Pith. sign in

REVIEW 3 major objections 4 minor 60 references

This survey organizes progress reward modeling for robotic learning into an interface, a construction mechanism, and a validation pipeline, making dense robot-learning rewards comparable across methods.

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

2026-08-01 09:39 UTC pith:DD35A4VX

load-bearing objection A useful organizing survey whose taxonomy is genuinely helpful; the unity claim is organizational, not ontological, so the stress-test worry doesn't land — but citation linkage and missing-reference sloppiness need fixing before it's a clean reference. the 3 major comments →

arxiv 2607.21655 v1 pith:DD35A4VX submitted 2026-07-22 cs.RO cs.CL

Progress Reward Modeling for Robotic Learning: A Comprehensive Survey

Jianshu Zhang , Keliang Wu , Haoran Lu , Anbang Liu , Ce Zhang , Weijie Yin , Chengxuan Qian , Xiyuan Yang
show 3 more authors
Zhenyu Pan Guo Ye Han Liu
This is my paper
classification cs.RO cs.CL
keywords progress rewardrobotic learningreward modelingreinforcement learningvision-language modelslong-horizon manipulationreward shapingsurvey
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 argues that the scattered field of progress reward modeling for robotic learning can be united under one descriptive framework. It treats every progress model as a task-conditioned black box with three faces: what state evidence it reads, how the goal is specified, and what output it emits—a score, a delta, a ranking, or a program. It then classifies construction methods into four paradigms—frozen foundation-model scoring, temporal or relative supervision, instruction-tuned progress prediction, and programmatic reward generation—and ties each to the data pipeline and evaluation protocol that actually validate it. A sympathetic reader comes away with a shared vocabulary: what a progress model is, how it is built, and what its numbers actually prove. The reason to care is that terminal success signals are too sparse for long-horizon robot learning, so a common account of dense progress signals is the field's missing piece.

Core claim

The central claim is taxonomic but consequential: although methods look disparate—success probabilities, CLIP similarities, pairwise preferences, reward programs—they can all be read as measurements of a single latent, history-dependent quantity called task progress, and their differences fall along three axes: interface, construction mechanism, and validation. The paper's organizing premise is that progress is not directly observable in a frame; it must be inferred from evidence about completed subgoals, current activity, and undone work. On that premise, the survey supplies a common input–output interface, four mechanisms for estimating progress, and a map showing which benchmark results c

What carries the argument

The load-bearing device is the interface abstraction: a progress model read as a task-conditioned map from state representation (single observation, temporal context, relational comparison, or state-access API) and goal specification (language, vision, or programmatic) to an output form (state-wise score, progress delta, ranking, or executable reward). Around this interface, the survey groups construction methods into four paradigms and attaches each to a supervision pipeline (human-driven, human-in-the-loop, fully automated) and an evaluation category (fidelity, robustness and generalization, downstream utility). This three-step organization is what lets methods with different architectures

Load-bearing premise

The entire framework assumes that very different outputs—success logits, CLIP similarities, preference rankings, delta scores, reward programs—all measure one underlying latent quantity called 'progress'; if they track different things, the unified view collapses into a labeling scheme.

What would settle it

Take two methods from different construction paradigms that the survey classifies under the same interface—say, a frozen foundation-model scorer and an instruction-tuned predictor, both single-observation with language goals—train them on the same task, and compare downstream policy success. If one consistently improves policies while the other only matches temporal order, 'progress' is not one construct and the framework's comparison logic fails. Alternatively, a benchmark with deliberately non-monotonic executions that shows a calibrated progress estimator and a monotonicity-fitting estimato

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

If this is right

  • Success signals alone underdetermine credit assignment; progress rewards can provide dense feedback before task completion and improve policy learning in long-horizon tasks.
  • Methods with different output forms—scores, deltas, rankings, programs—can be compared through a common interface, enabling direct method-to-method comparison.
  • Progress fidelity, robustness and generalization, and downstream utility are distinct validation targets: a reward can improve a policy without faithfully representing progress, and an accurate progress estimator may still be unsuitable for closed-loop control.
  • The framework exposes that current models lack fine-grained sensitivity, adaptive temporal modeling, efficient online inference, and long-horizon memory, pointing to concrete next steps.
  • Evaluation protocols that check temporal consistency do not certify calibration; benchmarks should separate these properties.

Where Pith is reading between the lines

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

  • If progress is truly a single latent construct, then a progress model trained for one task could serve as a general evaluator—monitoring, reranking, data filtering, planning—not just a reward; this is the paper's stated aspiration, but its feasibility depends on the unproven commensurability of different output forms.
  • The taxonomy implies a testable prediction: methods from different construction paradigms that share an interface and supervision source should be largely interchangeable in downstream performance; if they are not, the framework has missed a relevant axis.
  • An immediate extension of the paper's logic is a benchmark that deliberately uses non-monotonic executions—plateaus, reversals, retries, undoing—to separate chronological consistency from genuine progress estimation, since current benchmarks often conflate the two.
  • The paper's own limitation list implies that latency and memory, not accuracy alone, will determine whether progress models become usable at every control step; a two-tier design with a cheap local model and an expensive one called at transitions is one obvious direction implied by its discussion.

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

3 major / 4 minor

Summary. This survey proposes a unified framework for progress reward modeling in robotic learning. It organizes the field into three connected perspectives: the interface of a progress model (input representation, goal specification, output form), the construction mechanisms (frozen foundation-model scoring, temporal/relative supervision, instruction-tuned progress prediction, programmatic reward generation), and the data/benchmark pipelines (human-driven vs. automated supervision; fidelity, robustness, and downstream-utility evaluation). The central claim is that these three perspectives provide a shared vocabulary and allow methods that otherwise differ in architecture, supervision, and evaluation to be compared on a common footing.

Significance. If the framework is accepted, it would give a growing and fragmented literature a useful common language and a basis for comparing methods, which is a real contribution. The survey is honest about several important distinctions, notably the separation between progress fidelity and downstream utility (§4.4), and it identifies concrete open problems (non-monotonic execution, latency, long-horizon memory). The paper also ships an open-source repository of collected methods, which aids reproducibility of the survey's coverage. However, the strength of the unified-view claim is currently undercut by an unexamined commensurability assumption across output types and by citation inconsistencies in the taxonomy. These are fixable, but they are load-bearing for the paper's main thesis.

major comments (3)
  1. [§1, §2.3, §4.4] The paper's central claim is a 'unified view' in which all surveyed outputs—success probabilities, CLIP similarities, pairwise preferences, signed deltas, and executable reward programs—are treated as measurements of a single latent 'progress state' (§1). But §2.3 catalogs four output forms without an argument that they are commensurable surrogates of the same construct. The problem is not merely rhetorical: §4.4 explicitly concedes that 'a reward may improve a policy without faithfully representing progress, while an accurate progress estimator may still be unsuitable for closed-loop control.' This concession implies these signals are not interchangeable measurements of one quantity. Without an operational definition of the common latent variable, or an explicit mapping from each output type to that variable, comparative statements such as 'state-wise scores are the most direct' (§2.3)
  2. [§3.1 and References] In §3.1, the sentence 'GVL and OpenGVL further use VLMs to estimate temporal progress or frame ordering from videos (Ma et al., 2025; Budzianowski et al., 2025)' attributes GVL to Ma et al. (2025). The reference entry for Ma et al. (2025) is 'Vision language models are in-context value learners', not a paper titled or described as GVL. There is no separate GVL entry in the bibliography. Figure 1 also lists 'GVL' without any citation. Since the survey's contribution is precisely to classify methods under named categories, an attribution error of this kind is not a cosmetic issue: it undermines the verifiability of the taxonomy. All entries in Figure 1 and all method names used in the text must have a corresponding, correctly cited reference.
  3. [§2.3, §4.2, §4.4] The evaluation taxonomy in §4.2 lists five distinct fidelity dimensions (scalar calibration, temporal consistency, relative progress, task grounding, uncertainty), and §4.4 adds downstream-utility dimensions. The paper does not specify which of these evaluation dimensions are applicable to which output form from §2.3. For example, temporal consistency is meaningful for scalar and delta outputs, but is ill-posed for ranking outputs unless a total order is assumed; 'relative progress' benchmarks (§4.2) apply naturally to preference models but not to programmatic reward functions. Conversely, downstream-utility benchmarks may be the only sensible evaluation for programmatic rewards. Without an explicit mapping between output interface and evaluation dimensions, the promised connection between 'what a progress model is' and 'how its quality is validated' remains incomplete. Please add a tabl
minor comments (4)
  1. [Abstract and §1] Typos: 'providea' appears in both the abstract and the first line of §1; 'trajactory' appears in Figure 2.
  2. [References] The reference list contains inconsistent author formatting, e.g., 'Physical Intelligence, Kevin Black, ...' versus 'Physical Intelligence. Pi-0.7 ...'. The citation 'Intelligence et al., 2025' in §1 is confusing; consider using the full author list for the first citation.
  3. [Figure 1] Several entries in Figure 1 (e.g., DVD, Goal-Prox, FM Task-Spec, E2E-Robot-RL, Unsup-PR) are not explicitly connected to references in the text. Since the figure is the paper's visual summary, please add a legend or pointer to the relevant references in the caption.
  4. [§2.3] The phrase 'state-wise scores provide the most direct and reusable output form' is a value judgment that needs qualification; a preference output may be more directly learnable from human data, as the paper itself notes in the same subsection. Consider softening or providing a criterion for 'directness'.

Circularity Check

0 steps flagged

No circular reduction in this survey; minor self-citations anchor examples but do not carry the taxonomy.

full rationale

The paper is a survey, not a derivation: it contains no fitted parameters, no equations, and no quantity that is predicted from its own inputs. Its central claim is that progress reward methods can be organized along three axes — interface, construction paradigm, and data/evaluation pipeline — and this taxonomy is a retrospective classification of the cited literature, not a result derived from it. Categories such as 'frozen foundation-model scoring' and 'instruction-tuned progress prediction' are descriptive groupings; the paper does not reduce any surveyed output to a definition of progress, nor does it fit a parameter and then call a closely related quantity a prediction. The only overlap of concern is that several co-authored works (ProgressLM, RoboDojo, MagicSim, GarmentLab) appear as examples and framing citations. These are used as instances of the categories, not as justification for the categories, so this is a self-citation bias consideration rather than load-bearing circularity. The paper's assertion that 'Progress is therefore a latent and history-dependent task state' is a framing premise, not a derived theorem, and Section 4.4 itself concedes that reward utility and progress fidelity can diverge, which undercuts any claim that the framework is forced by a single construct. That is a coherence/scope concern, not a circular step. Accordingly, the score reflects only the minor presence of self-citations, not any circular derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

Survey paper: no free parameters fitted to data and no invented entities. The framework rests on three domain assumptions: (1) 'progress' is a single latent quantity commensurable across method outputs (Section 1); (2) the cited literature is characterized faithfully and coverage is representative (all sections); (3) temporal or relative ordering is a meaningful progress proxy, used in the benchmark taxonomy (§4.2) even while the fixed-rate view is criticized (§5).

axioms (3)
  • domain assumption The diverse surveyed methods all measure a common latent quantity — 'progress' — that can be abstracted into one interface without loss.
    Section 1 asserts 'Progress is therefore a latent and history-dependent task state'; the entire Section 2 interface abstraction presupposes that success probabilities, similarity scores, preference rankings, and reward programs are commensurable progress measurements.
  • domain assumption The cited literature is characterized faithfully and coverage is representative.
    The survey classifies every method into its categories and claims comprehensiveness; misattributions (GVL → Ma et al. 2025 in §3.1) and figure entries without references (REDS, R2VLM) show this premise is imperfect.
  • domain assumption Temporal or relative ordering in demonstrations is a meaningful progress proxy, adopted as a definitional benchmark family.
    §4.1: 'it assumes that task progress is sufficiently correlated with time'; the survey uses this as the standard automated supervision and defines temporal-consistency benchmarks (§4.2) on it, while §5 itself criticizes the fixed-rate assumption — an internal tension.

pith-pipeline@v1.3.0-alltime-deepseek · 15904 in / 17585 out tokens · 134764 ms · 2026-08-01T09:39:46.191907+00:00 · methodology

0 comments
read the original abstract

Robotic learning takes place in dynamic environments with large behavior spaces. A terminal success signal only tells the robot whether the task is completed. It does not explain whether the current behavior is making progress, remaining unchanged, or undoing earlier progress. For this reason, recent studies have increasingly explored progress rewards that provide feedback during task execution. However, the current literature lacks a shared framework. Existing methods use different observations, goal specifications, output signals, supervision sources, and evaluation protocols. This makes it difficult to compare them and understand what their results actually validate. In this survey, we provide a unified view of progress reward modeling for robotic learning. We organize the field in three connected steps. We first study the interface of a progress model. This defines the problem from the outside by asking what information the model receives and what form of progress signal it produces. We then move inside the model and study the methods used to construct this signal. This reveals the different assumptions and mechanisms behind progress estimation and reward generation. Finally, we examine the data and benchmarks that support these methods. This shows how progress supervision is obtained and what different evaluations actually measure. Together, these three perspectives connect what a progress model is, how it is built, and how its quality is validated. We further summarize the main limitations of current approaches and discuss future research directions.

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

60 extracted references · 3 linked inside Pith

  1. [1]

    2026 , eprint=

    ARM: Advantage Reward Modeling for Long-Horizon Manipulation , author=. 2026 , eprint=

    2026

  2. [2]

    2026 , eprint=

    Generalizable Dense Reward for Long-Horizon Robotic Tasks , author=. 2026 , eprint=

    2026

  3. [3]

    2026 , eprint=

    Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons , author=. 2026 , eprint=

    2026

  4. [4]

    2026 , eprint=

    TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics , author=. 2026 , eprint=

    2026

  5. [5]

    2026 , eprint=

    PROGRESSLM: Towards Progress Reasoning in Vision-Language Models , author=. 2026 , eprint=

    2026

  6. [6]

    2026 , eprint=

    RoboReward: General-Purpose Vision-Language Reward Models for Robotics , author=. 2026 , eprint=

    2026

  7. [7]

    arXiv preprint arXiv:2512.23703 , year=

    Robo-Dopamine: General Process Reward Modeling for High-Precision Robotic Manipulation , author=. arXiv preprint arXiv:2512.23703 , year=

    arXiv

  8. [8]

    2025 , eprint=

    SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation , author=. 2025 , eprint=

    2025

  9. [9]

    arXiv preprint arXiv:2509.17321 , year=

    OpenGVL -- Benchmarking Visual Temporal Progress for Data Curation , author=. arXiv preprint arXiv:2509.17321 , year=

    arXiv

  10. [10]

    arXiv preprint arXiv:2509.15937 , year=

    A Vision-Language-Action-Critic Model for Robotic Real-World Reinforcement Learning , author=. arXiv preprint arXiv:2509.15937 , year=

    arXiv

  11. [11]

    2025 , eprint=

    ReWiND: Language-Guided Rewards Teach Robot Policies without New Demonstrations , author=. 2025 , eprint=

    2025

  12. [12]

    The Thirteenth International Conference on Learning Representations , year=

    Subtask-Aware Visual Reward Learning from Segmented Demonstrations , author=. The Thirteenth International Conference on Learning Representations , year=

  13. [13]

    The Thirteenth International Conference on Learning Representations , year=

    Vision Language Models are In-Context Value Learners , author=. The Thirteenth International Conference on Learning Representations , year=

  14. [14]

    2025 , volume=

    Chen, Letian and Moorman, Nina Marie and Gombolay, Matthew Craig , booktitle=. 2025 , volume=

    2025

  15. [15]

    2025 , eprint=

    On-Robot Reinforcement Learning with Goal-Contrastive Rewards , author=. 2025 , eprint=

    2025

  16. [16]

    2025 , eprint=

    Video-Language Critic: Transferable Reward Functions for Language-Conditioned Robotics , author=. 2025 , eprint=

    2025

  17. [17]

    The Thirteenth International Conference on Learning Representations , year=

    VICtoR: Learning Hierarchical Vision-Instruction Correlation Rewards for Long-horizon Manipulation , author=. The Thirteenth International Conference on Learning Representations , year=

  18. [18]

    2024 , eprint=

    Rank2Reward: Learning Shaped Reward Functions from Passive Video , author=. 2024 , eprint=

    2024

  19. [19]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    RL-VLM-F: Reinforcement Learning from Vision Language Foundation Model Feedback , author=. Proceedings of the 41st International Conference on Machine Learning , pages=. 2024 , volume=

    2024

  20. [20]

    Proceedings of the 41st International Conference on Machine Learning , year=

    Code as Reward: Empowering Reinforcement Learning with VLMs , author=. Proceedings of the 41st International Conference on Machine Learning , year=

  21. [21]

    2024 , url=

    Guan, Lin and Zhou, Yifan and Liu, Denis and Zha, Yantian and Ben Amor, Heni and Kambhampati, Subbarao , booktitle=. 2024 , url=

    2024

  22. [22]

    Video Prediction Models as Rewards for Reinforcement Learning , url=

    Escontrela, Alejandro and Adeniji, Ademi and Yan, Wilson and Jain, Ajay and Peng, Xue Bin and Goldberg, Ken and Lee, Youngwoon and Hafner, Danijar and Abbeel, Pieter , booktitle=. Video Prediction Models as Rewards for Reinforcement Learning , url=

  23. [23]

    arXiv preprint arXiv:2312.09187 , year=

    Vision-Language Models as a Source of Rewards , author=. arXiv preprint arXiv:2312.09187 , year=

    Pith/arXiv arXiv

  24. [24]

    The Twelfth International Conference on Learning Representations , year=

    Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning , author=. The Twelfth International Conference on Learning Representations , year=

  25. [25]

    The Twelfth International Conference on Learning Representations , year=

    Eureka: Human-Level Reward Design via Coding Large Language Models , author=. The Twelfth International Conference on Learning Representations , year=

  26. [26]

    2023 , eprint=

    Robot Fine-Tuning Made Easy: Pre-Training Rewards and Policies for Autonomous Real-World Reinforcement Learning , author=. 2023 , eprint=

    2023

  27. [27]

    The Twelfth International Conference on Learning Representations , year=

    Text2Reward: Reward Shaping with Language Models for Reinforcement Learning , author=. The Twelfth International Conference on Learning Representations , year=

  28. [28]

    2023 , eprint=

    Robotic Offline RL from Internet Videos via Value-Function Pre-Training , author=. 2023 , eprint=

    2023

  29. [29]

    2023 , volume=

    Ma, Yecheng Jason and Kumar, Vikash and Zhang, Amy and Bastani, Osbert and Jayaraman, Dinesh , booktitle=. 2023 , volume=

    2023

  30. [30]

    2023 , editor=

    Nair, Suraj and Rajeswaran, Aravind and Kumar, Vikash and Finn, Chelsea and Gupta, Abhinav , booktitle=. 2023 , editor=

    2023

  31. [31]

    Proceedings of The 7th Conference on Robot Learning , pages=

    Language to Rewards for Robotic Skill Synthesis , author=. Proceedings of The 7th Conference on Robot Learning , pages=. 2023 , editor=

    2023

  32. [32]

    2024 , volume=

    Liu, Runze and Du, Yali and Bai, Fengshuo and Lyu, Jiafei and Li, Xiu , booktitle=. 2024 , volume=

    2024

  33. [33]

    Proceedings of The 2nd Conference on Lifelong Learning Agents , pages=

    Vision-Language Models as Success Detectors , author=. Proceedings of The 2nd Conference on Lifelong Learning Agents , pages=. 2023 , editor=

    2023

  34. [34]

    2023 , eprint=

    VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training , author=. 2023 , eprint=

    2023

  35. [35]

    Proceedings of the 39th International Conference on Machine Learning , pages=

    Zero-Shot Reward Specification via Grounded Natural Language , author=. Proceedings of the 39th International Conference on Machine Learning , pages=. 2022 , editor=

    2022

  36. [36]

    Proceedings of The 4th Annual Learning for Dynamics and Control Conference , pages=

    Can Foundation Models Perform Zero-Shot Task Specification For Robot Manipulation? , author=. Proceedings of The 4th Annual Learning for Dynamics and Control Conference , pages=. 2022 , volume=

    2022

  37. [37]

    Generalizable Imitation Learning from Observation via Inferring Goal Proximity , url=

    Lee, Youngwoon and Szot, Andrew and Sun, Shao-Hua and Lim, Joseph J , booktitle=. Generalizable Imitation Learning from Observation via Inferring Goal Proximity , url=

  38. [38]

    Proceedings of the 5th Conference on Robot Learning , pages=

    Learning Language-Conditioned Robot Behavior from Offline Data and Crowd-Sourced Annotation , author=. Proceedings of the 5th Conference on Robot Learning , pages=. 2022 , volume=

    2022

  39. [39]

    In-The-Wild

    Learning Generalizable Robotic Reward Functions from "In-The-Wild" Human Videos , author=. 2021 , eprint=

    2021

  40. [40]

    2020 , eprint=

    Scaling data-driven robotics with reward sketching and batch reinforcement learning , author=. 2020 , eprint=

    2020

  41. [41]

    2019 , eprint=

    End-to-End Robotic Reinforcement Learning without Reward Engineering , author=. 2019 , eprint=

    2019

  42. [42]

    2017 , eprint=

    Unsupervised Perceptual Rewards for Imitation Learning , author=. 2017 , eprint=

    2017

  43. [43]

    2026 , eprint=

    ProgressVLA: Progress-Guided Diffusion Policy for Vision-Language Robotic Manipulation , author=. 2026 , eprint=

    2026

  44. [44]

    2026 , eprint=

    Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress , author=. 2026 , eprint=

    2026

  45. [45]

    GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation , url =

    Lu, Haoran and Wu, Ruihai and Li, Yitong and Li, Sijie and Zhu, Ziyu and Ning, Chuanruo and Shen, Yan and Luo, Longzan and Chen, Yuanpei and Dong, Hao , booktitle =. GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation , url =. doi:10.52202/079017-0379 , editor =

    doi:10.52202/079017-0379

  46. [46]

    ArXiv , year=

    Learning to Feel the Future: DreamTacVLA for Contact-Rich Manipulation , author=. ArXiv , year=

  47. [47]

    Neural Dynamics Augmented Diffusion Policy , year=

    Wu, Ruihai and Chen, Haozhe and Zhang, Mingtong and Lu, Haoran and Li, Yitong and Li, Yunzhu , booktitle=. Neural Dynamics Augmented Diffusion Policy , year=

  48. [48]

    Pi-0.7: A Steerable Generalist Robotic Foundation Model with Emergent Capabilities , year =

  49. [49]

    GEN-0: Embodied Foundation Models That Scale with Physical Interaction , year =

  50. [50]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots , year =

  51. [51]

    ArXiv , year=

    0.5: a Vision-Language-Action Model with Open-World Generalization , author=. ArXiv , year=

  52. [52]

    2026 , url=

    World-Value-Action Model: Implicit Planning for Vision-Language-Action Systems , author=. 2026 , url=

    2026

  53. [53]

    2026 , eprint=

    Fast-WAM: Do World Action Models Need Test-time Future Imagination? , author=. 2026 , eprint=

    2026

  54. [54]

    2026 , howpublished =

    Ye, Seonghyeon and Ge, Yunhao and Zheng, Kaiyuan and Jang, Joel , title =. 2026 , howpublished =

    2026

  55. [55]

    2023 , eprint=

    LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning , author=. 2023 , eprint=

    2023

  56. [56]

    arXiv preprint arXiv:2607.04434 , year=

    RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies , author=. arXiv preprint arXiv:2607.04434 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2506.18088 , year=

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation , author=. arXiv preprint arXiv:2506.18088 , year=

    Pith/arXiv arXiv

  58. [58]

    2026 , eprint=

    MagicSim: A Unified Infrastructure for Executable Embodied Interaction , author=. 2026 , eprint=

    2026

  59. [59]

    2026 , eprint=

    RMBench: Memory-Dependent Robotic Manipulation Benchmark with Insights into Policy Design , author=. 2026 , eprint=

    2026

  60. [60]

    ArXiv , year=

    Mind to Hand: Purposeful Robotic Control via Embodied Reasoning , author=. ArXiv , year=