBERO-90LIBERO-90-NovelLIBERO-Long
OpenVLAStandard fine-tuning69.66% ± 1.15%9.67% ± 1.42%3.33% ± 1.26%
OpenVLAPrimitiveVLA79.90% ± 0.16%44.58% ± 0.80%37.75% ± 2.46%
π0.5Standard fine-tuning96.27% ± 0.46%48.58% ± 6.77%45.00% ± 13.34%
π0.5PrimitiveVLA95.29% ± 0.78%72.50% ± 2.70%79.92% ± 3.01%

On LIBERO-90-Novel and LIBERO-Long, PrimitiveVLA improves over standard fine-tuning at every evaluated seed for both backbones. The smallest observed gains were 19.50 and 20.00 percentage points, respectively. On LIBERO-90, mean success increases from 69.66% to 79.90% with OpenVLA, and remains close to standard fine-tuning with π0.5 (95.29% vs. 96.27%).

We will report variability across training seeds and update the abstract and data efficiency discussion accordingly.

Comments on the paper

1 comment
Area_Chair_TmR7OpenReview ↗

Reminder: Please Review Authors’ Rebuttal and Provide Feedback

Dear Reviewers,

Thank you for your reviews. If you have not already done so, please read the authors' rebuttal and share your feedback in the discussion thread at your earliest convenience. Your timely responses would be greatly appreciated.

Best,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.