Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

This paper claims that world knowledge in LLMs can replace human annotation in demonstration augmentation, turning as few as one human demonstration into a large robot dataset for imitation learning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 15:12 UTC pith:TWJ7TOEJ

load-bearing objection Solid incremental contribution: LLM annotation plus a bandit optimizer over reusable annotations is new and useful, but the 'consistently outperforms' claim is too strong and the perception dependence is under-examined. the 5 major comments →

arxiv 2509.20070 v2 pith:TWJ7TOEJ submitted 2025-09-24 cs.RO

LLM Trainer: Automated Robotic Data Generation via Demonstration Augmentation using LLMs

Abraham George , Amir Barati Farimani This is my paper
classification cs.RO
keywords LLM Trainerdemonstration augmentationimitation learningkeypose retargetingThompson samplingrobot manipulationdata generationLLM world knowledge
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces LLM Trainer, a pipeline that takes one (or a few) human demonstrations and a one-sentence task description, and uses an LLM to (1) annotate the demonstration with keyframes, salient objects, and pose-object relationships, and (2) retarget those keyposes to a new scene's initial observation. The retargeted keyposes warp the original trajectory to generate new demonstrations, which are executed and saved when successful. Because the annotation is independent of the new scene, it can be treated as an 'arm' in a Thompson-sampling bandit; optimizing over annotations improves generation success 2-3 times and lets the method exceed expert-annotated baselines on most evaluated tasks. This matters because imitation learning is data-hungry and manual demonstration annotation is a bottleneck; if LLM Trainer works as claimed, robot datasets can be scaled from a single human demonstration without additional human effort.

Core claim

The central claim is that the world knowledge encoded in LLMs is sufficient to replace human expert annotation in demonstration augmentation. The system decomposes data generation into an offline annotation function F that, given a demonstration and task description, outputs keyframes, salient objects, and modification instructions, and an online retargeting function G that, given that annotation plus an initial observation of a new scene, outputs adjusted keyposes. These keyposes define a warp of the original trajectory; executing the warped trajectory yields a new demonstration, saved only on success. The paper reports that this fully automated pipeline outperforms expert-annotated baselin

What carries the argument

The load-bearing mechanism is the two-stage LLM annotation-retargeting decomposition. F converts a demonstration into a processed annotation (keyposes, objects, retargeting instructions) once; G reuses that annotation to map an initial observation of a new scene to a new set of keyposes, which then drive a rigid keypose-based trajectory warp with rotation interpolation. Because F is reusable, annotations become arms in a Thompson-sampling multi-armed bandit; the bandit balances exploring new annotations with exploiting known-good ones, and the best annotation doubles or triples generation success. The system also introduces a magnitude-aware cosine similarity to ensemble the LLM feed-forward

Load-bearing premise

The pipeline assumes it can obtain accurate object poses from the initial observation of each new scene, using automated segmentation and pose registration; if those poses are wrong, the retargeted keyposes and warped trajectories are wrong regardless of the LLM's plan.

What would settle it

Run the data-generation pipeline twice on the same set of new scenes, once with ground-truth object poses and once with the paper's estimated poses from RGB-D; if generation success rates are close, pose perception is adequate, but if estimated poses cause a large drop, the key assumption is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single unannotated human demonstration plus a one-sentence task description can generate hundreds of successful demonstrations across varied scenes.
  • Optimizing annotations with Thompson sampling improves generation success by 2-3x over unoptimized LLM annotations and beats hand-crafted expert baselines on most tasks.
  • Imitation learning agents trained on LLM-generated data reach success rates on par with or better than agents trained on expert-annotated data at the same dataset size.
  • Ensembling the LLM feed-forward plan with a learned feedback controller yields higher success in low-data regimes and better robustness to dynamic scenes than either alone.
  • The pipeline transfers to physical robots, achieving 75.8% generation success on a long-horizon mug-cleanup task without task-specific engineering.

Where Pith is reading between the lines

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

  • The method's real bottleneck is likely perception: if object pose extraction from RGB-D fails (occlusion, novel objects, poor segmentation or registration), retargeted trajectories will be wrong even when the LLM's plan is correct.
  • The paper itself flags reliance on physical rollouts (time-consuming, may need supervision) and ignoring LLM/prompt cost in its optimization objective as limitations, setting the practical operating envelope.
  • Because the offline annotation is reusable across scenes, the same annotated demonstration could in principle be shared across different robot embodiments or tasks with similar geometric structure, further amortizing the LLM cost.
  • A natural testable extension is to apply the pipeline to non-rigid or articulated objects, where the rigid keypose warp and pose-object relations would need to be generalized; the current method assumes rigid object poses.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper presents LLM Trainer, a pipeline for automated robot demonstration generation. An LLM first annotates a human demonstration by selecting keyframes, listing relevant objects, and describing pose-object relations; then, given an initial observation of a new scene, a separate LLM-based retargeting step adapts those keypoints to the new scene. The original trajectory is warped to the new keypoints, executed, and saved if successful. The annotation is treated as a reusable arm in a Thompson-sampling multi-armed bandit, which the authors use to optimize annotation quality. The method is evaluated on RoboMimic/MimicGen and PandaGym/OneACTPlay simulation tasks, comparing generation success rates and downstream imitation-learning performance with the MimicGen and OneACTPlay baselines. The paper also proposes an ensemble controller that switches between the LLM feedforward policy and a learned IL feedback policy, and reports a hardware experiment on a Franka Emika Panda for a mug-cleanup task.

Significance. If the reported results hold, LLM Trainer would be a meaningful step toward reducing the human effort in data generation for imitation learning, replacing expert trajectory annotation with LLM-based keyframe extraction and retargeting while retaining the benefits of demo-augmentation approaches. The paper has several strengths: it explicitly separates no-optimization, best-annotation, and total generation success rates; it retrains MimicGen baselines rather than relying solely on published numbers; it evaluates the generated data by training downstream IL agents; and it includes an end-to-end hardware demonstration. The overall concept is timely and the optimization formulation is interesting. However, several load-bearing details are missing or insufficiently validated, including the identity and prompts of the LLM, the reliability of the perception pipeline in new scenes, the statistical reliability of the generation success rates, and the sensitivity of the bandit and ensemble heuristics.

major comments (5)
  1. [Section IV, Table II] The central data-generation comparison is presented as single point estimates with no variance, no number of independent data-collection runs, and no trial counts for the No RL, Best Annotation, and Total columns. The claim that the method "significantly outperforms" expert baselines is therefore not statistically supported. Moreover, the Stack task Total success rate is 79.1%, below the baseline's 80.0%; the text acknowledges this, but the abstract and conclusion phrase the outperformance without this qualification. In addition, "Best Annotation" is selected after all rollouts are observed, making it an oracle upper bound rather than an online achievable rate. The headline claim should be based on the Total (online) success rate, or the paper should explicitly state when the claim refers to the oracle-selected annotation.
  2. [Section III-A, III-F] The "fully automated" claim depends on the assumption that an initial observation of each new scene includes correct object poses. In simulation these poses come directly from the simulator, so the LLM and warping components are evaluated under perfect perception. On hardware, object poses are estimated via LLM descriptions, SAM/Grounding DINO segmentation, and RANSAC/ICP registration, but the paper reports only end-to-end generation success (75.8%) on one mug-cleanup task, with no pose accuracy, no failure analysis, and no robustness evaluation under partial occlusion, novel objects, or lighting changes. Since the retargeting function G (Section III-B) maps poses to keypoints, incorrect poses will produce incorrect warped trajectories even if the LLM's planning is correct. The Limitations section does not list perception robustness as a limitation. Please either add a systematic evaluat
  3. [Section III-B] The manuscript never specifies which LLM, model version, or prompt templates are used. All quantitative results depend on the LLM's ability to extract keyframes, follow formatting instructions, and reason about object relations; without this information the experiments cannot be reproduced or compared across LLM choices. This is a central reproducibility issue for a paper whose main contribution is an LLM-based pipeline. Please provide the model identifier(s), sampling parameters, and the full prompts used for functions F and G, or otherwise make them available in supplementary material.
  4. [Section III-D, Eq. (6)] The bandit optimization is central to the claimed 2-3x improvement, but the new-arm prior estimation is not rigorously validated. Equation (6) fits a Beta prior by maximizing a product of expectations of the Beta density under each arm's posterior; this is not a standard marginal-likelihood or moment-matching objective, and the paper provides no evidence that the fitted prior matches the observed distribution of new-arm success rates. The horizon T is approximated as (n_suc_goal - n_suc_cur)/max(P), which depends on max(P), the quantity being estimated. The approximations k=1000 and the number of samples m are given without sensitivity analysis. Since the optimization method is a contribution, please justify the objective, validate the fit, and include sensitivity studies for k, m, and T.
  5. [Section III-E, Eqs. (7)-(8)] The ensemble method is one of the three stated contributions, but the switching heuristic is not validated or fully reproducible. The similarity threshold tau is never reported, the cooldown value is fixed at 5 timesteps without analysis, and no ablation compares the proposed magnitude-aware cosine similarity and reattachment rule against simpler alternatives. The ensemble results in Tables III and IV are interesting, but without these details it is unclear how sensitive the performance is to the heuristic's free parameters. Please report tau, provide an ablation over tau, and preferably compare against a single fixed-switch or no-reattachment baseline.
minor comments (6)
  1. [Section III-D] Typo: "Thomposon" should be "Thompson".
  2. [Section IV] The text refers to "RobotMimic" in one place; this should be "RoboMimic" for consistency with the references.
  3. [Section IV] The paper says it uses the same human demonstrations as in [6] and [5], but does not clarify how many demonstrations are used for each environment beyond the statement that RoboMimic tasks use 10 and OneACTPlay tasks use 1. Please make this explicit in the main text near Table II.
  4. [Section IV, Table III/IV] Tables III and IV are extremely dense and hard to read. Consider splitting them into separate tables per environment or using clearer column/row grouping.
  5. [Section IV-A] Typo: "feedward" should be "feedforward".
  6. [Section IV] The "N/A" entry for Stack Flipped baseline is not explained; please state whether the baseline was run and failed or was not attempted.

Circularity Check

0 steps flagged

No significant circularity: the central claims are evaluated by measured rollouts against external baselines, and the self-cited components are specified in the paper rather than reduced to the citations.

full rationale

I walked the derivation chain: the annotation functions F and G are implemented via LLMs and are not defined in terms of the predicted outcome; the keypoint warping in Eqs. (2)-(5) is stated explicitly in the paper, so reusing the 'same warping mechanism proposed in [5]' is not a load-bearing reduction to a self-citation. The Thompson-sampling optimization is driven by binary rollout rewards, and the beta-distribution fit in Eq. (6) is an internal exploration/estimation heuristic, not a fitted parameter subsequently reported as a prediction. All headline success rates—simulation and hardware (75.8% generation success)—are measured from actual executions, and the main comparisons are against external MimicGen baselines [6] and OneACTPlay [5], with the MimicGen agents retrained and re-evaluated rather than taken on faith. The assumption of reliable object poses in new scenes (Section III-A and III-F) is a real correctness risk, but it is an input condition, not a circular derivation of the results. No step in the paper's own equations or argument reduces a claimed prediction to its fitted inputs or to an unverified self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The central claim rests on the prior warping framework, the perception assumption, and LLM capability. No new physical entities are postulated. The Thompson sampling and ensembling components add several heuristic hyperparameters.

free parameters (5)
  • Thompson sampling horizon estimate T
    T is approximated as (n_suc_goal - n_suc_cur)/max(P), using the current maximum success probability; affects the decision to add new annotations.
  • Virtual rollout sample count k = 1000
    Used in brute-force approximation of expected Thompson sampling value; larger k reduces variance but is arbitrary.
  • Beta fit sample count m
    Number of success-rate samples drawn from each arm to fit the new-arm beta prior; value not reported.
  • Ensemble similarity threshold tau
    Threshold for policy disagreement in Eq. 8; no value given, affects switching behavior.
  • Timestep sampling noise
    Small random noise added to the 5-timestep sampling to avoid LLM anchoring; magnitude unspecified.
axioms (7)
  • domain assumption The trajectory warping transform from [5] (Eqs. 2-5) produces valid robot trajectories when applied to keypoints.
    Borrowed from prior work; no re-derivation, central to data generation.
  • domain assumption Observations contain the poses of key objects (Section III-A).
    Required for LLM retargeting; in simulation given, on hardware estimated via perception.
  • domain assumption LLM world knowledge suffices to identify keyframes and write retargeting instructions from a text and image summary (Section III-B).
    The core capability claim; if LLMs fail, the pipeline fails.
  • standard math Arm success rates follow a Bernoulli process with Beta prior (Section III-D).
    Standard Thompson sampling model; assumed.
  • ad hoc to paper The brute-force simulation of Thompson sampling with k=1000 samples accurately estimates expected future reward (Section III-D).
    Approximation used to decide whether to add a new annotation; no error analysis.
  • ad hoc to paper Policy disagreement measured by Eq. 7 indicates feedforward error, and reattachment via Eq. 8 is safe (Section III-E).
    Heuristic ensembling rule; no theoretical justification or sensitivity analysis.
  • domain assumption RANSAC/ICP on SAM/Grounding DINO point clouds yields consistent object poses (Section III-F).
    Hardware perception; not validated for robustness.

pith-pipeline@v1.3.0-alltime-deepseek · 13162 in / 12875 out tokens · 84890 ms · 2026-08-04T15:12:56.744264+00:00 · methodology

0 comments
read the original abstract

We present LLM Trainer, a fully automated pipeline that leverages the world knowledge of Large Language Models (LLMs) to transform a small number of human demonstrations (as few as one) into a large robot dataset for imitation learning. Our approach decomposes demonstration generation into two steps: (1) offline demonstration annotation that extracts keyframes, salient objects, and pose-object relations; and (2) online keypose retargeting that adapts those keyframes to a new scene, given an initial observation. Using these modified keypoints, our system warps the original demonstration to generate a new trajectory, which is then executed, and the resulting demo, if successful, is saved. Because the annotation is reusable across scenes, we use Thompson sampling to optimize the annotation, significantly improving generation success rate. We evaluate our method on a range of tasks, and find that our data annotation method consistently outperforms expert-engineered baselines. We further show an ensemble policy that combines the optimized LLM feed-forward plan with a learned feedback imitation learning controller. Finally, we demonstrate hardware feasibility on a Franka Emika Panda robot. For additional materials and demonstration videos, please see the project website: https://sites.google.com/andrew.cmu.edu/llm-trainer

Figures

Figures reproduced from arXiv: 2509.20070 by Abraham George, Amir Barati Farimani.

Figure 1
Figure 1. Figure 1: Diagram of our approach. First, the LLM annotates [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diagram of the LLM-based demo augmentation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Outline of the ensembling method. The agent be [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Simulation evaluation tasks. Top: RoboMimic tasks [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Mug Clean-up task used for the hardware experi [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. In-situ process monitoring for defect detection in wire-arc additive manufacturing: an agentic AI approach

    cs.AI 2026-04 unverdicted novelty 6.0

    A multi-agent AI framework using processing and acoustic agents achieves 91.6% accuracy and 0.821 F1 score for in-situ porosity defect detection in wire-arc additive manufacturing.

  2. A Survey on Evaluating Quality and Trustworthiness in LLM-Generated Data

    cs.AI 2026-01 conditional novelty 5.0

    A metric-oriented survey that classifies intrinsic quality and trustworthiness metrics for LLM-generated data across six modalities and documents systematic evaluation gaps in the current literature.

Reference graph

Works this paper leans on

28 extracted references · 16 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances,

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausmanet al., “Do As I Can, Not As I Say: Grounding Language in Robotic Affordances,”arXiv preprint arXiv:2204.01691, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    Plato: Planning with llms and affordances for tool manipulation,

    A. Car, S. S. Yarlagadda, A. Bartsch, A. George, and A. B. Farimani, “Plato: Planning with llms and affordances for tool manipulation,” arXiv preprint arXiv:2409.11580, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Llm-craft: Robotic crafting of elasto-plastic objects with large language models,

    A. Bartsch and A. B. Farimani, “Llm-craft: Robotic crafting of elasto-plastic objects with large language models,”arXiv preprint arXiv:2406.08648, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Minimizing human assistance: Augmenting a single demonstration for deep reinforcement learning,

    A. George, A. Bartsch, and A. B. Farimani, “Minimizing human assistance: Augmenting a single demonstration for deep reinforcement learning,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 5027–5033

    2023

  5. [5]

    One act play: Single demonstration behavior cloning with action chunking transformers,

    A. George and A. B. Farimani, “One act play: Single demonstration behavior cloning with action chunking transformers,”arXiv preprint arXiv:2309.10175, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    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,”arXiv preprint arXiv:2310.17596, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Efficient exploration for dialog policy learning with deep bbq networks\& replay buffer spiking,

    Z. C. Lipton, J. Gao, L. Li, X. Li, F. Ahmed, and L. Deng, “Efficient exploration for dialog policy learning with deep bbq networks\& replay buffer spiking,”CoRR abs/1608.05081, 2016

    Pith/arXiv arXiv 2016

  8. [8]

    Curriculum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” inProceedings of the 26th Annual International Conference on Machine Learning, ser. ICML ’09. New York, NY , USA: Association for Computing Machinery, 2009, p. 41–48. [Online]. Available: https://doi.org/10.1145/1553374.1553380

    arXiv 2009

  9. [9]

    What matters in learning from offline human demonstrations for robot manipula- tion,

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Mart´ın-Mart´ın, “What matters in learning from offline human demonstrations for robot manipula- tion,” inarXiv preprint arXiv:2108.03298, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 07 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning,

    Z. Jiang, Y . Xie, K. Lin, Z. Xu, W. Wan, A. Mandlekar, L. J. Fan, and Y . Zhu, “Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 16 923–16 930

    2025

  12. [12]

    Skillmimicgen: Automated demonstration generation for efficient skill learning and deployment,

    C. Garrett, A. Mandlekar, B. Wen, and D. Fox, “Skillmimicgen: Automated demonstration generation for efficient skill learning and deployment,”arXiv preprint arXiv:2410.18907, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Large language models for multi- robot systems: A survey,

    P. Li, Z. An, S. Abrar, and L. Zhou, “Large language models for multi- robot systems: A survey,”arXiv preprint arXiv:2502.03814, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    A survey on integration of large language models with intelligent robots,

    Y . Kim, D. Kim, J. Choi, J. Park, N. Oh, and D. Park, “A survey on integration of large language models with intelligent robots,” Intelligent Service Robotics, vol. 17, no. 5, pp. 1091–1107, 2024

    2024

  15. [15]

    Large language models for robotics: Opportunities, challenges, and perspectives,

    J. Wang, E. Shi, H. Hu, C. Ma, Y . Liu, X. Wang, Y . Yao, X. Liu, B. Ge, and S. Zhang, “Large language models for robotics: Opportunities, challenges, and perspectives,”Journal of Automation and Intelligence, vol. 4, no. 1, pp. 52–64, 2025

    2025

  16. [16]

    Code as policies: Language model programs for embodied control,

    J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,”arXiv preprint arXiv:2209.07753, 2022

    Pith/arXiv arXiv 2022

  17. [17]

    Llm- drone: aerial additive manufacturing with drones planned using large language models,

    C. Merrill, A. Raman, A. George, and A. Barati Farimani, “Llm- drone: aerial additive manufacturing with drones planned using large language models,”Construction Robotics, vol. 9, no. 2, pp. 1–14, 2025

    2025

  18. [18]

    Semantic intelligence: Integrating gpt-4 with a planning in low-cost robotics,

    J. Barkley, A. George, and A. B. Farimani, “Semantic intelligence: Integrating gpt-4 with a planning in low-cost robotics,”arXiv preprint arXiv:2505.01931, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Gen2sim: Scaling up robot learning in simulation with generative models,

    P. Katara, Z. Xian, and K. Fragkiadaki, “Gen2sim: Scaling up robot learning in simulation with generative models,” in2024 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 6672–6679

    2024

  20. [20]

    Gensim: Generating robotic simulation tasks via large language models,

    L. Wang, Y . Ling, Z. Yuan, M. Shridhar, C. Bao, Y . Qin, B. Wang, H. Xu, and X. Wang, “Gensim: Generating robotic simulation tasks via large language models,”arXiv preprint arXiv:2310.01361, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Gensim2: Scaling robot data generation with multi-modal and reasoning llms,

    P. Hua, M. Liu, A. Macaluso, Y . Lin, W. Zhang, H. Xu, and L. Wang, “Gensim2: Scaling robot data generation with multi-modal and reasoning llms,”arXiv preprint arXiv:2410.03645, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Humanoidgen: Data generation for bimanual dexterous manipulation via llm reason- ing,

    Z. Jing, S. Yang, J. Ao, T. Xiao, Y . Jiang, and C. Bai, “Humanoidgen: Data generation for bimanual dexterous manipulation via llm reason- ing,”arXiv preprint arXiv:2507.00833, 2025

    arXiv 2025

  23. [23]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

    2023

  24. [24]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhuet al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,”arXiv preprint arXiv:2303.05499, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    A tutorial on thompson sampling,

    D. J. Russo, B. Van Roy, A. Kazerouni, I. Osband, Z. Wenet al., “A tutorial on thompson sampling,”F oundations and Trends® in Machine Learning, vol. 11, no. 1, pp. 1–96, 2018

    2018

  26. [26]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,”Communications of the ACM, vol. 24, no. 6, pp. 381–395, 1981

    1981

  27. [27]

    Method for registration of 3-d shapes,

    P. J. Besl and N. D. McKay, “Method for registration of 3-d shapes,” inSensor fusion IV: control paradigms and data structures, vol. 1611. Spie, 1992, pp. 586–606

    1992

  28. [28]

    Multi-goal reinforcement learning environments for simulated franka emika panda robot,

    Q. Gallou ´edec, N. Cazin, E. Dellandr ´ea, and L. Chen, “Multi-goal reinforcement learning environments for simulated franka emika panda robot,”arXiv preprint arXiv:2106.13687, 2021

    Pith/arXiv arXiv 2021