pith. sign in

arxiv: 2606.24078 · v1 · pith:65OIAINOnew · submitted 2026-06-23 · 💻 cs.RO

MinInter: Minimizing Trajectory Interpolation During Data Augmentation for Imitation Learning

Qingyang Wang , Xingang Liu , Changwei Yao , Zikai Ouyang , Junwei Liu , Haibo Lu , Wei Zhang This is my paper

Pith reviewed 2026-06-26 00:45 UTC · model grok-4.3

classification 💻 cs.RO
keywords imitation learningdata augmentationtrajectory interpolationrobot manipulationsynthetic demonstrationsMinInterMimicGen benchmark
0
0 comments X

The pith

Selecting the demonstration needing the least interpolation for each new start state produces higher-quality synthetic trajectories for imitation learning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MinInter as a selection rule for recombining expert demonstrations into new trajectories under varied initial conditions. For any given starting configuration it picks the source demonstration that requires the shortest interpolation segment to reach a complete path. This reduces the fraction of non-expert transitions inserted during augmentation. Experiments on twelve manipulation tasks with twenty-six variants show consistent rises in both the fraction of valid generated trajectories and the success rate of the resulting policies. The gains are largest on contact-rich, long-horizon, and high-variance problems, and the method remains compatible with existing recombination pipelines.

Core claim

MinInter selects, for each sampled initial configuration, the source demonstration that minimizes the length or presence of interpolation segments needed to form a complete trajectory. The resulting synthetic demonstrations contain fewer non-expert segments and therefore train policies that achieve higher success rates on the original manipulation tasks.

What carries the argument

The MinInter selection rule that, for a given initial state, chooses the expert demonstration requiring the least interpolation to complete a full trajectory.

If this is right

  • Data generation success rates increase because fewer attempted recombinations fail due to excessive interpolation.
  • Policy success rates rise on the target tasks, with the largest improvements occurring on contact-rich, long-horizon, and high-variance variants.
  • The method integrates directly into existing trajectory-recombination frameworks without altering their interpolation engines.
  • It produces higher policy success rates than the SkillGen framework despite remaining conceptually simpler.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same least-interpolation selection idea could be tested in other sequential data-augmentation settings outside robotics.
  • The result suggests that the expert-to-non-expert ratio inside each training trajectory may matter more than simply increasing the number of distinct start states.
  • Future pipelines might combine MinInter with adaptive interpolation methods that further shorten or eliminate non-expert segments.

Load-bearing premise

That the amount of interpolation is the dominant factor controlling the quality of the synthetic trajectories rather than other unmeasured properties of the chosen demonstrations.

What would settle it

A controlled comparison in which trajectories chosen by MinInter yield no higher data-generation or policy success rates than trajectories chosen by random selection or by an alternative non-minimizing rule on the same MimicGen tasks.

Figures

Figures reproduced from arXiv: 2606.24078 by Changwei Yao, Haibo Lu, Junwei Liu, Qingyang Wang, Wei Zhang, Xingang Liu, Zikai Ouyang.

Figure 1
Figure 1. Figure 1: Data generation pipeline with MinInter. Illustration of the proposed MinInter method for selecting demonstration trajectories. For each sampled initial object configuration, trajectories are generated from all demonstrations by transforming their subtask segments and adding necessary interpolations. The total interpolation of each candidate is computed, and the trajectory with the smallest interpolation is… view at source ↗
Figure 2
Figure 2. Figure 2: Task visualization. Visualization of the 12 manipulation tasks used in our experiments, sourced from the open benchmark [2]. These tasks span a range of difficulty and behavior types, including basic tasks (a–b), contact-rich tasks (c–h), and long-horizon tasks (i–l). Environments are implemented using the MuJoCo simulator. each task is run under three initial state distributions (D0, D1, D2), with difficu… view at source ↗
Figure 3
Figure 3. Figure 3: Interpolation reduction. Average change in interpolations between MinInter and the baseline across all evaluated task variants. Interpolations refer to the transition segments used in the baseline methods and are evaluated in translation and rotation as defined in our methodology. MinInter consistently reduces interpolations in the generated trajectories. portion of successful trajectories and enabling hig… view at source ↗
Figure 4
Figure 4. Figure 4: Success rate improvement. Change in policy success rate between MinInter and baseline across all evaluated task variants. For instance, Stack D0 (100.0%→100.0%) and Threading D0 (98.0%→100.0%) show only marginal gains, which consequently lowers the overall average. Furthermore, the improvements observed in policy success rates are consistent with the reduction in interpolations ( [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 5
Figure 5. Figure 5: Policy Success rate improvement by task difficulty. Average success rate improvement of MinInter over the baseline, grouped by two difficulty-related factors. (a) Reset distribution: D0 represents the default object initialization, D1 increases variation in position and orientation, and D2 introduces broader and more randomized sampling. (b) Task Group: Basic (B) tasks involve simple manipulation, Contact-… view at source ↗
read the original abstract

Imitation learning enables robots to acquire complex manipulation skills from demonstrations, but its effectiveness is limited by the cost of collecting high-quality data. Trajectory-level data augmentation methods alleviate this challenge by recombining expert demonstrations under varied initial states. However, such methods typically insert interpolations or other non-expert transition segments between disjoint parts, and such non-expert segments could reduce the quality of the generated data. This paper introduces Minimizing Interpolation (MinInter), an effective trajectory selection method that, for each sampled initial configuration, chooses the source demonstration requiring the least interpolation to form a complete trajectory. By explicitly minimizing interpolations during data generation, MinInter produces higher-quality synthetic demonstrations while remaining compatible with existing data generation frameworks. Experiments on 12 manipulation tasks with 26 variants from the MimicGen benchmark show that MinInter consistently improves both data generation success rates and policy success rates, with the largest gains on contact-rich, long-horizon and high-variance settings. Compared to the recent SkillGen framework, MinInter achieves higher policy success rates despite its conceptual simplicity, underscoring the value of interpolation minimization for data augmentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces MinInter, a simple trajectory selection heuristic for trajectory-level data augmentation in imitation learning. For each sampled initial state, it selects the expert demonstration that requires the shortest interpolation segment to produce a complete trajectory, with the goal of reducing non-expert transitions. The method is shown to be compatible with existing augmentation pipelines such as SkillGen. Experiments on 12 manipulation tasks (26 variants) from the MimicGen benchmark report consistent gains in both data-generation success and downstream policy success rates, with larger improvements on contact-rich, long-horizon, and high-variance settings.

Significance. If the central empirical claim holds after proper controls, the work supplies a lightweight, framework-agnostic improvement to data augmentation that directly targets a plausible source of quality degradation (interpolation segments). The evaluation on a public benchmark with a sizable number of task variants and the explicit comparison to SkillGen are positive features; the approach requires no new parameters or learned components.

major comments (2)
  1. [Experiments] Experiments section: the claim that MinInter's gains arise specifically from minimizing interpolation length is not isolated from other demonstration properties. No ablation is described that holds trajectory length, state-space overlap with the target initial condition, or expert-segment reliability fixed while varying only the interpolation criterion; without this, the reported improvements could be driven by correlated selection effects rather than interpolation minimization itself.
  2. [Abstract / Experiments] Abstract and Experiments section: success-rate improvements are described only qualitatively (“consistently improves,” “largest gains”) with no numerical deltas, confidence intervals, or statistical significance tests provided for either data-generation or policy success rates across the 26 variants.
minor comments (1)
  1. [Method] Method section: the precise definition and measurement of “interpolation length” (e.g., Euclidean distance, time steps, or a task-specific metric) should be stated explicitly with an equation or pseudocode to allow exact reproduction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful review and for highlighting areas where the presentation and analysis can be strengthened. We address each major comment below and will incorporate revisions to improve clarity and rigor.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the claim that MinInter's gains arise specifically from minimizing interpolation length is not isolated from other demonstration properties. No ablation is described that holds trajectory length, state-space overlap with the target initial condition, or expert-segment reliability fixed while varying only the interpolation criterion; without this, the reported improvements could be driven by correlated selection effects rather than interpolation minimization itself.

    Authors: We agree that an explicit ablation isolating interpolation length from correlated factors such as trajectory length or state overlap would provide stronger evidence for the mechanism. The current experiments demonstrate consistent gains across 26 task variants when using the interpolation-minimization criterion, but do not include controls that vary only that criterion. In the revision we will add a targeted ablation comparing MinInter against length-based and overlap-based selection heuristics on a subset of tasks, while reporting the resulting interpolation lengths to help separate the effects. revision: yes

  2. Referee: [Abstract / Experiments] Abstract and Experiments section: success-rate improvements are described only qualitatively (“consistently improves,” “largest gains”) with no numerical deltas, confidence intervals, or statistical significance tests provided for either data-generation or policy success rates across the 26 variants.

    Authors: We accept that the current text relies on qualitative descriptors. The revision will replace these with concrete numerical results (mean success-rate deltas and standard deviations across the 26 variants for both data-generation and policy success), and will include 95% confidence intervals or equivalent for the main comparisons against SkillGen and the baseline. revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical selection rule validated on external benchmark

full rationale

The paper introduces a heuristic for choosing source demonstrations that minimize interpolation length during trajectory recombination, then reports success-rate improvements on the MimicGen benchmark. No equations, fitted parameters, or derivation chain exist. The central claim is an empirical comparison (MinInter vs. baselines on 12 tasks), not a reduction of any quantity to itself by definition or self-citation. External benchmark results are independent of any internal construction, satisfying the self-contained criterion.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no equations, fitted constants, or new postulated entities; the ledger is therefore empty.

pith-pipeline@v0.9.1-grok · 5743 in / 1042 out tokens · 18730 ms · 2026-06-26T00:45:14.289761+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 9 canonical work pages · 2 internal anchors

  1. [1]

    Cyberdemo: Augmenting simulated human demonstration for real-world dexterous manipulation,

    J. Wang, Y . Qin, K. Kuang, Y . Korkmaz, A. Gurumoorthy, H. Su, and X. Wang, “Cyberdemo: Augmenting simulated human demonstration for real-world dexterous manipulation,” in2024 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 2024, pp. 17 952–17 963

    2024

  2. [2]

    Mimicgen: A data generation system for scalable robot learning using human demonstrations,

    A. Mandlekar, S. Nasiriany, B. Wen, I. Akinola, Y . Narang, L. Fan, Y . Zhu, and D. Fox, “Mimicgen: A data generation system for scalable robot learning using human demonstrations,” inConference on Robot Learning. PMLR, 2023, pp. 1820–1864

    2023

  3. [3]

    Skillgen: Auto- mated demonstration generation for efficient skill learning and deploy- ment,

    C. R. Garrett, A. Mandlekar, B. Wen, and D. Fox, “Skillgen: Auto- mated demonstration generation for efficient skill learning and deploy- ment,” in2nd CoRL Workshop on Learning Effective Abstractions for Planning

  4. [4]

    Dexmimicgen: Automated data generation for biman- ual dexterous manipulation via imitation learning,

    Z. Jiang, Y . Xie, K. Lin, Z. Xu, W. Wan, A. Mandlekar, L. Fan, and Y . Zhu, “Dexmimicgen: Automated data generation for biman- ual dexterous manipulation via imitation learning,”arXiv preprint arXiv:2410.24185, 2024

    work page arXiv 2024

  5. [5]

    Data augmentation for manipulation,

    P. Mitrano and D. Berenson, “Data augmentation for manipulation,” Robotics: Science and Systems 2022, 2022

    2022

  6. [6]

    A comprehensive survey on data augmentation,

    Z. Wang, P. Wang, K. Liu, P. Wang, Y . Fu, C.-T. Lu, C. C. Aggarwal, J. Pei, and Y . Zhou, “A comprehensive survey on data augmentation,” CoRR, 2024

    2024

  7. [7]

    A framework for efficient robotic manipulation,

    A. Zhan, R. Zhao, L. Pinto, P. Abbeel, and M. Laskin, “A framework for efficient robotic manipulation,” inDeep RL Workshop NeurIPS 2021, 2021

    2021

  8. [8]

    Reinforcement learning with augmented data,

    M. Laskin, K. Lee, A. Stooke, L. Pinto, P. Abbeel, and A. Srinivas, “Reinforcement learning with augmented data,”Advances in neural information processing systems, vol. 33, pp. 19 884–19 895, 2020

    2020

  9. [9]

    Image augmentation is all you need: Regularizing deep reinforcement learning from pixels,

    I. Kostrikov, D. Yarats, and R. Fergus, “Image augmentation is all you need: Regularizing deep reinforcement learning from pixels,”arXiv preprint arXiv:2004.13649, 2020

    work page arXiv 2004

  10. [10]

    Visual imitation made easy,

    S. Young, D. Gandhi, S. Tulsiani, A. Gupta, P. Abbeel, and L. Pinto, “Visual imitation made easy,” inConference on Robot learning. PMLR, 2021, pp. 1992–2005

    2021

  11. [11]

    Retinagan: An object-aware approach to sim-to-real transfer,

    D. Ho, K. Rao, Z. Xu, E. Jang, M. Khansari, and Y . Bai, “Retinagan: An object-aware approach to sim-to-real transfer,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 10 920–10 926

    2021

  12. [12]

    S4rl: Surprisingly simple self-supervision for offline reinforcement learning in robotics,

    S. Sinha, A. Mandlekar, and A. Garg, “S4rl: Surprisingly simple self-supervision for offline reinforcement learning in robotics,” in Conference on Robot Learning. PMLR, 2022, pp. 907–917

    2022

  13. [13]

    Expert data augmentation in imitation learning (student abstract),

    F. Han and Z. Zhang, “Expert data augmentation in imitation learning (student abstract),” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 13, 2023, pp. 16 220–16 221

    2023

  14. [14]

    Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking,

    H. Bharadhwaj, J. Vakil, M. Sharma, A. Gupta, S. Tulsiani, and V . Ku- mar, “Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking,” inFirst Workshop on Out-of-Distribution Generalization in Robotics at CoRL 2023

    2023

  15. [15]

    Physically- based lighting augmentation for robotic manipulation,

    S. Jin, L. Wang, B. Temming, and F. T. Pokorny, “Physically- based lighting augmentation for robotic manipulation,”arXiv preprint arXiv:2508.01442, 2025

    work page arXiv 2025

  16. [16]

    Diffusion meets dagger: Supercharging eye-in-hand imitation learning,

    X. Zhang, M. Chang, P. Kumar, and S. Gupta, “Diffusion meets dagger: Supercharging eye-in-hand imitation learning,” inRobotics science and systems. Robotics science and systems, 2024

    2024

  17. [17]

    Zero-shot robotic manipulation with pre-trained image- editing diffusion models,

    K. Black, M. Nakamoto, P. Atreya, H. Walke, C. Finn, A. Kumar, and S. Levine, “Zero-shot robotic manipulation with pre-trained image- editing diffusion models,” inNeurIPS 2023 Workshop on Goal- Conditioned Reinforcement Learning

    2023

  18. [18]

    Cacti: A framework for scalable multi-task multi-scene visual imitation learning,

    Z. Mandi, H. Bharadhwaj, V . Moens, S. Song, A. Rajeswaran, and V . Kumar, “Cacti: A framework for scalable multi-task multi-scene visual imitation learning,” inCoRL 2022 Workshop on Pre-training Robot Learning

    2022

  19. [19]

    Learning robust real-world dexterous grasping policies via implicit shape augmentation,

    Q. Chen, K. Van Wyk, Y .-W. Chao, W. Yang, A. Mousavian, A. Gupta, and D. Fox, “Learning robust real-world dexterous grasping policies via implicit shape augmentation,” inConference on Robot Learning. PMLR, 2023, pp. 1222–1232

    2023

  20. [20]

    Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning,

    L. Y . Chen, C. Xu, K. Dharmarajan, R. Cheng, K. Keutzer, M. Tomizuka, Q. Vuong, and K. Goldberg, “Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning,” in8th Annual Conference on Robot Learning

  21. [21]

    Imitating task and motion planning with visuomotor transformers,

    M. Dalal, A. Mandlekar, C. R. Garrett, A. Handa, R. Salakhutdinov, and D. Fox, “Imitating task and motion planning with visuomotor transformers,” inConference on Robot Learning. PMLR, 2023, pp. 2565–2593

    2023

  22. [22]

    Nod-tamp: Multi-step manipulation planning with neural object descriptors,

    S. Cheng, C. R. Garrett, A. Mandlekar, and D. Xu, “Nod-tamp: Multi-step manipulation planning with neural object descriptors,” in CoRL 2023 Workshop on Learning Effective Abstractions for Planning (LEAP)

    2023

  23. [23]

    Hybridgen: Vlm-guided hybrid planning for scalable data generation of imitation learning,

    W. Wang and N. Tan, “Hybridgen: Vlm-guided hybrid planning for scalable data generation of imitation learning,”arXiv preprint arXiv:2503.13171, 2025

    work page arXiv 2025

  24. [24]

    Physics-driven data generation for contact-rich manipulation via trajectory optimization,

    L. Yang, H. Suh, T. Zhao, B. P. Graesdal, T. Kelestemur, J. Wang, T. Pang, and R. Tedrake, “Physics-driven data generation for contact-rich manipulation via trajectory optimization,”arXiv preprint arXiv:2502.20382, 2025

    work page arXiv 2025

  25. [25]

    DreamGen: Unlocking Generalization in Robot Learning through Video World Models

    J. Jang, S. Ye, Z. Lin, J. Xiang, J. Bjorck, Y . Fang, F. Hu, S. Huang, K. Kundalia, Y .-C. Linet al., “Dreamgen: Unlocking generaliza- tion in robot learning through video world models,”arXiv preprint arXiv:2505.12705, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    Real2render2real: Scaling robot data without dynamics simulation or robot hardware,

    J. Yu, L. Fu, H. Huang, K. El-Refai, R. A. Ambrus, R. Cheng, M. Z. Irshad, and K. Goldberg, “Real2render2real: Scaling robot data without dynamics simulation or robot hardware,”arXiv preprint arXiv:2505.09601, 2025

    work page arXiv 2025

  27. [27]

    Robogen: Towards unleashing infinite data for automated robot learning via generative simulation,

    Y . Wang, Z. Xian, F. Chen, T.-H. Wang, Y . Wang, K. Fragkiadaki, Z. Erickson, D. Held, and C. Gan, “Robogen: Towards unleashing infinite data for automated robot learning via generative simulation,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 51 936–51 983

    2024

  28. [28]

    Counterfactual data augmentation using locally factored dynamics,

    S. Pitis, E. Creager, and A. Garg, “Counterfactual data augmentation using locally factored dynamics,”Advances in Neural Information Processing Systems, vol. 33, pp. 3976–3990, 2020

    2020

  29. [29]

    Mocoda: Model- based counterfactual data augmentation,

    S. Pitis, E. Creager, A. Mandlekar, and A. Garg, “Mocoda: Model- based counterfactual data augmentation,”Advances in Neural Infor- mation Processing Systems, vol. 35, pp. 18 143–18 156, 2022

    2022

  30. [30]

    Rocoda: Coun- terfactual data augmentation for data-efficient robot learning from demonstrations,

    E. Ameperosa, J. A. Collins, M. Jain, and A. Garg, “Rocoda: Coun- terfactual data augmentation for data-efficient robot learning from demonstrations,”arXiv preprint arXiv:2411.16959, 2024

    work page arXiv 2024

  31. [31]

    Offline imitation learning through graph search and retrieval,

    Z. Yin and P. Abbeel, “Offline imitation learning through graph search and retrieval,” inRobotics: Science and Systems (RSS), 2024

    2024

  32. [32]

    Skillmimic-v2: Learning robust and generalizable interaction skills from sparse and noisy demonstrations,

    R. Yu, Y . Wang, Q. Zhao, H. W. Tsui, J. Wang, P. Tan, and Q. Chen, “Skillmimic-v2: Learning robust and generalizable interaction skills from sparse and noisy demonstrations,” inProceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, 2025, pp. 1–11

    2025

  33. [33]

    Model-based trajectory stitching for improved behavioural cloning and its applications,

    C. A. Hepburn and G. Montana, “Model-based trajectory stitching for improved behavioural cloning and its applications,”Machine Learning, vol. 113, pp. 647–674, 2023

    2023

  34. [34]

    Miles: Making imitation learning easy with self-supervision,

    G. Papagiannis and E. Johns, “Miles: Making imitation learning easy with self-supervision,” in8th Annual Conference on Robot Learning

  35. [35]

    In- tervengen: Interventional data generation for robust and data-efficient robot imitation learning,

    R. Hoque, A. Mandlekar, C. Garrett, K. Goldberg, and D. Fox, “In- tervengen: Interventional data generation for robust and data-efficient robot imitation learning,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 2840– 2846

    2024

  36. [36]

    robosuite: A Modular Simulation Framework and Benchmark for Robot Learning

    Y . Zhu, J. Wong, A. Mandlekar, R. Martín-Martín, A. Joshi, S. Nasiri- any, and Y . Zhu, “robosuite: A modular simulation framework and benchmark for robot learning,”arXiv preprint arXiv:2009.12293, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2009