acceptodds
Under review as a conference paper at ICLR 2027

PGOT: Physically Grounded Optimal Transport for Distilling Vision-Language-Action Models

Abstract

Vision–language–action (VLA) models bring large-scale visual and semantic knowledge to embodied control, but their computational demands motivate distillation into compact counterparts for responsive deployment. Existing distillation objectives, however, fall short in capturing the physical semantics and decision structure of embodied control: categorical matching of token distributions treats physical commands as unstructured labels, while unselective supervision dilutes learning at critical decision points and passes imperfect teacher decisions directly to students. We introduce **Physically Grounded Optimal Transport (PGOT)**, a distillation framework that addresses these gaps through three complementary mechanisms. **Geometry** aligns action distributions through optimal transport of probability mass, grounding teacher–student discrepancies in physical action space. **Criticality** exploits the temporal structure of expert trajectories to identify pivotal decision points and concentrate learning on them. **Reliability** adjusts teacher guidance according to its physical agreement with expert demonstrations, favoring trustworthy supervision across states and action dimensions. Across UAV navigation and robotic manipulation, PGOT distills 7B/9B teachers into 0.8B students that retain teacher-level success with approximately 9–11 fewer parameters. Over the strongest distillation baselines, it improves success rate by 2.67 and 2.40 percentage points and three-out-of-three completion by 7.33 and 6.00 points, respectively, exceeding even the teachers on the latter. Analyses of learning dynamics and closed-loop behavior further reveal that physical transport supplies optimization directions largely distinct from categorical matching, while selective supervision improves outcomes at critical maneuvers and limits the inheritance of teacher errors, with the largest gains where the teacher itself is unreliable.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Knowledge Distillation, Vision-Language-Action Models, Optimal Transport, Embodied AI, Model Compression, Robot Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_r2sWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PGOT distills discrete-action VLA models using an action-space Wasserstein loss, higher weights at changes in expert behavior, and teacher supervision filtered by agreement with demonstrations. The method compresses 7B/9B teachers into 0.8B students without changing student inference. Experiments on simulated UAV navigation and LIBERO-10 show improved success and three-out-of-three completion, supported by component ablations and action-distribution diagnostics.

Critical strengths
  1. Using decoded action distances is a sensible improvement over treating all token errors equally. The one-dimensional transport loss is simple to compute, and the method handles the categorical gripper separately.

  2. The gains appear in two different control domains. Table 1 shows that the small students retain teacher-level success while improving repeated completion, which is relevant to practical deployment.

  3. The component analysis is unusually detailed. Leave-one-out ablations, paired completion counts, and measurements of distant action errors help explain where the improvements come from.

Critical weaknesses
  1. All models use training seed 42. The reported standard deviations cover three decoding rounds, not independent training runs (Section 4.1). This leaves the reliability of the 2.40–2.67-point success gains unclear.

  2. Pass3 measures consistency under token sampling at temperature 0.3. Without greedy decoding or a temperature sweep, it is hard to separate better control from reduced sensitivity to decoding randomness.

  3. The method is sensitive to the transport weight: on LIBERO-10, success drops from 49.47% at weight 10 to 43.00% or 41.87% at weights 5 or 20 (Appendix H). The selection protocol and comparable tuning budgets for the adapted baselines are not clear enough.

Questions for authors
  1. Do the gains over FKL and Shallow-π persist across independent training seeds, with uncertainty computed over both trained models and evaluation configurations?

  2. Does PGOT retain its advantage under greedy decoding and across sampling temperatures, when SR and Pass3 are reported together?

  3. How were the loss weights selected without using the final evaluation tasks, and what tuning budget was given to each baseline?

5 comments
AuthorsOpenReview ↗

Response to Reviewer r2sW (1/3)

We thank the reviewer for the detailed review. We appreciate the recognition that using decoded action distances is a sensible improvement over treating all token errors equally, that the gains appear in two different control domains with relevance to practical deployment, and that our component analysis is unusually detailed. The questions on training seeds, decoding, and hyperparameter selection are important for the reliability of our conclusions, and we address each with new experiments below.

1. Independent training seeds (W1, Q1). We retrained PGOT, FKL, and Shallow-π on CoNav-UAV with seeds 43 and 44 (App. J, Table 12). The aggregate SR uncertainty is computed over all nine runs (3 trained models × 3 evaluation rounds), so it covers both training and evaluation variation.

PolicySR, seeds 42 / 43 / 44SR, all 9 runsPass³, seeds 42 / 43 / 44Pass³, mean ± std over seeds
FKL85.33 / 87.11 / 86.0086.15 ± 1.5974.67 / 76.67 / 78.0076.44 ± 1.68
Shallow-π85.33 / 82.44 / 83.1183.63 ± 1.6074.00 / 73.33 / 70.6772.67 ± 1.76
PGOT88.00 / 87.56 / 87.1187.56 ± 1.9482.00 / 80.67 / 80.6781.11 ± 0.77
  • PGOT is best under every seed. It has the highest SR and the highest Pass³ for each of seeds 42, 43, and 44.
  • The Pass³ gain is robust to seed variation. Averaged over seeds, PGOT improves Pass³ by +4.67 over FKL and +8.44 over Shallow-π. Its lowest seed (80.67) exceeds the highest seed of either baseline (78.00), and it has the smallest seed-to-seed standard deviation (0.77). The differences are significant under Welch's t-test over the three seeds (p = 0.025 against FKL and p = 0.006 against Shallow-π), so the improvement in execution consistency is not an effect of a particular seed.
  • The Pass³ gain is also significant on LIBERO-10. In paired comparisons over the 500 initial states (exact McNemar test on the counts in App. F.3), p = 0.024 against FKL and p < 0.001 against Shallow-π.
  • The SR gain is consistent across seeds. Averaged over seeds, PGOT improves SR by +1.41 over FKL and +3.93 over Shallow-π, and its seed-averaged SR (87.56) equals that of the 7B teacher. This matches the main claim of the paper: PGOT retains teacher-level success, and its advantage over the baselines lies primarily in consistency.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer r2sW (2/3)

2. Greedy decoding and temperature sweep (W2, Q2). We evaluated all methods with greedy decoding (T = 0) and at sampling temperatures T ∈ {0.3, 0.5, 0.7} (App. K, Table 13), covering the teacher, the Student SFT reference, and all baselines. Entries are SR / Pass³. "Best over T" gives each method's best value over the four settings, separately for each metric; bold marks the best student value:

PolicyT = 00.30.50.7Avg.Best over T
Teacher SFT84.89 / 74.6787.56 / 78.6782.00 / 70.0081.33 / 64.6783.94 / 72.0087.56 / 78.67
Student SFT50.67 / 41.3353.33 / 38.6757.33 / 42.0050.67 / 33.3353.00 / 38.8357.33 / 42.00
FKL82.89 / 73.3385.33 / 74.6784.67 / 74.0084.44 / 68.6784.33 / 72.6785.33 / 74.67
RKL81.78 / 74.6778.00 / 68.0083.11 / 74.6776.67 / 62.0079.89 / 69.8383.11 / 74.67
ABKD80.44 / 70.6780.89 / 67.3379.78 / 66.6780.44 / 65.3380.39 / 67.5080.89 / 70.67
ToDi74.89 / 63.3375.56 / 62.0078.22 / 67.3376.67 / 64.6776.33 / 64.3378.22 / 67.33
Shallow-π83.11 / 76.0085.33 / 74.0085.78 / 74.6783.78 / 72.0084.50 / 74.1785.78 / 76.00
VLA-AD74.00 / 62.0074.89 / 58.0079.33 / 65.3374.22 / 56.6775.61 / 60.5079.33 / 65.33
PGOT84.22 / 75.3388.00 / 82.0089.11 / 76.6787.11 / 77.3387.11 / 77.8389.11 / 82.00

The reviewer's question is whether PGOT's gain comes only from a more concentrated action distribution that is less affected by sampling, in which case greedy decoding would remove the gap. The sweep shows that this is not the case:

  • Sampling is beneficial, so greedy decoding is not a substitute. Every policy in the table reaches its highest SR at T = 0.3 or 0.5, not under greedy decoding (e.g., Teacher SFT 84.89 → 87.56, FKL 82.89 → 85.33, Shallow-π 83.11 → 85.78, PGOT 84.22 → 89.11). Moderate sampling is therefore the better operating regime for these policies. Switching to greedy decoding would not close the gap: it lowers the SR of every method relative to its best setting.
  • PGOT leads without sampling. Under greedy decoding, PGOT has the highest student SR (84.22 vs. 83.11 for Shallow-π and 82.89 for FKL), and its Pass³ is on par with the best baseline (75.33 vs. 76.00, one task out of 150). Part of the gain therefore comes from better most-likely actions, not only from robustness to sampling.
  • PGOT leads at every sampling temperature and under each method's best setting. It has the best student SR and Pass³ at T = 0.3, 0.5, and 0.7. With the best temperature chosen separately for each method and metric, PGOT leads the best baseline by 3.33 SR and 6.00 Pass³ points, so its advantage cannot be removed by changing the decoding configuration of the baselines.
  • PGOT keeps the benefit of sampling while reducing its downside. Sampling raises SR, but it occasionally executes actions far from the most likely one. Geometry reduces the student's probability on such actions (17.26% less mass at 32 or more bins from the expert, §4.4.1). PGOT therefore gains more from sampling than the two strongest baselines: +4.89 SR from greedy decoding to its best temperature, against +2.44 for FKL and +2.67 for Shallow-π. Robustness to sampling is thus a desirable property of the learned action distribution: the policy benefits from sampling without the failures that far-off sampled actions would cause.

Finally, Pass³ remains below SR at T = 0 because the simulator is not fully deterministic: images rendered at the same pose differ slightly between runs, so repeated greedy executions can diverge. We revised the description of Pass³ in §4.1, which previously attributed the gap between SR and Pass³ to decoding noise alone.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer r2sW (3/3)

3. Loss-weight selection and tuning budgets (W3, Q3).

  • Selection. For PGOT, we searched the weight of its core component, Geometry (w₁), over {5, 10, 20} using a single training seed (42) and decoding temperature (0.3). The same value, w₁ = 10, is best in both domains and is used throughout (App. H).
  • Protocol. All methods share data, learning rate, batch size (128), and number of steps (9,785 on CoNav-UAV and 10,000 on LIBERO-10). They are evaluated at their final checkpoints without early stopping or checkpoint selection (§4.1).
  • Equal budget. To ensure a fair comparison, we gave the two strongest baselines the same search budget under this protocol: one key coefficient each, searched over three values (new Table 3, App. C):
SettingSRPass³CRMMBC
FKL, λ_FKL = 1 (default)85.3374.6711.785.27
FKL, λ_FKL = 2 / 373.56 / 75.7854.67 / 58.6726.22 / 23.563.54 / 3.72
Shallow-π, λ_attn = 1 (default)85.3374.0010.895.09
Shallow-π, λ_attn = 2 / 370.00 / 73.5653.33 / 59.3329.56 / 25.333.52 / 3.81
PGOT, w₁ = 10 (default)88.0082.008.005.72
PGOT, w₁ = 5 / 2082.44 / 83.1171.33 / 75.3311.33 / 11.565.41 / 5.46
  • The comparison uses matched tuning budgets. Both baselines perform best at their default settings, which are those reported in the main table. Table 1 therefore compares the best of three settings for PGOT with the best of three for each baseline, and PGOT leads on every metric shown.
  • The advantage holds in experiments not used to select w₁. The seed and temperature studies in our responses 1 and 2 above were run afterwards with w₁ = 10 unchanged. PGOT still achieves the highest student SR under seeds 43 and 44 and at T ∈ {0, 0.5, 0.7}, so its advantage is not limited to the setting in which w₁ was selected.
  • The sensitivity is not specific to PGOT. Doubling the coefficient reduces SR by 4.89 points for PGOT, against 11.77 for FKL and 15.33 for Shallow-π. The drop observed for w₁ therefore reflects a general property of this setting, in which a compact student must balance expert supervision against teacher supervision, and PGOT is the least sensitive of the three objectives (App. H).

We hope these additional experiments address your concerns about the reliability of our results. If so, we would be grateful if you would consider raising your rating, and we are happy to provide any further details.

Reviewer_1MA6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PGOT, a knowledge distillation framework tailored for discrete-action Vision-Language-Action (VLA) models, transferring control capabilities from a large teacher VLA to a lightweight student VLA. Aiming to address the limitation that standard categorical KL distillation ignores physical semantics in action spaces, the authors introduce three complementary mechanisms: (1) Geometry, which employs the Wasserstein distance to measure discrepancies between teacher and student action distributions grounded in physical space; (2) Criticality, which assigns larger training weights to critical decision states based on the magnitude of action variations between adjacent steps in expert demonstrations; and (3) Reliability, which downweights unreliable teacher supervision based on the physical distance between the teacher's prediction and the expert demonstration action. Experiments on CoNav-UAV and LIBERO-10 demonstrate competitive distillation performance.

Critical strengths

1.Incorporating physical semantics into the distillation supervision via optimal transport, state criticality, and teacher reliability is both intuitive and empirically effective.

2.The proposed framework requires neither architectural modifications to the student policy nor costly online environment interactions during distillation, making it clean to implement and computationally efficient to train.

Critical weaknesses

1.The proposed method appears strictly limited to autoregressive VLA models that discretize actions into categorical bins (e.g., 256 bins per dimension). However, mainstream state-of-the-art VLA models predominantly adopt continuous action heads trained via flow matching or diffusion, which do not output explicit categorical distributions over discrete bins. This significantly restricts the broader applicability of PGOT.

2.On LIBERO-10 (Table 1), the performance gains appear quite limited, and the baseline performance is unconvincingly low. Specifically, the 9B Teacher SFT only achieves a 49.13% success rate (with the 0.8B Student SFT already at 45.47%, and PGOT reaching 49.47%). Given that LIBERO is a relatively simple and saturated benchmark where mainstream VLA models routinely achieve success rates above 90%–96%, the current setup suggests either that this discrete single-step paradigm struggles on manipulation tasks, or that the authors should adopt a much stronger teacher VLA. Moreover, evaluating on more challenging benchmarks with higher generalization demands (such as LIBERO-Plus or RoboTwin 2.0) would be much more convincing.

Questions for authors

1.Referring to Weakness 1, how can PGOT be adapted to continuous-action VLA models (e.g., flow-matching or diffusion-based action experts), given that continuous action generation is currently the dominant paradigm for VLA models?

2.Referring to Weakness 2, why is the Teacher SFT performance on LIBERO-10 so low (49.13%) compared to standard VLA baselines? Could the authors evaluate PGOT using a stronger teacher VLA on LIBERO-10, and consider benchmarking on more challenging manipulation suites with greater generalization requirements, such as LIBERO-Plus or RoboTwin 2.0?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1MA6 (1/3)

We sincerely thank the reviewer for the positive assessment. We appreciate the recognition that incorporating physical semantics through optimal transport, state criticality, and teacher reliability is both intuitive and empirically effective, and that PGOT is clean to implement and efficient to train without architectural changes or online interaction. The two questions are very constructive and give us the opportunity to clarify the positioning and experimental design of our work.

1. Applicability to continuous-action VLAs (W1, Q1). We fully agree that VLAs with continuous action generation, such as flow-matching and diffusion action experts, have seen extensive research and remarkable progress. At the same time, discrete autoregressive (AR) VLAs remain an important line among state-of-the-art models. They have distinctive advantages that make distilling them worthwhile:

  • Competitive and actively developed. AR VLAs range from RT-2 (Brohan et al., 2023), OpenVLA (Kim et al., 2025b), and π0-FAST (Pertsch et al., 2025) to the recent ActionCodec (Dong et al., 2026) and G0.5 (Liu et al., 2026). π0-FAST matches the performance of diffusion VLAs while reducing training time by up to 5×. A purely AR SmolVLM2-2.2B with ActionCodec reaches 90.6% on LIBERO-10, and G0.5 reports 98.9% on LIBERO and 82.5% on DROID under zero-shot transfer to an unseen environment and objects. Distillation for this paradigm, however, has received little attention, and PGOT is designed specifically for it.
  • Native compatibility with LLM/VLM progress. AR VLAs keep the VLM's next-token interface. Distilled students can therefore directly adopt advances in LLM post-training, such as on-policy distillation and interleaved reasoning tokens (as in G0.5), and can scale up as VLM backbones improve.
  • Largest benefit from compression. Token-by-token decoding makes AR VLAs slower at inference than flow-based policies: about 750 ms versus 100 ms per action chunk for π0-FAST and π0 (Pertsch et al., 2025). Distilling them into compact students therefore addresses their main practical limitation.
  • Lightweight, architecture-agnostic distillation. PGOT requires no action head or architectural change. Our UAV teacher (Qwen2.5-VL-7B) and student (Qwen3.5-0.8B) even come from different model generations and are aligned only through action-token strings.

Regarding continuous-action VLAs (Q1), Geometry is designed around the explicit, ordered action distributions that AR VLAs expose. Criticality and Reliability, in contrast, only require expert actions and the teacher's predicted action. They can therefore directly serve as per-sample weights for the distillation losses of flow-matching or diffusion action experts. Extending physically grounded, selective supervision to continuous-action VLAs is a promising direction, which we now discuss in §5.

References

  • Anthony Brohan, Noah Brown, Justice Carbajal, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. In Proceedings of the 7th Conference on Robot Learning (CoRL), 2023.
  • Zibin Dong, Yicheng Liu, Shiduo Zhang, et al. ActionCodec: What makes for good action tokenizers. arXiv preprint arXiv:2602.15397, 2026.
  • Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, et al. OpenVLA: An open-source vision-language-action model. In Proceedings of the 8th Conference on Robot Learning (CoRL), 2025b.
  • Yicheng Liu, Zibin Dong, Baijun Ye, et al. G0.5: One autoregressive stream for robot reasoning and action. arXiv preprint arXiv:2608.11739, 2026.
  • Karl Pertsch, Kyle Stachowicz, Brian Ichter, et al. FAST: Efficient action tokenization for vision-language-action models. In Proceedings of Robotics: Science and Systems (RSS), 2025.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer 1MA6 (2/3)

2. Absolute success and gains on LIBERO-10 (W2, Q2). Thank you for this question, which helps us clarify our experimental design. Our LIBERO-10 success rates are lower than those of mainstream VLAs mainly because the setup is deliberately minimal:

  • A minimal setup isolates the distillation objective. We did not optimize the teacher through robot pretraining, additional data, richer inputs, alternative output formats, or hyperparameter search, because our goal is to test whether the distillation objective itself is effective.
    • We train only on the OpenVLA-processed LIBERO-10 demonstrations (379 episodes, 101,469 samples), with no other LIBERO suite or robot data.
    • The model sees one agent-view image, with no wrist camera or proprioceptive state, and outputs one action per step (no action chunking) as uniform 256-bin tokens. There is no robot-specific pretraining.
    • This setting is close to OpenVLA's single-step 256-bin interface (Kim et al., 2025b). OpenVLA, which is additionally pretrained on large-scale robot data, reaches 53.7% on LIBERO-10 (Kim et al., 2025a), so our 9B teacher's 49.13% is in the expected range.
  • AR VLAs reach much higher success when tuned further. As noted above, ActionCodec (90.6% on LIBERO-10; Dong et al., 2026) and G0.5 (98.9% on LIBERO; Liu et al., 2026) improve the input and output design and the training procedure while remaining fully autoregressive. We plan to invest more compute in these directions in future work. We added these details to §4.1 and App. F.1.
  • The comparison between distillation methods is controlled. All distillation methods use the same teacher, student, interface, and training protocol, and differ only in the distillation objective. The differences in Table 1 are therefore attributable to the objective, regardless of the absolute level of the teacher.
  • Gains relative to the available headroom. The gap between Teacher SFT and Student SFT on LIBERO-10 is only 3.66 SR points (49.13 vs. 45.47), which bounds what distillation can recover in SR. PGOT closes this gap entirely (49.47), whereas five of the six distillation baselines fall below the undistilled student. On Pass³, PGOT reaches 45.40, which is 14.00 points above Student SFT, 6.00 above the best baseline, and 8.20 above the 9B teacher (37.20). These Pass³ gains are statistically significant in paired comparisons over the 500 initial states (exact McNemar test on the counts in App. F.3): p = 0.024 against FKL and p < 0.001 against Shallow-π.

References

  • Zibin Dong, Yicheng Liu, Shiduo Zhang, et al. ActionCodec: What makes for good action tokenizers. arXiv preprint arXiv:2602.15397, 2026.
  • Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. In Proceedings of Robotics: Science and Systems (RSS), 2025a.
  • Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, et al. OpenVLA: An open-source vision-language-action model. In Proceedings of the 8th Conference on Robot Learning (CoRL), 2025b.
  • Yicheng Liu, Zibin Dong, Baijun Ye, et al. G0.5: One autoregressive stream for robot reasoning and action. arXiv preprint arXiv:2608.11739, 2026.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer 1MA6 (3/3)

2. Absolute success and gains on LIBERO-10 (continued).

  • Why the student can match or exceed the teacher. This is a well-documented phenomenon in both offline and on-policy distillation:

    • Distilled 770M T5 models outperform their 540B PaLM teacher (Hsieh et al., 2023).
    • MiniLLM students exceed their teachers on several benchmarks (Gu et al., 2024).
    • Generalized on-policy distillation can yield students that surpass their teachers in certain settings (Yang et al., 2026).

    An established explanation is that the teacher's soft distribution is a much denser supervision signal than one-hot labels. It tells the student how plausible every alternative action is, providing more information per training sample and lower-variance gradients (Hinton et al., 2015). Meanwhile, the student still learns from the expert labels through the cross-entropy term. PGOT strengthens this effect in two ways. Reliability filters out teacher guidance that departs from the expert, so the student does not inherit the teacher's errors. Criticality concentrates the dense signal on the decisions that matter most.

  • Stronger teachers and harder suites (LIBERO-Plus, RoboTwin 2.0). We agree that a stronger teacher and benchmarks with greater generalization demands would further strengthen the paper. These experiments require training a new teacher and retraining PGOT and all distillation baselines, which we could not complete within the rebuttal period. The experiment with a stronger teacher on LIBERO-10 is in progress, and we will include its results in the final version. The PGOT objective itself does not depend on the teacher's strength and applies unchanged to any teacher with binned action tokens. We list evaluation on more challenging benchmarks as future work (§5).

We hope these clarifications address your questions. Thank you again for the encouraging assessment and the helpful suggestions, which have improved both the positioning and the clarity of the paper. If our responses have resolved your concerns, we would be grateful if you would consider raising your rating, and we are happy to address any further questions.

References

  • Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. MiniLLM: Knowledge distillation of large language models. In International Conference on Learning Representations (ICLR), 2024.
  • Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
  • Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, et al. Distilling step-by-step! Outperforming larger language models with less training data and smaller model sizes. In Findings of the Association for Computational Linguistics: ACL 2023, 2023.
  • Wenkai Yang, Weijie Liu, Ruobing Xie, et al. Learning beyond teacher: Generalized on-policy distillation with reward extrapolation. arXiv preprint arXiv:2602.12125, 2026.
Reviewer_8itgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PGOT distills discrete-action vision-language-action models through three mechanisms: aligning action distributions using physically grounded optimal transport, emphasizing critical decisions, and filtering teacher guidance according to its agreement with expert demonstrations. Experiments on CoNav-UAV and LIBERO-10 show that 0.8B students achieve success rates comparable to their 7B/9B teachers, while improving success rate and three-out-of-three completion over the strongest distillation baselines. Ablation studies and diagnostic analyses further examine each component’s contribution.

Critical strengths
  1. PGOT considers action distances, decision importance, and teacher reliability when training the student model.
  2. Ablation experiments show that all three components contribute to the performance improvements.
  3. On both simulation benchmarks, the smaller student models achieve success rates comparable to their teachers and complete tasks more consistently than competing distillation methods.
Critical weaknesses
  1. Performance drops substantially when the transport weight changes from 10 to 5 or 20. Please explain how the default coefficients were selected and whether the baselines received comparable tuning effort.
  2. Large action changes do not necessarily indicate important decisions; maintaining a steady grip may be critical despite little action change. Please explain why action change is a suitable importance proxy, using the existing evidence, and discuss when this assumption may fail.
  3. Removing Geometry also removes an extra training loss, so the performance drop alone does not show why Geometry helps. Please explain how the existing results support the value of action distances and whether changes in training strength could also contribute to the improvement.
Questions for authors
  1. What evidence shows that larger changes in expert actions indicate decisions more critical to task success?
  2. How does Reliability distinguish incorrect teacher predictions from valid alternatives to the expert demonstration?
  3. Do Geometry’s gains come from meaningful action distances rather than increased supervision strength? Can a matched-strength control verify this?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 8itg (1/4)

We thank the reviewer for the careful and constructive review. We appreciate the recognition that PGOT accounts for action distances, decision importance, and teacher reliability, that the ablations show every component contributes, and that our compact students achieve teacher-level success while completing tasks more consistently than competing distillation methods. The questions on hyperparameter selection, the Criticality proxy, and the source of Geometry's gains are insightful. We ran a new experiment for each and summarize the results below.

1. Hyperparameter selection and baseline tuning (W1). For PGOT, we searched the weight of its core component, Geometry (w₁), over {5, 10, 20} using a single training seed (42) and decoding temperature (0.3). The same value, w₁ = 10, is best in both domains and is used throughout (App. H). To ensure that the comparison is not affected by unequal tuning effort, we gave the two strongest baselines the same budget under the identical training and evaluation protocol. We searched the FKL weight λ_FKL of FKL and the attention-alignment weight λ_attn of Shallow-π, each over three values (new Table 3, App. C). All methods share the training data, optimizer schedule, batch size, and number of steps, and are evaluated at their final checkpoints.

SettingSRPass³CRMMBC
FKL, λ_FKL = 1 (default)85.3374.6711.785.27
FKL, λ_FKL = 2 / 373.56 / 75.7854.67 / 58.6726.22 / 23.563.54 / 3.72
Shallow-π, λ_attn = 1 (default)85.3374.0010.895.09
Shallow-π, λ_attn = 2 / 370.00 / 73.5653.33 / 59.3329.56 / 25.333.52 / 3.81
PGOT, w₁ = 10 (default)88.0082.008.005.72
PGOT, w₁ = 5 / 2082.44 / 83.1171.33 / 75.3311.33 / 11.565.41 / 5.46
  • The comparison uses matched tuning budgets. Both baselines perform best at their default settings, which are exactly those reported in our main table. Table 1 therefore compares the best of three settings for PGOT with the best of three for each baseline, and PGOT leads on every metric shown (e.g., 88.00 vs. 85.33 SR; 82.00 vs. 74.67 Pass³).
  • The advantage holds in experiments not used to select w₁. The additional training seeds (App. J) and decoding settings (App. K) were run afterwards with w₁ = 10 unchanged. PGOT still achieves the highest student SR under seeds 43 and 44 and at T ∈ {0, 0.5, 0.7}, so its advantage is not limited to the setting in which w₁ was selected.
  • The sensitivity is not specific to PGOT. Doubling the coefficient reduces SR by 4.89 points for PGOT, against 11.77 for FKL and 15.33 for Shallow-π. The drop observed for w₁ therefore reflects a general property of this setting, in which a compact student must balance expert supervision against teacher supervision, and PGOT is the least sensitive of the three objectives (App. H).

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer 8itg (2/4)

2. Action change as an importance proxy, and when it fails (W2, Q1). Action change is computed from the demonstrations alone: it requires no annotation, no teacher, and no environment interaction, and the same formula applies to both domains. Using changes in demonstrated behavior to locate important moments also has precedent in autonomous driving, imitation learning, manipulation, and aerial navigation (Bojarski et al., 2016; Wen et al., 2021; James & Davison, 2022; Shridhar et al., 2022; Gao et al., 2026). Criticality shares this intuition as a bounded, continuous per-sample weight.

Beyond the correlational evidence in §4.4.2, we tested this directly with a perturbation study (new App. L). Criticality scores each frame by q (Eq. 4), the change between the expert action at that frame and at its neighboring frames. On 30 long validation routes, we take two states from each route:

  • a straight-flight state with low q (bottom validation quartile), and
  • a turning state with high q (top validation quartile).

At each state, we shift the UAV toward a nearby building and let the Student SFT policy continue from the shifted position. We use Student SFT because it is trained without Criticality weighting, so it measures the effect of the shift at the two types of states without bias from our method. We compare against an unperturbed control at the same state, with three matched seeds (540 episodes). The two shift sizes are set relative to the median expert step of 5.42 m: 2.71 m (half a step) is smaller than a typical expert step, and 8.13 m (1.5 steps) is larger. Entries are SR / Pass³, with the change from the control in parentheses; each column compares the two state types under the same shift:

StateControl (no shift)2.71 m shift8.13 m shift
Straight (low q)60.00 / 50.0063.33 (+3.33) / 43.33 (−6.67)64.44 (+4.44) / 50.00 (0.00)
Turning (high q)56.67 / 43.3353.33 (−3.33) / 33.33 (−10.00)52.22 (−4.44) / 36.67 (−6.67)
Extra change at turns–−6.67 / −3.33−8.89 / −6.67

In all four comparisons (two shift sizes × two metrics), the same displacement is more harmful at turning states than at straight states. A deviation of identical physical size therefore matters more for task success where expert actions change, which is the assumption underlying Criticality. This agrees with our existing analysis in §4.4.2:

  • High-weight states are harder to imitate. The top quartile by Criticality weight has 69.6% higher expected physical error than the bottom quartile.
  • Sharp maneuvers precede failures. Late sharp turns occur before 73.9% of collisions but only 35.3% of successes.
  • In manipulation, the score concentrates on gripper switches. They make up 51.25% of the top-5% states, against 2.57% overall.

When the proxy can fail. We agree with the reviewer that action change can miss critical phases in which the expert action barely changes, such as holding a steady grasp or flying straight through a narrow passage, where a small deviation can still cause failure. Even so, the experiments above show that deviations are on average more harmful at states with large action changes. In addition, Criticality never reduces supervision, because the Criticality weight h_n ≥ 1: low-change states keep their full weight, and the weight ratio between any two samples is at most 1.5. Such steady phases also span many consecutive, similar frames, so they are already well represented in the training data, whereas transition frames are rare. Identifying these phases requires task semantics or richer context than the action sequence provides, which we leave to future work. We discuss this limitation in §5.

References

  • Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016.
  • Yunpeng Gao, Chenhui Li, Zhongrui You, et al. OpenFly: A comprehensive platform for aerial vision-language navigation. In International Conference on Learning Representations (ICLR), 2026.
  • Stephen James and Andrew J. Davison. Q-attention: Enabling efficient learning for vision-based robotic manipulation. IEEE Robotics and Automation Letters, 7(2):1612–1619, 2022.
  • Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-Actor: A multi-task transformer for robotic manipulation. In Proceedings of the 6th Conference on Robot Learning (CoRL), 2022.
  • Chuan Wen, Jierui Lin, Jianing Qian, et al. Keyframe-focused visual imitation learning. In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer 8itg (3/4)

3. Physical distances vs. extra loss strength (W3, Q3). We thank the reviewer for suggesting this control. Removing Geometry also removes one loss term, so its effect could in principle come from reduced supervision strength rather than from physical distances. To separate the two, we trained a matched-strength variant (new App. I). It removes Geometry and raises the filtered-FKL weight from 1 to 1.3737, where 0.3737 is the median ratio of the transport loss to the filtered-FKL loss during PGOT training. Its total distillation loss therefore has the same magnitude as PGOT's, and the only remaining difference is whether discrepancies are measured by physical distance.

SRPass³CRMMBC (m)
PGOT88.00 ± 2.9182.008.005.72
w/o Geometry82.22 ± 2.1472.6713.565.25
w/o Geometry, matched strength85.33 ± 2.6776.009.115.25
  • Supervision strength accounts for part of the drop. Restoring the strength raises SR from 82.22 to 85.33 and Pass³ from 72.67 to 76.00, recovering 3.11 of the 5.78 SR points and 3.33 of the 9.33 Pass³ points lost in the original ablation.
  • Geometry contributes beyond supervision strength. At matched strength, PGOT still leads by 2.67 SR and 6.00 Pass³ points, and MMBC does not improve at all: it stays at 5.25 m, as in the original w/o Geometry ablation, against 5.72 m for PGOT. Stronger categorical supervision therefore cannot substitute for Geometry. FKL measures only how much probability is misplaced; the transport loss also measures how far it is from where the teacher places it, and penalizes it in proportion to that physical distance.
  • The gradient analysis supports the same conclusion (§4.4.1). If Geometry only increased supervision strength, its gradient would be parallel to that of the FKL term. The measured mean angle between the two is 72–74° in both domains, although the transport gradient has only 10–15% of the FKL gradient's magnitude. Geometry therefore mainly changes the direction of the update, which rescaling FKL cannot reproduce.
  • This direction is specific to physical distances (§4.4.1). With Geometry, student probability at least 32 bins from the expert action falls by 17.26% on CoNav-UAV and 44.93% on LIBERO-10, while probability 3–15 bins away rises by 10.5% and 12.0%. Geometry thus moves probability from distant bins to bins near the expert action. FKL treats all incorrect bins alike regardless of their distance from the expert action, so increasing its weight does not specifically reduce probability on distant bins.

(continued in the next comment)

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged that all three reviewers found the paper "clear and easy to follow", and that they recognized the following strengths:

  • The idea. Incorporating physical semantics into the distillation supervision is "both intuitive and empirically effective" (1MA6), and using decoded action distances is "a sensible improvement over treating all token errors equally" (r2sW).
  • The practicality. PGOT "requires neither architectural modifications to the student policy nor costly online environment interactions", which makes it "clean to implement and computationally efficient to train" (1MA6).
  • The empirical results. "The gains appear in two different control domains", and the improvement in repeated completion "is relevant to practical deployment" (r2sW). The students "achieve success rates comparable to their teachers and complete tasks more consistently than competing distillation methods" (8itg).
  • The analysis. "All three components contribute to the performance improvements" (8itg), and "the component analysis is unusually detailed" (r2sW).

The reviewers' questions concern the statistical robustness of the gains, the fairness of the comparison, the evidence for each component, and the scope of the method. We conducted five new experiments and added statistical tests to address them; all changes are marked in blue in the revised PDF.

New experimentConcern addressedMain findingLocation
Independent training seedsWhether gains hold across trained models (r2sW W1/Q1)PGOT has the best SR and Pass³ for every seed. Its seed-averaged Pass³ gain is +4.67 over FKL and +8.44 over Shallow-π (Welch's t-test over seeds: p = 0.025 and 0.006), and its lowest seed exceeds the highest seed of either baselineApp. J
Greedy decoding and temperature sweepBetter control vs. reduced sensitivity to decoding (r2sW W2/Q2)PGOT has the best student SR at every setting, including greedy decoding. With the best temperature chosen for each method, it leads the best baseline by 3.33 SR and 6.00 Pass³ pointsApp. K
Equal-budget hyperparameter searchLoss-weight selection and baseline tuning effort (8itg W1; r2sW W3/Q3)Under the same three-value budget, both baselines are best at their reported defaults and PGOT remains best; its advantage also holds on the seeds and temperatures not used to select w₁App. C, H
Matched-strength control for GeometryPhysical distances vs. extra loss strength (8itg W3/Q3)With the loss magnitude matched, removing Geometry still lowers SR / Pass³ by 2.67 / 6.00 points, and MMBC does not recoverApp. I
Criticality perturbation studyAction change as a proxy for decision importance (8itg W2/Q1)In all four comparisons (two shift sizes × two metrics), the same displacement is more harmful at high-action-change turns than at straight segmentsApp. L

Main revisions to the paper:

  • Positioning and setup (§2, §3.4, §4.1, App. F.1): the role of discrete autoregressive VLAs, related work on action-change keyframes, the shared training and evaluation protocol, the LIBERO-10 interface with its effect on absolute success, and a corrected description of Pass³ (1MA6, 8itg, r2sW).
  • Results and analysis (§4.2–§4.4): robustness to training seeds and decoding, why distilled students can match or exceed their teachers, the matched-strength result for Geometry, and perturbation evidence for Criticality (all reviewers).
  • Statistical tests (App. F.3, J): Pass³ significance across training seeds on CoNav-UAV, and exact McNemar tests over the 500 initial states on LIBERO-10 (1MA6, r2sW).
  • Limitations and future work (§5): continuous-action VLAs, the Criticality proxy, and Reliability vs. valid teacher alternatives (all reviewers).
  • New appendix material (App. C, I, J, K, L) with complete tables for all new experiments.

Detailed responses are provided under each review. We thank the reviewers again for helping us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.