pith. machine review for the scientific record. sign in

arxiv: 2605.10118 · v1 · submitted 2026-05-11 · 💻 cs.RO

Recognition: no theorem link

Plan in Sandbox, Navigate in Open Worlds: Learning Physics-Grounded Abstracted Experience for Embodied Navigation

Zhixuan Shen , Jiawei Du , Ziyu Guo , Han Luo , Lilan Peng , Joey Tianyi Zhou , Haonan Luo , Tianrui Li
Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:27 UTC · model grok-4.3

classification 💻 cs.RO
keywords embodied navigationphysics-grounded abstractionsemantic environmentsreinforcement learningpolicy transfervision-language modelsrobot deployment
0
0 comments X

The pith

Planning in simplified physics abstractions improves transfer of navigation policies to real robots and open worlds.

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

Vision-language models reason well in general but face limits in embodied navigation from scarce aligned vision-control data. Photorealistic simulators generate training data yet often yield policies that fail to carry over to physical robots. The work shows that rehearsing plans inside a physics-grounded semantic abstraction, much like human mental simulation, produces policies that succeed more often in navigation tasks and deploy successfully on real indoor robots.

Core claim

SAGE lets agents learn by first building diverse physics-constrained semantic environments, then distilling policies via reinforcement learning that uses asymmetric adaptive clipping for stable updates, and finally bridging the resulting abstract policy to open-world control and physical deployment.

What carries the argument

The physics-grounded semantic abstraction that replaces photorealistic visuals during policy learning, organized through the Genesis, Evolution, and Navigation phases.

If this is right

  • Policies refined in the abstract setting achieve higher success in planner-assisted embodied navigation.
  • The learned behaviors transfer to physical indoor robot hardware without retraining on photorealistic data.
  • Asymmetric adaptive clipping keeps reinforcement learning updates stable while distilling experience.
  • Diverse training experiences arise from semantic environments that avoid the cost of detailed visual rendering.

Where Pith is reading between the lines

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

  • The same abstraction tactic could cut the data volume needed to train other embodied skills such as grasping or locomotion.
  • Testing the method on outdoor or multi-robot scenarios would show how much semantic detail is required for reliable transfer.
  • The approach may generalize to tasks that combine planning with language instructions beyond pure navigation.

Load-bearing premise

That the simplified semantic and physics rules in the abstract environments capture enough real-world dynamics for the learned policy to work when moved to actual robots and open settings.

What would settle it

A physical robot trial in which the policy trained only on the abstracted experiences fails to navigate where unmodeled factors such as exact friction or lighting appear.

Figures

Figures reproduced from arXiv: 2605.10118 by Han Luo, Haonan Luo, Jiawei Du, Joey Tianyi Zhou, Lilan Peng, Tianrui Li, Zhixuan Shen, Ziyu Guo.

Figure 1
Figure 1. Figure 1: (a): Our SAGE framework utilizes a physics-grounded sandbox for self-evolving data generation and policy optimization, enabling the agent to bridge the gap between sandbox and open-world. (b): We demonstrate real-world robotic demonstrations powered by SAGE, showcasing its Sim2Real generalization capabilities. (c): Unlike other VLM or RL paradigms, our approach uniquely combines physics-grounded interactio… view at source ↗
Figure 2
Figure 2. Figure 2: The SAGE Framework. The system operates in three phases: (a) Genesis: A sandbox environment ES synthesizes task-oriented experience rules Kexp. (b) Evolution: The policy πθ is optimized via a hybrid prompt-augmented sampling strategy, utilizing both standard and augmented contexts. (c) Navigation: The embodied navigation policy relies on evolving policy πθ and retrieved experience Cret to execute a ∗ t in … view at source ↗
Figure 3
Figure 3. Figure 3: Asymmetric Adaptive Clipping (AAC). While both standard and augmented samples share a conservative lower bound (1 − ϵstd) to prevent policy collapse under A < 0, augmented experience samples feature an expanded upper bound ϵexp under A > 0. The shaded region indicates the additional optimization space allocated for aggressive knowledge absorption from high￾reward augmented trajectories. bound dynamically: … view at source ↗
Figure 4
Figure 4. Figure 4: (a)&(b): Impact of fixed and dynamic experience-injection probabilities on navigation performance. We compare fixed η ∈ {0.0, 0.5, 0.8, 1.0} with a validation-dependent dynamic schedule, where the red star curve uses ηinit = 0.8, ηmin = 0.0, and Rtarget = 1.5. (c)&(d): Impact of upper clipping threshold ϵexp on navigation performance. All experiments use the model with 2B parameters on A-EQA [PITH_FULL_IM… view at source ↗
Figure 5
Figure 5. Figure 5: (a): Comparison of data composition strategies. (b): Impact of data scale on model performance. All experiments use the model with 2B parameters on A-EQA. ing complementary environments during the Genesis phase, the agent learns more robust navigation priors that gener￾alize better to unseen test scenarios, effectively preventing overfitting to specific simulator artifacts. Scalability with Minimal Data. T… view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of sandbox task categories. (a) Sandbox Q&A (b) A-EQA [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the word cloud. rules using regular expressions. The entire trajectory is discarded if the generated output fails to match the required templates, specifically the defined Task/Question/Answer format for EQA or the logical IF-AND-THEN structure for experience descriptions. Only candidates that successfully pass these rigorous formatting checks without raising parsing exceptions are seriali… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative example of the SAGE agent on A-EQA task. Given a natural language query, the SAGE agent maintains a Frontier Buffer Ft for exploration candidates and a Memory Buffer Mt for semantic history. (Top) The evolving top-down occupancy map and the robot’s trajectory (blue line), showing the progressive exploration of the environment.(Middle) The Frontier Buffer (blue block) stores candidate nodes for … view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative example of the SAGE agent on GOAT-Bench task. The agent is instructed to find a specific object based on a spatial description. (Top) The robot’s trajectory (blue line) navigates through multiple rooms to locate the target. (Middle) The distinct Frontier Buffer Ft and Memory Buffer Mt. Throughout the steps, the agent actively selects frontier nodes to traverse the hallway and enter the bedroom.… view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of SAGE’s deployment strategy for real-world robot navigation platforms. Third-person Perspective & Top-Down Map Frontier Buffer Memory Buffer Steps Selected Action What is the object on the table beside the large wooden door? Frontier View Memory View Top-Down Map 1 1 1 2 1 2 1 2 1 2 3 3 3 A potted plant. 2 3 3 1 2 3 4 1 1 2 1 2 2 [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization in real-world environment. Best viewed when zoomed in. J.2. Computational Cost and Latency Analysis The real-world deployment relies on a client-server architecture. We provide a breakdown of the computational cost and latency for a single navigation step in [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Prompt for Sandbox Task Synthesis. The placeholders {objects} and {core relationship} are replaced by the detected objects and the scene graph. {target task} is replaced by the eight distinct task categories. <IMG> is replaced by the front view of the final waypoint. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Prompt for Sandbox Task Synthesis. The placeholders {Q}, {objects} and {core relationship} are replaced by the synthetic task, detected objects and the scene graph. <IMG> is replaced by the front view of the final waypoint. <EXP> Guidance from Memory: {retrieved_experience} (Instruction: Carefully check if any candidate image contains the visual cues mentioned in the 'IF' condition of this experience. If … view at source ↗
Figure 14
Figure 14. Figure 14: Experience template. The placeholder {retrieved experience} is replaced by the retrieved sandbox experience. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Prompt for Sandbox Task Synthesis. The placeholders {Q}, {exp section} and {objects} are replaced by the sandbox synthetic task, experience (see [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
read the original abstract

Vision-Language Models (VLMs) have demonstrated exceptional general reasoning capabilities. However, their performance in embodied navigation remains hindered by a scarcity of aligned open-world vision and robot control data. Despite simulators providing a cost-effective alternative for data collection, the inherent reliance on photorealistic simulations often limits the transferability of learned policies. To this end, we propose \textit{\textbf{S}andbox-\textbf{A}bstracted \textbf{G}rounded \textbf{E}xperience} (\textbf{\textit{SAGE}}), a framework that enables agents to learn within a physics-grounded semantic abstraction rather than a photorealistic simulation, mimicking the human capacity for mental simulation where plans are rehearsed in simplified physics abstractions before execution. \textit{SAGE} system operates via three synergistic phases: (1) \textit{Genesis}: constructing diverse, physics-constrained semantic environments to bootstrap experience; (2) \textit{Evolution}: distilling experiences through Reinforcement Learning (RL), utilizing a novel asymmetric adaptive clipping mechanism to stabilize updates; (3) \textit{Navigation}: bridging the abstract policy to open-world control. We demonstrate that \textit{SAGE} significantly improves planner-assisted embodied navigation, achieving a 53.21\% LLM-Match Success Rate on A-EQA (+9.7\% over baseline), while showing encouraging transfer to physical indoor robot deployment.

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 / 2 minor

Summary. The paper proposes SAGE, a three-phase framework (Genesis, Evolution, Navigation) for learning physics-grounded semantic abstractions to improve embodied navigation with VLMs. Genesis builds diverse physics-constrained semantic environments; Evolution distills experiences via RL with a novel asymmetric adaptive clipping mechanism; Navigation bridges the resulting policy to open-world control. The central empirical claim is a 53.21% LLM-Match Success Rate on A-EQA (+9.7% over baseline) with encouraging transfer to physical indoor robots.

Significance. If the transfer results hold under rigorous evaluation, the work would be significant for sim-to-real robotics: it offers a pathway to policy learning that avoids photorealistic simulation costs while leveraging abstracted physics, potentially improving data efficiency and generalization for VLM-driven navigation. The specific A-EQA gains and the three-phase pipeline design are clear strengths.

major comments (2)
  1. [Physical deployment results] Physical deployment results (likely §5 or equivalent): The claim of 'encouraging transfer to physical indoor robot deployment' is supported only by qualitative description with no success rates, trial counts, ablation on domain gaps, or analysis of sensor noise/actuation delays. This is load-bearing for the paper's core distinction from pure simulation training, as the skeptic correctly notes that unquantified gaps leave the Navigation phase's effectiveness unverified.
  2. [Evolution phase] Evolution phase and asymmetric adaptive clipping: The manuscript describes this as a novel mechanism to stabilize RL updates but provides no equation, pseudocode, or ablation isolating its contribution to the reported +9.7% gain. Without these details, it is impossible to assess whether the improvement stems from the clipping innovation or other factors in the pipeline.
minor comments (2)
  1. [Abstract] Abstract: The performance numbers (53.21%, +9.7%) are presented without reference to the exact baseline method or number of evaluation episodes, reducing immediate interpretability.
  2. [Evaluation metrics] Notation and terminology: 'LLM-Match Success Rate' is used without an explicit definition or reference to how matching is computed, which should be clarified in the methods or evaluation section.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and describe the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: Physical deployment results (likely §5 or equivalent): The claim of 'encouraging transfer to physical indoor robot deployment' is supported only by qualitative description with no success rates, trial counts, ablation on domain gaps, or analysis of sensor noise/actuation delays. This is load-bearing for the paper's core distinction from pure simulation training, as the skeptic correctly notes that unquantified gaps leave the Navigation phase's effectiveness unverified.

    Authors: We agree that the physical deployment is presented qualitatively and that additional quantification would strengthen the claim. The current manuscript positions the physical transfer as a preliminary demonstration of the Navigation phase's bridging capability rather than a comprehensive real-world evaluation. In the revision we will expand the deployment section with more details on the robot platform, trial protocol, observed effects of sensor noise and actuation delays, and how the abstracted policy is mapped to low-level control. We will also add an explicit limitations paragraph noting the absence of full success-rate statistics and ablations. Because the physical experiments were limited in scope, we cannot supply new quantitative metrics; the revision will therefore be partial and descriptive. revision: partial

  2. Referee: Evolution phase and asymmetric adaptive clipping: The manuscript describes this as a novel mechanism to stabilize RL updates but provides no equation, pseudocode, or ablation isolating its contribution to the reported +9.7% gain. Without these details, it is impossible to assess whether the improvement stems from the clipping innovation or other factors in the pipeline.

    Authors: We thank the referee for identifying this omission. The asymmetric adaptive clipping is a core technical contribution of the Evolution phase; it modifies the standard PPO clipping range asymmetrically according to the sign and magnitude of the advantage estimate to reduce destructive policy updates in physics-constrained semantic environments. In the revised manuscript we will insert the exact mathematical formulation, the corresponding pseudocode, and a dedicated ablation that compares SAGE against an otherwise identical pipeline using symmetric clipping. This will isolate the mechanism's contribution to the observed performance gain. revision: yes

standing simulated objections not resolved
  • Quantitative success rates, trial counts, and domain-gap ablations for physical robot deployment, because only qualitative demonstrations were performed.

Circularity Check

0 steps flagged

No circularity: empirical framework supported by reported metrics

full rationale

The paper proposes a three-phase framework (Genesis for semantic environments, Evolution via RL with adaptive clipping, Navigation for bridging to open-world control) and supports its claims with concrete experimental results including a 53.21% LLM-Match Success Rate on A-EQA (+9.7% over baseline) plus qualitative transfer notes. No equations, fitted parameters presented as predictions, self-definitional constructs, or load-bearing self-citations appear in the provided text. The derivation chain consists of system design followed by benchmark evaluation rather than any reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Main unstated premise is transferability of abstracted experience; no free parameters or invented entities detailed in abstract.

axioms (1)
  • domain assumption Semantic abstraction with basic physics constraints captures sufficient information for policy transfer to real open worlds
    Invoked to justify bridging abstract training to physical deployment.

pith-pipeline@v0.9.0 · 5570 in / 1070 out tokens · 27582 ms · 2026-05-12T03:27:52.804461+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

60 extracted references · 60 canonical work pages · 8 internal anchors

  1. [1]

    International conference on machine learning , pages=

    Universal value function approximators , author=. International conference on machine learning , pages=. 2015 , organization=

    work page 2015

  2. [2]

    Goal-conditioned re- inforcement learning: Problems and solutions.arXiv preprint arXiv:2201.08299,

    Goal-conditioned reinforcement learning: Problems and solutions , author=. arXiv preprint arXiv:2201.08299 , year=

    work page arXiv

  3. [3]

    Habitat-Matterport 3D Dataset (HM3D): 1000 Large-scale 3D Environments for Embodied AI

    Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai , author=. arXiv preprint arXiv:2109.08238 , year=

    work page internal anchor Pith review arXiv

  4. [4]

    2025 , howpublished =

    InteriorGS: A 3D Gaussian Splatting Dataset of Semantically Labeled Indoor Scenes , author =. 2025 , howpublished =

    work page 2025

  5. [5]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Habitat-matterport 3d semantics dataset , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  6. [6]

    DeepSeek-V3 Technical Report

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    GPT-4 Technical Report

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Qwen3 Technical Report

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Goat: Go to any thing,

    Goat: Go to any thing , author=. arXiv preprint arXiv:2311.06430 , year=

    work page arXiv

  10. [10]

    IEEE Robotics and Automation Letters , volume=

    Openin: Open-vocabulary instance-oriented navigation in dynamic domestic environments , author=. IEEE Robotics and Automation Letters , volume=. 2025 , publisher=

    work page 2025

  11. [11]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

    work page

  12. [12]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Unigoal: Towards universal zero-shot goal-oriented navigation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  13. [13]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Embodied question answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    work page

  14. [14]

    arXiv preprint arXiv:2403.15941 , year=

    Explore until confident: Efficient exploration for embodied question answering , author=. arXiv preprint arXiv:2403.15941 , year=

    work page arXiv

  15. [15]

    arXiv preprint arXiv:2410.20263 , year=

    Efficienteqa: An efficient approach for open vocabulary embodied question answering , author=. arXiv preprint arXiv:2410.20263 , year=

    work page arXiv

  16. [16]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    TANGO: training-free embodied AI agents for open-world tasks , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  17. [17]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Openeqa: Embodied question answering in the era of foundation models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  18. [18]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0 , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    work page 2024

  19. [19]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    3D-mem: 3D scene memory for embodied exploration and reasoning , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  20. [20]

    arXiv preprint arXiv:2510.20310 , year=

    Multi-Step Reasoning for Embodied Question Answering via Tool Augmentation , author=. arXiv preprint arXiv:2510.20310 , year=

    work page arXiv

  21. [21]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  22. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Goat-bench: A benchmark for multi-modal lifelong navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  23. [23]

    arXiv preprint arXiv:2410.02751 , year=

    Relic: A recipe for 64k steps of in-context reinforcement learning for embodied ai , author=. arXiv preprint arXiv:2410.02751 , year=

    work page arXiv

  24. [24]

    arXiv preprint arXiv:2412.11484 , year=

    Efficient policy adaptation with contrastive prompt ensemble for embodied agents , author=. arXiv preprint arXiv:2412.11484 , year=

    work page arXiv

  25. [25]

    arXiv preprint arXiv:2505.08361 , year=

    Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning , author=. arXiv preprint arXiv:2505.08361 , year=

    work page arXiv

  26. [26]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Roomtour3d: Geometry-aware video-instruction tuning for embodied navigation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  27. [27]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Navigation world models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  28. [28]

    arXiv preprint arXiv:2509.25687 , year=

    Omninav: A unified framework for prospective exploration and visual-language navigation , author=. arXiv preprint arXiv:2509.25687 , year=

    work page arXiv

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Object goal navigation using goal-oriented semantic exploration , author=. Advances in Neural Information Processing Systems , volume=

    work page

  30. [30]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Instance-aware exploration-verification-exploitation for instance imagegoal navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  31. [31]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    IGL-Nav: Incremental 3D Gaussian Localization for Image-goal Navigation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  32. [32]

    arXiv preprint arXiv:2412.14480 , year=

    Grapheqa: Using 3d semantic scene graphs for real-time embodied question answering , author=. arXiv preprint arXiv:2412.14480 , year=

    work page arXiv

  33. [33]

    arXiv preprint arXiv:2503.11117 , year=

    Beyond the destination: A novel benchmark for exploration-aware embodied question answering , author=. arXiv preprint arXiv:2503.11117 , year=

    work page arXiv

  34. [34]

    arXiv preprint arXiv:2508.09423 , year=

    Distilling LLM Prior to Flow Model for Generalizable Agent's Imagination in Object Goal Navigation , author=. arXiv preprint arXiv:2508.09423 , year=

    work page arXiv

  35. [35]

    arXiv preprint arXiv:2506.01031 , year=

    NavBench: Probing Multimodal Large Language Models for Embodied Navigation , author=. arXiv preprint arXiv:2506.01031 , year=

    work page arXiv

  36. [36]

    Zhang, Lingfeng and Hao, Xiaoshuai and Tang, Yingbo and Fu, Haoxiang and Zheng, Xinyu and Wang, Pengwei and Wang, Zhongyuan and Ding, Wenbo and Zhang, Shanghang , journal=. NavA\^

    work page

  37. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Towards learning a generalist model for embodied navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  38. [38]

    International Conference on Machine Learning , pages=

    Esc: Exploration with soft commonsense constraints for zero-shot object navigation , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    work page 2023

  39. [39]

    2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    L3mvn: Leveraging large language models for visual target navigation , author=. 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2023 , organization=

    work page 2023

  40. [40]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    work page 2025

  41. [41]

    arXiv preprint arXiv:2505.01458 , year=

    A Survey of Robotic Navigation and Manipulation with Physics Simulators in the Era of Embodied AI , author=. arXiv preprint arXiv:2505.01458 , year=

    work page arXiv

  42. [42]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    Advances in neural information processing systems , volume=

    Habitat 2.0: Training home assistants to rearrange their habitat , author=. Advances in neural information processing systems , volume=

    work page

  44. [44]

    and Peng, Y

    A comprehensive survey of self-evolving ai agents: A new paradigm bridging foundation models and lifelong agentic systems , author=. arXiv preprint arXiv:2508.07407 , year=

    work page arXiv

  45. [45]

    arXiv preprint arXiv:2502.04780 , year=

    Sirius: Self-improving multi-agent systems via bootstrapped reasoning , author=. arXiv preprint arXiv:2502.04780 , year=

    work page arXiv

  46. [46]

    arXiv preprint arXiv:2410.16946 , year=

    Self-evolving multi-agent collaboration networks for software development , author=. arXiv preprint arXiv:2410.16946 , year=

    work page arXiv

  47. [47]

    Advances in Neural Information Processing Systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in Neural Information Processing Systems , volume=

    work page

  48. [48]

    arXiv preprint arXiv:2310.05915 , year=

    Fireact: Toward language agent fine-tuning , author=. arXiv preprint arXiv:2310.05915 , year=

    work page arXiv

  49. [49]

    arXiv preprint arXiv:2405.17418 , year=

    A Self-Correcting Vision-Language-Action Model for Fast and Slow System Manipulation , author=. arXiv preprint arXiv:2405.17418 , year=

    work page arXiv

  50. [50]

    Eureka: Human-Level Reward Design via Coding Large Language Models

    Eureka: Human-level reward design via coding large language models , author=. arXiv preprint arXiv:2310.12931 , year=

    work page internal anchor Pith review arXiv

  51. [51]

    arXiv preprint arXiv:2507.13152 , year=

    Se-vln: A self-evolving vision-language navigation framework based on multimodal large language models , author=. arXiv preprint arXiv:2507.13152 , year=

    work page arXiv

  52. [52]

    arXiv preprint arXiv:2509.24910 , year=

    Learning Goal-Oriented Language-Guided Navigation with Self-Improving Demonstrations at Scale , author=. arXiv preprint arXiv:2509.24910 , year=

    work page arXiv

  53. [53]

    Agentevolver: Towards efficient self-evolving agent system.arXiv, 2025

    AgentEvolver: Towards Efficient Self-Evolving Agent System , author=. arXiv preprint arXiv:2511.10395 , year=

    work page arXiv

  54. [54]

    arXiv preprint arXiv:2412.08467 , year=

    Bootstrapping language-guided navigation learning with self-refining data flywheel , author=. arXiv preprint arXiv:2412.08467 , year=

    work page arXiv

  55. [55]

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

    Flashattention-2: Faster attention with better parallelism and work partitioning , author=. arXiv preprint arXiv:2307.08691 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2501.12345 , year=

    Easyr1: An efficient, scalable, multi-modality rl training framework , author=. arXiv preprint arXiv:2501.12345 , year=

    work page arXiv

  57. [57]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Yolo-world: Real-time open-vocabulary object detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  58. [58]

    Poliformer: Scaling on-policy rl with transformers results in masterful navigators.arXiv preprint arXiv:2406.20083,

    Poliformer: Scaling on-policy rl with transformers results in masterful navigators , author=. arXiv preprint arXiv:2406.20083 , year=

    work page arXiv

  59. [59]

    GPT-4o System Card

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Rest-mcts*: Llm self-training via process reward guided tree search , author=. Advances in Neural Information Processing Systems , volume=

    work page