pith. machine review for the scientific record. sign in

arxiv: 2605.13119 · v1 · submitted 2026-05-13 · 💻 cs.RO · cs.AI· cs.CV

Recognition: no theorem link

Towards Long-horizon Embodied Agents with Tool-Aligned Vision-Language-Action Models

Changxing Liu, Minhao Xiong, Siheng Chen, Weixin Li, Yichen Xiong, Yuanzhuo Ding, Zhipeng Zhang, Zixing Lei

Authors on Pith no claims yet

Pith reviewed 2026-05-14 18:30 UTC · model grok-4.3

classification 💻 cs.RO cs.AIcs.CV
keywords vision-language-action modelslong-horizon tasksembodied agentstool userobot planningpost-training
0
0 comments X

The pith

Splitting long robot tasks between a high-level planner and specialized action tools raises success rates on extended sequences.

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

The paper claims that single vision-language-action models struggle on long-horizon tasks because they must simultaneously manage extended temporal planning and varied physical operations. It proposes dividing the work so a high-level vision-language model handles scene analysis, global planning, and recovery while a family of specialized VLA tools each execute one bounded subtask. A new tool-family interface provides explicit selection and progress feedback, allowing the planner to replan on events rather than polling continuously. Tool-Aligned Post-Training creates training data and adapters that make the tools follow high-level calls more reliably. If the split succeeds, embodied agents could complete longer sequences without the performance collapse that occurs when one model tries to do everything.

Core claim

VLAs-as-Tools distributes long-horizon burden by letting a high-level VLM agent perform temporal reasoning and recovery while each specialized VLA tool handles a bounded physical subtask, with coupling achieved through a VLA tool-family interface that supplies explicit tool selection and in-execution progress feedback and with tool specialization obtained via Tool-Aligned Post-Training that builds invocation-aligned data units and residual adapters.

What carries the argument

The VLA tool-family interface that exposes explicit tool selection and progress feedback to enable event-triggered replanning without continuous polling.

If this is right

  • Long-horizon success improves because the planner only reasons at key events instead of generating every low-level action.
  • Invocation fidelity rises when training units are built to match the exact format of agent calls.
  • Efficient replanning occurs whenever a tool reports progress, removing the need for constant agent polling.
  • Specialized tools can be swapped or added without retraining the entire high-level planner.

Where Pith is reading between the lines

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

  • The same split could be tested on tasks that require recovery from external disturbances such as moved objects or human interruptions.
  • Residual adapters might allow new tools to be trained on small datasets collected from real robots rather than simulation.
  • Progress feedback could be extended to include uncertainty estimates so the planner can decide whether to continue or replan earlier.

Load-bearing premise

The specialized VLA tools produced by the interface and post-training will follow high-level invocations reliably enough that they do not introduce new failure modes during closed-loop execution.

What would settle it

A trial on LIBERO-Long or RoboTwin in which the measured Non-biased Rate of tool invocation stays below the reported 15-point gain or where overall task success falls because tool execution errors accumulate faster than the planner can recover.

Figures

Figures reproduced from arXiv: 2605.13119 by Changxing Liu, Minhao Xiong, Siheng Chen, Weixin Li, Yichen Xiong, Yuanzhuo Ding, Zhipeng Zhang, Zixing Lei.

Figure 1
Figure 1. Figure 1: Motivation and overview of VLAs-as-Tools. Classic code-as-policy systems benefit from high-level reasoning, but are limited by manually specified and restricted robot tool sets. End-to-end VLAs provide stronger robot-control capabilities, but make long-horizon execution vulnerable to demonstration biases and recovery failures. Our method combines the strengths of both paradigms: a general VLM agent decompo… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the closed-loop system with the VLAs-as-Tools strategy. A VLM agent [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Main reinforcement-learning results on LIBERO-Long and CALVIN. LIBERO-Long [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of long-horizon failure recovery. Our method prevents cascading [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the automatic labeling pipeline output. Raw robot trajectories are segmented [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
read the original abstract

Vision-language-action (VLA) models are effective robot action executors, but they remain limited on long-horizon tasks due to the dual burden of extended closed-loop planning and diverse physical operations. We therefore propose VLAs-as-Tools, a strategy that distributes this burden across a high-level vision language model (VLM) agent for temporal reasoning and a family of specialized VLA tools for diverse local physical operations. The VLM handles scene analysis, global planning, and recovery, while each VLA tool executes a bounded subtask. To tightly couple agent planning with VLA tool execution in long-horizon tasks, we introduce a VLA tool-family interface that exposes explicit tool selection and in-execution progress feedback, enabling efficient event-triggered agent replanning without continuous agent polling. To obtain diverse specialized VLA tools that faithfully follow agent invocations, we further propose Tool-Aligned Post-Training (TAPT), which constructs invocation-aligned training units for instruction following and adopts tool-family residual adapters for efficient tool specialization. Experiments show that VLAs-as-Tools improves the success rate of $\pi_{0.5}$ by 4.8 points on LIBERO-Long and 23.1 points on RoboTwin, and further enhances invocation fidelity by 15.0 points as measured by Non-biased Rate. Code will be released.

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 VLAs-as-Tools, a framework that decomposes long-horizon robotic tasks by pairing a high-level VLM agent (for scene analysis, global planning, and recovery) with a family of specialized VLA tools (for bounded physical subtasks). It introduces a VLA tool-family interface exposing explicit selection and in-execution progress feedback for event-triggered replanning, plus Tool-Aligned Post-Training (TAPT) that builds invocation-aligned units and uses residual adapters for efficient specialization. Experiments report that the approach raises success rate of π_{0.5} by 4.8 points on LIBERO-Long and 23.1 points on RoboTwin while improving invocation fidelity by 15.0 points on Non-biased Rate.

Significance. If the central claims hold, the modular separation of high-level temporal reasoning from low-level execution, combined with an explicit interface for replanning, offers a practical route to more scalable long-horizon embodied agents. The numeric gains on two established benchmarks and the commitment to release code are concrete strengths that would aid reproducibility and follow-on work.

major comments (2)
  1. [Experiments] Experiments section: the reported gains (4.8/23.1 success-rate points and 15.0 Non-biased Rate) rest on the assumption that the tool-family interface and TAPT produce tools that execute reliably in closed loop without compounding errors from progress feedback or selection mismatches; no ablation isolating interface-driven replanning versus continuous polling is presented, leaving the efficiency claim unsupported.
  2. [Experiments] Experiments section: error propagation across sequential agent-tool calls is not quantified, yet the central claim requires that specialized VLA tools maintain low error rates over long horizons; without such analysis the 23.1-point RoboTwin gain cannot be confidently attributed to the proposed interface rather than isolated tool improvements.
minor comments (2)
  1. [Abstract] Abstract: the baseline π_{0.5} should be briefly characterized (architecture, training data) so readers can assess the magnitude of the reported deltas.
  2. [Abstract] Notation: the term 'Non-biased Rate' is introduced without a formal definition or equation; a short definition or reference to its computation would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point-by-point below and have revised the Experiments section to incorporate additional analyses that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the reported gains (4.8/23.1 success-rate points and 15.0 Non-biased Rate) rest on the assumption that the tool-family interface and TAPT produce tools that execute reliably in closed loop without compounding errors from progress feedback or selection mismatches; no ablation isolating interface-driven replanning versus continuous polling is presented, leaving the efficiency claim unsupported.

    Authors: We agree that an explicit ablation isolating the contribution of event-triggered replanning (via the tool-family interface) versus continuous polling would strengthen the efficiency claims. In the revised manuscript we have added this ablation on both LIBERO-Long and RoboTwin. The new results show that event-triggered replanning reduces VLM invocations by 38-42% while preserving success rates within 1.2 points of the continuous-polling baseline, confirming that the interface enables efficient replanning without introducing compounding errors from progress feedback or selection mismatches. We have also clarified the interface design in Section 3.2 to emphasize how explicit progress signals prevent error accumulation. revision: yes

  2. Referee: [Experiments] Experiments section: error propagation across sequential agent-tool calls is not quantified, yet the central claim requires that specialized VLA tools maintain low error rates over long horizons; without such analysis the 23.1-point RoboTwin gain cannot be confidently attributed to the proposed interface rather than isolated tool improvements.

    Authors: We acknowledge that a direct quantification of error propagation across sequential calls is necessary to attribute gains to the interface rather than solely to tool specialization. We have added a new subsection in the Experiments section that reports per-step success rates and cumulative horizon-wise error on RoboTwin. The analysis shows that the TAPT-specialized tools sustain >84% per-step success even at horizon lengths >15, with the interface-enabled replanning recovering from the remaining failures. This supports that the 23.1-point gain arises from the combined framework rather than isolated tool improvements alone. The revised text now explicitly links these metrics to the central long-horizon claim. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical gains measured on external benchmarks with no self-referential derivations or fitted predictions

full rationale

The paper's central claims rest on empirical success-rate improvements (4.8 points on LIBERO-Long, 23.1 on RoboTwin) and a Non-biased Rate metric, both evaluated against independent external task suites. No equations, fitted parameters, or self-citations are invoked to derive these quantities from the method itself; TAPT and the tool-family interface are presented as engineering choices whose value is demonstrated by closed-loop performance rather than by construction. The derivation chain therefore terminates in observable robot behavior on held-out benchmarks rather than reducing to any input definition or prior self-result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The central claim rests on the assumption that VLMs can reliably perform scene analysis and temporal planning, and that specialized VLA tools can be post-trained to follow invocations faithfully. No free parameters are explicitly listed in the abstract.

axioms (1)
  • domain assumption High-level VLMs can perform effective temporal reasoning, scene analysis, and recovery for long-horizon tasks
    Invoked to justify the VLM agent role.
invented entities (3)
  • VLAs-as-Tools framework no independent evidence
    purpose: Distribute planning and execution burden across VLM agent and VLA tools
    New proposed architecture for long-horizon control.
  • VLA tool-family interface no independent evidence
    purpose: Enable explicit tool selection and in-execution progress feedback
    New interface for coupling agent and tools.
  • Tool-Aligned Post-Training (TAPT) no independent evidence
    purpose: Construct invocation-aligned training units and use residual adapters for tool specialization
    New post-training procedure.

pith-pipeline@v0.9.0 · 5570 in / 1445 out tokens · 33294 ms · 2026-05-14T18:30:08.903263+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

34 extracted references · 34 canonical work pages · 18 internal anchors

  1. [1]

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

    Michael Ahn, Anthony Brohan, Noah Brown, et al. Do as I can, not as I say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Belkhale, T

    Suneel Belkhale, Tianli Ding, Ted Xiao, et al. RT-H: Action hierarchies using language.arXiv preprint arXiv:2403.01823,

    work page arXiv

  3. [3]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    Kevin Black, Noah Brown, Danny Driess, et al. π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

    Anthony Brohan et al. RT-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023a. Anthony Brohan et al. RT-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023b. Kaiyuan Chen, Shuangyu Xie, Zehan Ma, Pannag R Sanketi, and Ken Goldberg. Rob...

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Yiguo Fan, Pengxiang Ding, Shuanghao Bai, Xinyang Tong, Yuyang Zhu, Hongchao Lu, Fengqi Dai, Wei Zhao, Yang Liu, Siteng Huang, Zhaoxin Fan, Badong Chen, and Donglin Wang

    URL https://arxiv.org/abs/2505.15517. Yiguo Fan, Pengxiang Ding, Shuanghao Bai, Xinyang Tong, Yuyang Zhu, Hongchao Lu, Fengqi Dai, Wei Zhao, Yang Liu, Siteng Huang, Zhaoxin Fan, Badong Chen, and Donglin Wang. Long-VLA: Unleashing long-horizon capability of vision language action model for robot manipulation.arXiv preprint arXiv:2508.19958,

    work page arXiv

  6. [6]

    When vision overrides language: Evaluating and mitigating counterfactual failures in VLAs

    Yu Fang, Yuchun Feng, Dong Jing, Jiaqi Liu, Yue Yang, Zhenyu Wei, Daniel Szafir, and Mingyu Ding. When vision overrides language: Evaluating and mitigating counterfactual failures in VLAs. arXiv preprint arXiv:2602.17659,

    work page arXiv

  7. [7]

    LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models

    Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. Libero-plus: In-depth robustness analysis of vision-language-action models.arXiv preprint arXiv:2510.13626,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Hancock, Xindi Wu, Lihan Zha, Olga Russakovsky, and Anirudha Majumdar

    Asher J. Hancock, Xindi Wu, Lihan Zha, Olga Russakovsky, and Anirudha Majumdar. Actions as language: Fine-tuning VLMs into VLAs without catastrophic forgetting.arXiv preprint arXiv:2509.22195,

    work page arXiv

  9. [9]

    LoRA: Low-Rank Adaptation of Large Language Models

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Thinkact: Vision-language-action reasoning via reinforced visual latent planning.arXiv preprint arXiv:2507.16815,

    Chi-Pin Huang, Yueh-Hua Wu, Min-Hung Chen, Yu-Chiang Frank Wang, and Fu-En Yang. ThinkAct: Vision-language-action reasoning via reinforced visual latent planning.arXiv preprint arXiv:2507.16815,

    work page arXiv

  11. [11]

    Inner Monologue: Embodied Reasoning through Planning with Language Models

    doi: 10.1177/02783649231170897. Wenlong Huang, Fei Xia, Ted Xiao, et al. Inner monologue: Embodied reasoning through planning with language models.arXiv preprint arXiv:2207.05608,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1177/02783649231170897

  12. [12]

    BC-Z: Zero-shot task generalization with robotic imitation learning.arXiv preprint arXiv:2202.02005,

    Eric Jang, Alex Irpan, Mohi Khansari, et al. BC-Z: Zero-shot task generalization with robotic imitation learning.arXiv preprint arXiv:2202.02005,

    work page arXiv

  13. [13]

    DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset

    URL https://arxiv. org/abs/2403.12945. 10 Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, et al. OpenVLA: An open-source vision-language- action model.arXiv preprint arXiv:2406.09246,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Code as Policies: Language Model Programs for Embodied Control

    Jacky Liang, Wenlong Huang, Fei Xia, et al. Code as policies: Language model programs for embodied control.arXiv preprint arXiv:2209.07753,

    work page internal anchor Pith review arXiv

  16. [16]

    LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning

    Bo Liu, Yifeng Zhu, Chongkai Gao, et al. LIBERO: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation

    Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Towards long-lived robots: Continual learning VLA models via reinforcement fine-tuning.arXiv preprint arXiv:2602.10503, 2026a

    Yuan Liu, Haoran Li, Shuai Tian, Yuxing Qin, Yuhui Chen, Yupeng Zheng, Yongzhen Huang, and Dongbin Zhao. Towards long-lived robots: Continual learning VLA models via reinforcement fine-tuning.arXiv preprint arXiv:2602.10503, 2026a. Yuanzhe Liu, Jingyuan Zhu, Yuchen Mo, Gen Li, Xu Cao, Jin Jin, Yifan Shen, Zhengyuan Li, Tianjiao Yu, Wenzhen Yuan, Fangqiang...

    work page arXiv

  19. [19]

    What matters in language conditioned robotic imitation learning over unstructured data.arXiv preprint arXiv:2204.06252,

    Oier Mees, Lukas Hermann, and Wolfram Burgard. What matters in language conditioned robotic imitation learning over unstructured data.arXiv preprint arXiv:2204.06252,

    work page arXiv

  20. [20]

    Yao Mu, Tianxing Chen, Zanxin Chen, et al

    doi: 10.3390/e25040657. Yao Mu, Tianxing Chen, Zanxin Chen, et al. RoboTwin: Dual-arm robot benchmark with generative digital twins.arXiv preprint arXiv:2504.13059,

    work page doi:10.3390/e25040657

  21. [21]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    NVIDIA, Johan Bjorck, et al. GR00T N1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Octo: An Open-Source Generalist Robot Policy

    Octo Model Team, Dibya Ghosh, Homer Walke, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Open X-Embodiment: Robotic Learning Datasets and RT-X Models

    Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, et al. Open X-Embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    $\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization

    Physical Intelligence, Kevin Black, Noah Brown, et al. π0.5: a vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054,

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Toolformer: Language Models Can Teach Themselves to Use Tools

    11 Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, et al. Toolformer: Language models can teach themselves to use tools.arXiv preprint arXiv:2302.04761,

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y .K. Li, Y . Wu, et al. DeepSeekMath: Pushing the limits of mathemati- cal reasoning in open language models.arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Hi robot: Open-ended instruction following with hierarchical vision-language-action models,

    Lucy Xiaoyang Shi, Brian Ichter, Michael Equi, et al. Hi Robot: Open-ended instruction following with hierarchical vision-language-action models.arXiv preprint arXiv:2502.19417,

    work page arXiv

  28. [28]

    Zhe Su, Oliver Kroemer, Gerald E

    doi: 10.1007/978-3-319-43488-9_16. Zhe Su, Oliver Kroemer, Gerald E. Loeb, Gaurav S. Sukhatme, and Stefan Schaal. Learning manipulation graphs from demonstrations using multimodal sensory signals. In2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2758–2765. IEEE,

    work page doi:10.1007/978-3-319-43488-9_16

  29. [29]

    Xiaoquan Sun, Zetian Xu, Chen Cao, Zonghe Liu, Yihan Sun, Jingrui Pang, Ruijian Zhang, Zhen Yang, Kang Pang, Dingxin He, Mingqi Yuan, and Jiayu Chen

    doi: 10.1109/ICRA.2018.8461121. Xiaoquan Sun, Zetian Xu, Chen Cao, Zonghe Liu, Yihan Sun, Jingrui Pang, Ruijian Zhang, Zhen Yang, Kang Pang, Dingxin He, Mingqi Yuan, and Jiayu Chen. AtomVLA: Scalable post-training for robotic manipulation via predictive latent world models.arXiv preprint arXiv:2603.08519,

    work page doi:10.1109/icra.2018.8461121 2018

  30. [30]

    DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control

    URLhttps://proceedings.mlr.press/v229/walke23a.html. Junjie Wen, Yichen Zhu, Jinming Li, et al. DexVLA: Vision-language model with plug-in diffusion expert for general robot control.arXiv preprint arXiv:2502.05855,

    work page Pith review arXiv

  31. [31]

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

    John Yang, Carlos E. Jimenez, Alexander Wettig, et al. SWE-agent: Agent-computer interfaces enable automated software engineering.arXiv preprint arXiv:2405.15793,

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    LoHoVLA: A unified vision- language-action model for long-horizon embodied tasks.arXiv preprint arXiv:2506.00411, 2025a

    Yi Yang, Jiaxuan Sun, Siqi Kou, Yihan Wang, and Zhijie Deng. LoHoVLA: A unified vision- language-action model for long-horizon embodied tasks.arXiv preprint arXiv:2506.00411, 2025a. Yue Yang, Shuo Cheng, Yu Fang, Homanga Bharadhwaj, Mingyu Ding, Gedas Bertasius, and Daniel Szafir. LiLo-VLA: Compositional long-horizon manipulation via linked object-centric...

    work page arXiv

  33. [33]

    Agentic robot: A brain-inspired framework for vision-language-action models in embodied agents,

    Zhejian Yang, Yongchao Chen, Xueyang Zhou, Jiangyue Yan, Dingjie Song, Yinuo Liu, Yuting Li, Yu Zhang, Pan Zhou, Hechang Chen, and Lichao Sun. Agentic robot: A brain-inspired framework for vision-language-action models in embodied agents.arXiv preprint arXiv:2505.23450, 2025b. Shunyu Yao, Jeffrey Zhao, Dian Yu, et al. ReAct: Synergizing reasoning and acti...

    work page arXiv

  34. [34]

    Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, Hang Xu, Hefeng Wu, Liang Lin, and Xiaodan Liang

    doi: 10.1177/0278364919835020. Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, Hang Xu, Hefeng Wu, Liang Lin, and Xiaodan Liang. AtomicVLA: Unlocking the potential of atomic skill learning in robots.arXiv preprint arXiv:2603.07648, 2026a. Ninghao Zhang, Bin Zhu, Shijie Zhou, and Jingjing Chen. Restoring l...

    work page doi:10.1177/0278364919835020 2023