Pith. sign in

REVIEW 4 major objections 5 minor 25 references

UniETP claims that unifying four embodied simulators under one observation/action interface and an automatic task generator yields planners that generalize to unseen simulators and task templates.

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 16:12 UTC pith:RIODKMBE

load-bearing objection A genuinely unifying benchmark and task-generation pipeline with a credible SFT transfer result — but the scripted-planner feasibility filter is underspecified and needs to be opened up before the numbers can be trusted. the 4 major comments →

arxiv 2607.18062 v1 pith:RIODKMBE submitted 2026-07-20 cs.RO cs.CV

UniETP: Unifying Environments for Generalizable Embodied Task Planning

Peiran Xu , Jiaqi Zheng , Ziyou Wang , Yadong Mu This is my paper
classification cs.RO cs.CV
keywords embodied task planningunified benchmarksimulator unificationscene graphautomatic task generationvision-language modelsgeneralizationevaluation modes
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.

Embodied task planning research is fragmented: each simulator brings its own observations, actions, and evaluation, so results do not transfer and training data does not scale. UniETP is a unified benchmark that wraps four simulators (AI2-THOR, VirtualHome, Habitat, BEHAVIOR) in a single observation and action interface, and uses a shared scene-graph representation with a logic-based evaluator. It also provides an automatic pipeline that turns 138 task templates into 536 benchmark tasks and a 9,000-task training set across five capability splits. The paper's central claim is that this unification pays off: supervised fine-tuning on the 9,000-task set makes a small 4B vision-language model outperform far larger zero-shot models, and the gains persist when the training data excludes one simulator or one third of the task templates. If this is right, the field gains a common substrate for training and comparing embodied planners instead of a scatter of incompatible environments.

Core claim

The paper claims that a single standardized environment interface, combined with a logic-based evaluation over a unified scene graph and an automatic task generation pipeline, is sufficient to train embodied task planners that generalize across simulators and task templates. Concretely, UniETP unifies AI2-THOR, VirtualHome, Habitat, and BEHAVIOR, and the authors show that supervised fine-tuning of Qwen3.5-4B on its 9,000-task training set yields a planner that beats much larger zero-shot models, and that holding out an entire simulator (AI2-THOR) or about a third of the task templates in training still leaves clear improvements over the untrained base model.

What carries the argument

The load-bearing representation is the Unified Scene Graph (USG), a simulator-agnostic graph whose nodes are rooms and objects (each with ID, category, intrinsic properties, current state, and location) and whose edges encode spatial relations, updated after every executed action. A logic system over USG traces — first-order predicates with Boolean connectives, quantifiers, temporal operators, and counting aggregates — turns any task into an unambiguous goal formula, so evaluation is consistent across very different simulators. Projection functions φ_i and ψ_i translate each simulator's raw observations and actions into the unified observation and action space, letting one agent policy be tr

Load-bearing premise

The benchmark retains only tasks that a hand-written scripted planner can solve, so the diversity, difficulty ratings, and cross-simulator comparisons are all constrained by that planner's coverage of the task space.

What would settle it

Take a simulator whose native action semantics do not fit UniETP's 30+ unified action categories, fine-tune a model on the 9,000-task training set, and measure success on that simulator; if success stays near the untrained baseline, the claimed generalization is limited to simulators whose interfaces the authors happened to align.

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

If this is right

  • A single benchmark can rank embodied planners across four simulators under identical observation and action formats, removing simulator-specific tuning as a confound.
  • The automatic pipeline converts 138 templates into validated, grounded task instances, making large training sets cheap to produce and easy to expand to new scenes or templates.
  • The M1/M2/M3 evaluation modes separate planning from perception and exploration, so a model's failure can be attributed to reasoning, object grounding, or search rather than lumped together.
  • Fine-tuning on the unified interface transfers to held-out simulators and held-out task templates, suggesting cross-simulator training is a practical route to generalist embodied planners.
  • The logic-based goal evaluator supports temporal, quantified, and counting constraints, allowing capabilities to be tested that simple goal-state success cannot express.

Where Pith is reading between the lines

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

  • Because the pipeline keeps only tasks a hand-written scripted planner can complete, the benchmark may under-sample search-heavy or unusual-tool tasks; removing that filter would likely change which splits look hardest.
  • The M1→M3 difficulty gradient offers a natural training curriculum — start with full scene priors and composite actions, then progressively strip priors and force exploration — which the paper's results suggest could improve sample efficiency.
  • Instance Grounding being the hardest split points to fine-grained visual discrimination, not language understanding or high-level reasoning, as the current limiting factor for VLM-based planners.
  • Since a new simulator can be added by writing projector functions, the same interface and evaluation logic could extend to future simulators or real robots, turning UniETP into a living benchmark.

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

4 major / 5 minor

Summary. The paper proposes UniETP, a benchmark and environment layer that unifies four embodied simulators (AI2-THOR, VirtualHome, Habitat, BEHAVIOR) through a common observation space, a parameterized action space, a Unified Scene Graph (USG), and a logic-based evaluation system. On top of this layer, the authors construct an automatic task-generation pipeline that combines manually designed templates, per-simulator Affordance Knowledge Bases (AKBs), an LLM annotator, and a scripted expert policy for feasibility filtering. This yields 536 benchmark tasks across 138 templates and a 9k-task training set. The paper evaluates a range of VLMs under three evaluation modes (M1, M2, M3), analyzes performance by task split and simulator, ablates feedback signals, and shows that SFT of Qwen3.5-4B on the generated training set improves performance and appears to transfer to held-out simulators and held-out task templates.

Significance. If the construction is sound, UniETP is a potentially valuable community resource. Unifying four heterogeneous simulators under one interface, with multi-granularity observation/action modes and a USG-based logic evaluator, directly addresses a real fragmentation problem in embodied task planning. The 9k-task automatic generation pipeline and the cross-simulator SFT experiment are also substantive contributions, and the paper is careful to separate evaluation modes by difficulty. The main risk is that the task-generation pipeline is filtered by an underspecified scripted planner, so the benchmark's diversity, difficulty profile, and generalization results may reflect the planner's coverage rather than the space of natural embodied tasks. The absence of variance information further weakens quantitative comparisons.

major comments (4)
  1. [§3.3.1] The scripted expert policy π* is the sole feasibility gate for both the 536 benchmark tasks and the 9k SFT tasks, but it is never described. The manuscript does not state whether π* uses privileged USG information, what planning/search algorithm it runs, its success rate per simulator or template, or how many generated candidates are rejected. This is load-bearing: if π* is conservative or has simulator-specific blind spots, the claimed diversity (Fig. 4), difficulty profile (Table 5), and the transfer results (Table 7) are all properties of the filter, not of the underlying task distribution. Please specify π* completely, report rejection statistics per simulator/template, and provide a robustness check such as regenerating tasks with an independent planner or manually validating a random sample of rejected/accepted instances.
  2. [§4.1, Tables 3–7] All success rates are reported as point estimates with no repeated trials, variance, or confidence intervals. Since the evaluated VLMs are stochastic and the tasks involve closed-loop interaction, differences such as 72 vs. 73 in Table 3, or the SFT gains in Table 7, cannot be distinguished from noise. Please run multiple seeds/trials per condition and report mean and variance (or per-task pass/fail data). This is necessary to support the comparative claims about model rankings, simulator difficulty, and SFT improvement.
  3. [§3.2.2, M3 bbox matching] The M3 protocol says that model-generated bboxes are matched to visible objects by intersection-over-union, but no IoU threshold, tie-breaking rule, or behavior for unmatched boxes is specified. Without this, the M3 evaluation is not reproducible and the M3 results in Table 4 cannot be interpreted. Please state the exact matching rule, including how multiple visible objects of the same category are handled and what happens when no object meets the threshold. Also note that Table 4 covers only the Atomic split, so the broader conclusions about M3 difficulty are limited.
  4. [§4.4, Table 7] The 'unseen tasks' in Table 7 are generated by the same template/AKB/LLM/scripted-planner pipeline as the training tasks. This is a mild circularity: transfer to held-out templates demonstrates generalization within the template manifold, but not generalization to tasks outside that manifold. Please define precisely what 'Seen/Unseen' means for each row, give the exact template split proportions, and report per-template performance on the held-out templates. In particular, clarify the row 'SFT (w.o. unseen tasks)' — it is not clear what is excluded or how it differs from 'SFT (full dataset)'.
minor comments (5)
  1. [Abstract/§1] Typo: 'serveral' should be 'several'. Also Table 1 contains a duplicated parenthesis in 'BEHA VIOR))'.
  2. [Table 3] Many numeric entries appear concatenated (e.g., 'Gemini-er-1.6 9386 87 5073 76 45 23 33 30 24 29'), making the table hard to read. Please ensure proper column alignment in the camera-ready version.
  3. [Table 4] The table is missing a column header for model names and would benefit from a note that only the Atomic split is evaluated under M3.
  4. [Figure 4] The bar labels in panel (a) and the legend for splits are not legible in the current rendering; please enlarge and clarify.
  5. [§4.1] The ReAct-style inference protocol mentions 'one or more actions' per call, but it is not specified how multiple actions are parsed or whether an invalid multi-action output is discarded. Please state the parsing/fallback rule.

Circularity Check

0 steps flagged

No significant circularity: the benchmark results and SFT transfer study are externally evaluated on held-out tasks; the scripted-planner filter is a data-curation gate, not a fitted predictor.

full rationale

UniETP's central contributions are engineering (standardized observation/action interface, unified scene graph evaluation) and empirical benchmarking of external VLMs plus a supervised fine-tuning study. I find no load-bearing step that reduces to its own inputs by construction. The SFT experiment in Section 4.4 trains Qwen3.5-4B on a 9k-task training set and tests on the 536-task benchmark with non-overlapping scenes; the 'unseen simulator' and 'unseen task template' ablations explicitly hold out entire simulators or about one-third of templates, so test performance is a genuine held-out evaluation rather than a fitted prediction. The scripted-expert filter in Section 3.3.1 ('we employ a scripted expert policy π* to attempt task planning and retain only the tasks that can be successfully completed') is a feasibility/curation gate, not a fitted parameter or a prediction target; it may bias task coverage and difficulty, and the paper omits rejection statistics, but it does not make the VLM success rates or the Table 7 generalization deltas equivalent to the filter by construction. The same template/LLM/scripted-planner pipeline generating both training and benchmark tasks limits the external-validity claim to UniETP's generated distribution, but that is a scope limitation, not circularity. The only self-citations (Xu et al. 2026, Zhao et al. 2025) appear in related-work lists and are not load-bearing. Per the review rules, this is a non-circular finding: score 0, no circular steps.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claims are engineering/data claims rather than derivations, so the ledger is dominated by design choices and domain assumptions rather than fitted constants. The free parameters are the hand-set evaluation limits, mode configurations, an unreported M3 IoU threshold, and the manually chosen task distribution; the axioms concern simulator-state fidelity, action-semantics equivalence, scripted-planner filtering, and representativeness of the four simulators.

free parameters (4)
  • Episode limits H and failed-action cap H' = H=50/150, H'=10/30 for M1/M2/M3
    Hand-chosen cutoffs; they directly bound how far an agent can explore and therefore shape every success rate in Tables 3-5.
  • M3 bbox-IoU matching threshold = not reported
    Section 3.2.2 says actions in M3 match model bboxes to visible objects by IoU but never gives the acceptance threshold; a different threshold would change success rates.
  • Evaluation-mode configuration M1/M2/M3 = hand-selected
    Which global-information levels, navigation granularities, and composite actions are enabled is a design choice, so cross-mode difficulty comparisons are relative to these choices, not intrinsic.
  • Task-template and split distribution = 138 templates, 536 benchmark tasks; manual split sizes
    Aggregate average scores depend on the manually chosen mix of Atomic/Compositional/Logical/Linguistic/Instance-Grounding tasks and per-simulator task counts; no sampling error or sensitivity analysis is reported.
axioms (5)
  • standard math Standard POMDP formulation and first-order/temporal logic goal semantics.
    Section 3.1 models ETP as a POMDP and Section 3.2.1's logic system assumes standard semantics for quantifiers, temporal operators, and counting over USG traces; the paper does not derive these.
  • domain assumption The Unified Scene Graph built from simulator privileged information is an accurate state representation for evaluation.
    USG is the sole basis for success evaluation; if simulator metadata is incomplete or stale, reported success rates are invalid.
  • domain assumption The 30+ merged action categories preserve equivalent semantics across all four simulators.
    Section 3.2.1 merges actions by 'equivalent semantics'; cross-simulator comparability depends on the projectors ψ_i mapping unified actions to faithful simulator actions.
  • domain assumption Retaining only tasks solved by the scripted expert policy does not bias the benchmark.
    Section 3.3.1 discards tasks the scripted planner fails; the benchmark's complexity and diversity are therefore bounded by that planner's capabilities.
  • domain assumption The four chosen simulators are a representative basis for claims about generalizable embodied planning.
    The paper's generalization conclusions are drawn from this specific set of simulators and tasks; no argument establishes coverage of the broader space of embodied environments.
invented entities (2)
  • Unified Scene Graph (USG) no independent evidence
    purpose: A graph over rooms, objects, states, and spatial relations, used as the simulator-agnostic substrate for evaluation logic and data generation.
    This is a software abstraction introduced by the paper. It makes the evaluation and generation pipeline concrete, but it makes no falsifiable prediction outside the benchmark.
  • Affordance Knowledge Base (AKB) no independent evidence
    purpose: Per-simulator table mapping object categories to states, functionalities, and feasible spatial relations, used to instantiate task templates.
    Compiled from simulator documentation plus LLM queries; accuracy is not independently validated and could propagate systematic errors into generated tasks.

pith-pipeline@v1.3.0-alltime-deepseek · 16799 in / 16397 out tokens · 154883 ms · 2026-08-01T16:12:21.218532+00:00 · methodology

0 comments
read the original abstract

This paper focuses on the problem of Embodied Task Planning, where an agent is required to execute a sequence of atomic actions within an interactive environment to complete a user-specified task. Though a variety of simulators and datasets have previously been built for this task, these efforts are largely isolated, with each using its own observation format, action type, and task domain. This fragmentation complicates comprehensive model evaluation and hinders the scalability of training data. As an effort towards generalizable embodied planning, we propose UniETP, a unified interface integrating four commonly-used simulators (AI2-THOR, VirtualHome, Habitat, BEHAVIOR). UniETP is characterized by both standardization and diversity. On one hand, it formalizes all the simulators into a consistent observation and action space, and builds an evaluation system to support complicated task goal. On the other hand, it enhances task diversity and complexity across dimensions like task logic, instance grounding, and instruction understanding, constructing a new dataset with varied levels of difficulty in an automatic manner. Extensive experiments on the proposed benchmark are conducted to evaluate the embodied planning capabilities of recent models and analyze the performance bottlenecks. Codes and data will be available at https://github.com/woyut/UniETP .

Figures

Figures reproduced from arXiv: 2607.18062 by Jiaqi Zheng, Peiran Xu, Yadong Mu, Ziyou Wang.

Figure 1
Figure 1. Figure 1: We propose UniETP, an embodied task planning benchmark consisting of an environment [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Upper: the ETP inference pipeline in UniETP’s environment layer. Lower: the task [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example tasks on different splits. 3.3.2 EVALUATION DIMENSIONS When constructing the benchmark tasks, we design several capability-oriented evaluation splits to examine different aspects of ETP. (1) Atomic. It contains the most basic task types with a single predicate as goal formula, evaluating the understanding of basic object states and spatial relations. (2) Compositional. It consists of the conjunctio… view at source ↗
Figure 4
Figure 4. Figure 4: Key statistics of the UniETP benchmark. (a) Numbers of tasks across different simulators [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The distribution of number of steps under mode M1 (left) and M2 (right). [PITH_FULL_IMAGE:figures/full_fig_p009_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.

Reference graph

Works this paper leans on

25 extracted references · 12 linked inside Pith

  1. [1]

    Gemini robotics 1.5: Pushing the frontier of generalist robots with advanced embodied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03342,

    Abbas Abdolmaleki, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gon- zalez Arenas, Ashwin Balakrishna, Nathan Batchelor, Alex Bewley, Jeff Bingham, Michael Bloesch, et al. Gemini robotics 1.5: Pushing the frontier of generalist robots with advanced embodied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03342,

    Pith/arXiv arXiv

  2. [3]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    Pith/arXiv arXiv

  3. [4]

    Cookbench: A long-horizon embodied planning benchmark for complex cooking scenarios.arXiv preprint arXiv:2508.03232,

    Muzhen Cai, Xiubo Chen, Yining An, Jiaxin Zhang, Xuesong Wang, Wang Xu, Weinan Zhang, and Ting Liu. Cookbench: A long-horizon embodied planning benchmark for complex cooking scenarios.arXiv preprint arXiv:2508.03232,

    Pith/arXiv arXiv

  4. [5]

    Partnr: A benchmark for planning and reasoning in embodied multi-agent tasks

    Matthew Chang, Gunjan Chhablani, Alexander Clegg, Mikael Dallaire Cote, Ruta Desai, Michal Hlavac, Vladimir Karashchuk, Jacob Krantz, Roozbeh Mottaghi, Priyam Parashar, et al. Partnr: A benchmark for planning and reasoning in embodied multi-agent tasks. InInternational Con- ference on Learning Representations, volume 2025, pp. 65205–65268,

    2025

  5. [8]

    Ace-brain-0: Spatial intelligence as a shared scaffold for universal embodiments.arXiv preprint arXiv:2603.03198,

    Ziyang Gong, Zehang Luo, Anke Tang, Zhe Liu, Shi Fu, Zhi Hou, Ganlin Yang, Weiyun Wang, Xiaofeng Wang, Jianbo Liu, et al. Ace-brain-0: Spatial intelligence as a shared scaffold for universal embodiments.arXiv preprint arXiv:2603.03198,

    arXiv

  6. [9]

    Gemini Robotics-ER 1.6 Model Card.https://deepmind.google/ models/model-cards/gemini-robotics-er-1-6/, April 2026a

    Google DeepMind. Gemini Robotics-ER 1.6 Model Card.https://deepmind.google/ models/model-cards/gemini-robotics-er-1-6/, April 2026a. Google DeepMind. Gemma 4 Model Card.https://ai.google.dev/gemma/docs/ core/model_card_4, April 2026b. Xiaoshuai Hao, Lei Zhou, Zhijian Huang, Zhiwen Hou, Yingbo Tang, Lingfeng Zhang, Guang Li, Zheng Lu, Shuhuai Ren, Xianhui ...

    Pith/arXiv arXiv

  7. [10]

    Ai2-thor: An interactive 3d environment for visual ai.arXiv preprint arXiv:1712.05474,

    Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai.arXiv preprint arXiv:1712.05474,

    Pith/arXiv arXiv

  8. [13]

    Is-bench: Evaluating interactive safety of vlm-driven embodied agents in daily household tasks.arXiv preprint arXiv:2506.16402,

    Xiaoya Lu, Zeren Chen, Xuhao Hu, Yijin Zhou, Weichen Zhang, Dongrui Liu, Lu Sheng, and Jing Shao. Is-bench: Evaluating interactive safety of vlm-driven embodied agents in daily household tasks.arXiv preprint arXiv:2506.16402,

    arXiv

  9. [16]

    Bumble: Unifying reasoning and acting with vision-language models for building-wide mobile manipulation

    Rutav Shah, Albert Yu, Yifeng Zhu, Yuke Zhu, and Roberto Mart ´ın-Mart´ın. Bumble: Unifying reasoning and acting with vision-language models for building-wide mobile manipulation. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 13337–13345. IEEE,

    2025

  10. [17]

    World- aware planning narratives enhance large vision-language model planner.arXiv preprint arXiv:2506.21230, 2025a

    Junhao Shi, Zhaoye Fei, Siyin Wang, Qipeng Guo, Jingjing Gong, and Xipeng QIu. World- aware planning narratives enhance large vision-language model planner.arXiv preprint arXiv:2506.21230, 2025a. Lucy Xiaoyang Shi, Brian Ichter, Michael Equi, Liyiming Ke, Karl Pertsch, Quan Vuong, James Tanner, Anna Walling, Haohuan Wang, Niccolo Fusai, et al. Hi robot: O...

    Pith/arXiv arXiv

  11. [18]

    Robobrain 2.0 technical report.arXiv preprint arXiv:2507.02029,

    14 UniETP BAAI RoboBrain Team, Mingyu Cao, Huajie Tan, Yuheng Ji, Xiansheng Chen, Minglan Lin, Zhiyu Li, Zhou Cao, Pengwei Wang, Enshen Zhou, et al. Robobrain 2.0 technical report.arXiv preprint arXiv:2507.02029,

    Pith/arXiv arXiv

  12. [19]

    Seea-r1: Tree-structured reinforcement fine-tuning for self-evolving embodied agents.arXiv preprint arXiv:2506.21669,

    Wanxin Tian, Shijie Zhang, Kevin Zhang, Xiaowei Chi, Yulin Luo, Junyu Lu, Chunkai Fan, Qiang Zhou, Yiming Zhao, Ning Liu Siyu Lin, et al. Seea-r1: Tree-structured reinforcement fine-tuning for self-evolving embodied agents.arXiv preprint arXiv:2506.21669,

    arXiv

  13. [20]

    Grutopia: Dream general robots in a city at scale.arXiv preprint arXiv:2407.10943,

    Hanqing Wang, Jiahe Chen, Wensi Huang, Qingwei Ben, Tai Wang, Boyu Mi, Tao Huang, Siheng Zhao, Yilun Chen, Sizhe Yang, et al. Grutopia: Dream general robots in a city at scale.arXiv preprint arXiv:2407.10943,

    Pith/arXiv arXiv

  14. [21]

    World modeling makes a better planner: Dual preference optimization for embodied task plan- ning.arXiv preprint arXiv:2503.10480, 2025a

    Siyin Wang, Zhaoye Fei, Qinyuan Cheng, Shiduo Zhang, Panpan Cai, Jinlan Fu, and Xipeng Qiu. World modeling makes a better planner: Dual preference optimization for embodied task plan- ning.arXiv preprint arXiv:2503.10480, 2025a. Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et...

    Pith/arXiv arXiv

  15. [22]

    Vlaser: Vision-language-action model with synergistic embodied reasoning.arXiv preprint arXiv:2510.11027, 2025a

    Ganlin Yang, Tianyi Zhang, Haoran Hao, Weiyun Wang, Yibin Liu, Dehui Wang, Guanzhou Chen, Zijian Cai, Junting Chen, Weijie Su, et al. Vlaser: Vision-language-action model with synergistic embodied reasoning.arXiv preprint arXiv:2510.11027, 2025a. Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Haoran Tan, Chencheng Jiang, Jiamu Kang, Yuanhan Zhan...

    arXiv

  16. [23]

    Safeagentbench: A benchmark for safe task planning of embodied llm agents.arXiv preprint arXiv:2412.13178,

    Sheng Yin, Xianghe Pang, Yuanzhuo Ding, Menglan Chen, Yutong Bi, Yichen Xiong, Wenhao Huang, Zhen Xiang, Jing Shao, and Siheng Chen. Safeagentbench: A benchmark for safe task planning of embodied llm agents.arXiv preprint arXiv:2412.13178,

    arXiv

  17. [24]

    Embodied navigation foundation model.arXiv preprint arXiv:2509.12129, 2025a

    Jiazhao Zhang, Anqi Li, Yunpeng Qi, Minghan Li, Jiahang Liu, Shaoan Wang, Haoran Liu, Gengze Zhou, Yuze Wu, Xingxing Li, et al. Embodied navigation foundation model.arXiv preprint arXiv:2509.12129, 2025a. Lingfeng Zhang, Yuening Wang, Hongjian Gu, Atia Hamidizadeh, Zhanguang Zhang, Yuecheng Liu, Yutong Wang, David Gamaliel Arcos Bravo, Junyi Dong, Shunbo ...

    arXiv

  18. [25]

    Embodiedbrain: Expanding performance boundaries of task planning for embodied intelligence.arXiv preprint arXiv:2510.20578,

    Ding Zou, Feifan Wang, Mengyu Ge, Siyuan Fan, Zongbing Zhang, Wei Chen, Lingfeng Wang, Zhongyou Hu, Wenrui Yan, Zhengwei Gao, et al. Embodiedbrain: Expanding performance boundaries of task planning for embodied intelligence.arXiv preprint arXiv:2510.20578,

    arXiv

  19. [2017]

    Behavior-1k: A human- centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation.arXiv preprint arXiv:2403.09227, 2024a

    12 UniETP Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Mart ´ın- Mart´ın, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, et al. Behavior-1k: A human- centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation.arXiv preprint arXiv:2403.09227, 2024a. Manling Li, Shiyu Zhao, Qineng Wang...

    Pith/arXiv arXiv

  20. [2019]

    Jie Liu, Pan Zhou, Yingjun Du, Ah-Hwee Tan, Cees G. M. Snoek, Jan-Jakob Sonke, and Efstratios Gavves. Capo: Cooperative plan optimization for efficient embodied multi-agent cooperation. In The Thirteenth International Conference on Learning Representations, 2025a. Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Song XiXuan, Yifan Xu, Shudan Zhang, Hanyu La...

    Pith/arXiv arXiv

  21. [2021]

    Habitat 3.0: A co-habitat for humans, avatars, and robots

    Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Part- sey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. InInternational Conference on Learning Representations, vol- ume 2024, pp. 15306–15336,

    2024

  22. [2023]

    Vestabench: An embodied benchmark for safe long-horizon planning under multi-constraint and adversarial settings

    Tanmana Sadhu, Yanan Chen, and Ali Pesaranghader. Vestabench: An embodied benchmark for safe long-horizon planning under multi-constraint and adversarial settings. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pp. 2122–2145,

    2025

  23. [2024]

    Ronghao Dang, Jiayan Guo, Bohan Hou, Sicong Leng, Kehan Li, Xin Li, Jiangpin Liu, Yunxuan Mao, Zhikai Wang, Yuqian Yuan, et al

    URLhttps://openreview.net/forum? id=ADSxCpCu9s. Ronghao Dang, Jiayan Guo, Bohan Hou, Sicong Leng, Kehan Li, Xin Li, Jiangpin Liu, Yunxuan Mao, Zhikai Wang, Yuqian Yuan, et al. Rynnbrain: Open embodied foundation models.arXiv preprint arXiv:2602.14979,

    arXiv

  24. [2025]

    Cosmos-reason1: From physical common sense to embodied reasoning.arXiv preprint arXiv:2503.15558,

    Alisson Azzolini, Junjie Bai, Hannah Brandon, Jiaxin Cao, Prithvijit Chattopadhyay, Huayu Chen, Jinju Chu, Yin Cui, Jenna Diamond, Yifan Ding, et al. Cosmos-reason1: From physical common sense to embodied reasoning.arXiv preprint arXiv:2503.15558,

    Pith/arXiv arXiv

  25. [2026]

    Homesafebench: A benchmark for embodied vision-language models in free-exploration home safety inspection

    Siyuan Gao, Jiashu Yao, Haoyu Wen, Yuhang Guo, Zeming Liu, and Heyan Huang. Homesafebench: A benchmark for embodied vision-language models in free-exploration home safety inspection. arXiv preprint arXiv:2509.23690,

    arXiv