LLM-Enhanced Rapid-Reflex Async-Reflect Embodied Agent for Real-Time Decision-Making in Dynamically Changing Environments

Dingxin Zhang; Shunqi Mao; Weidong Cai; Yangqing Zheng

REVIEW 3 major objections 4 minor 20 references

LLM-Enhanced Rapid-Reflex Async-Reflect Embodied Agent for Real-Time Decision-Making in Dynamically Changing Environments

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper argues that decision latency, not just decision quality, determines success in dynamic disaster-response environments, and that an agent combining instant rule-based reflexes with asynchronous LLM reflection outperforms both…

desk verdict Worth engaging for the latency-aware evaluation protocol, but the headline performance gain rests on not charging the async reflector's wall-clock time. read the letter →

arxiv 2506.07223 v1 pith:F77O63BV submitted 2025-06-08 cs.AI

Yangqing Zheng , Shunqi Mao , Dingxin Zhang , Weidong Cai This is my paper

classification cs.AI

keywords embodiedagentslargelanguagemodelsreal-timedecision-makinglatency-awareevaluationtimeconversionmechanismHAZARDbenchmarkasynchronousreflectiondynamicenvironments

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that in fast-changing environments, the time an agent spends thinking is part of the problem: a correct decision that arrives late can be worse than a quick approximate one. It introduces a Time Conversion Mechanism that turns inference seconds into simulator frames, so that slow deliberation is penalized in the same units as environment motion. Using the HAZARD fire scenario, it reports that LLM-based and MCTS agents score poorly once that penalty is applied, while simple rule-based policies hold up. It then presents RRARA, which executes a low-latency rule-based action immediately and lets an LLM reflect asynchronously on whether to interrupt, and reports that this hybrid achieves the best value rate and lowest damage ratio in the comparison.

What carries the argument

The Time Conversion Mechanism (TCM) is the accounting device: it maps an inference delay $T_{\text{inf}}$ in seconds to $T_{\text{inf}} \times \text{FPS}$ simulator frames, placing cognitive cost and physical change on one FPS-based scale. The Rapid-Reflex Async-Reflect Agent (RRARA) is the control device: a rule-based policy fires an immediate reflexive action, while an asynchronous LLM Reflector receives observations and current/prior actions and either validates or interrupts the reflex. Together they carry the argument: TCM makes latency visible in the score, and the reflex/reflect split lets the agent spend the LLM's thinking time in the background rather than in the decision loop.

What would settle it

Measure the real wall-clock duration of each GPT-3.5 reflector call during a HAZARD fire run, charge the corresponding $T_{\text{inf}} \times 30$ frames to RRARA's episode before recomputing VR and DR, and check whether RRARA still beats the plain Greedy baseline.

Watch

Extended reading notes

Core claim

The central discovery is that when inference latency is converted into simulation time, the usual ranking of embodied agents reverses. Under the proposed Time Conversion Mechanism, seconds of decision delay become $T_{\text{inf}} \times \text{FPS}$ frames, so a planner that thinks for 2–4 seconds is charged with losing 60–120 frames at 30 FPS while the disaster continues to unfold. In this latency-aware evaluation, the paper finds that MCTS and LLM-based agents fall to value rates of 0.03–0.20, the rule-based Greedy policy holds at 0.22, and RRARA—an immediate Greedy reflex plus an asynchronous GPT-3.5 reflector—reaches 0.29 value rate with a 0.23 damage ratio. The paper states that the LLM reflector intervenes in roughly 60% of action steps without incurring critical latency, and interprets this as evidence that selective asynchronous reflection can improve a reactive policy without sacrificing responsiveness.

Load-bearing premise

The asynchronous LLM reflector is assumed to add zero effective decision latency, so RRARA's RL and LAR are reported as 0.00 even though a GPT-3.5 call takes seconds; if those seconds were converted to frames under TCM, RRARA's advantage could shrink or disappear.

Editorial extensions

If this is right

If inference time is charged as simulation frames, benchmarks that ignore latency will systematically overrate slow deliberative agents.
Latency-aware evaluation will push embodied LLM systems toward caching, parallel reasoning, distilled fast policies, or selective LLM overrides.
RRARA's roughly 60% intervention rate indicates that a planner can improve a reactive controller by editing its decisions, not by replacing it.
Reporting Respond Latency and Latency-to-Action Ratio makes hardware-dependent reasoning costs explicit, so fair comparison requires fixing hardware.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A fair test of the zero-latency assumption would charge the reflector's actual wall-clock response time to RRARA under TCM; the current 0.00 RL/LAR report treats background inference as free.
The reflex-then-reflect pattern could generalize to robot navigation and online control, but its value will hinge on whether the planner's advice arrives before the environment moves on.
A direct test of TCM would inject artificial delays of 0, 1, 2, and 4 seconds into one model and check that the score degrades monotonically as predicted.
The evidence currently covers only the HAZARD fire scenario; flood and wind scenarios may reward different reflex policies, so the claimed advantage is scenario-specific until tested.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Dear colleague,

Short version: the paper contributes a latency-aware evaluation protocol for HAZARD and a reflex/reflector hybrid agent. The evaluation idea is genuinely useful and easy to adopt; the agent is a sensible, if unsurprising, fast-and-slow combination. But the headline empirical result depends on an accounting choice that the paper never makes explicit: the async LLM reflector's 2.35 s wall-clock time is apparently charged nowhere, so TCM stops measuring the deliberation that drives roughly 60% of action switches.

What is new: TCM (F_inf = T_inf * FPS) is direct, but turning it into RL and LAR metrics and applying them to HAZARD is a real step forward. The paper is right that LLM baselines get unfairly graded when inference time is ignored. The RRARA design, a rule-based reflex that acts immediately while an LLM reflects asynchronously and interrupts when needed, is known in spirit, but the HAZARD instantiation is a legitimate test. Credit is also due for comparing against external baselines from HAZARD, not just self-constructed ones.

Soft spots: first, RL=0/LAR=0 for RRARA is not credible without a rule for charging the reflector's inference time and the aborted reflex frames. The GPT-3.5 baseline has RL=2.35 s, and the evaluator intervenes in ~60% of steps; if that time is free, TCM has a hole. This is load-bearing. Second, the empirical support is thin: one table, one scenario, no seeds, no error bars, no episode counts, no statistical tests. The VR/DR gaps could be noise. Third, the circularity concern is real but mild: the comparison is still against the same baselines under the same metric, so the bias is in the metric, not in cherry-picking. Still, it needs transparency. Fourth, no code or data are shipped, which would let a reader check the latency accounting.

Net: the central idea is right and the paper deserves engagement, but the current write-up does not support the claim of substantial improvement. A serious referee should ask for a precise TCM accounting rule for asynchronous reflection, plus multiple seeds and episode counts.

Referee Report

3 major / 4 minor

Summary. The paper proposes a Time Conversion Mechanism (TCM) that converts agent inference delays into equivalent simulation frames (F_inf = T_inf * FPS), and two metrics, Respond Latency (RL) and Latency-to-Action Ratio (LAR), for latency-aware evaluation in the HAZARD benchmark. It then introduces RRARA, a training-free agent that couples a low-latency rule-based reflex policy with an asynchronous GPT-3.5 'Reflector' that can interrupt and revise the reflex's actions. Experiments on the HAZARD fire scenario report that RRARA (Greedy) achieves VR 0.29 and DR 0.23, compared to Greedy's 0.22 and 0.24, and the authors claim that RRARA substantially outperforms existing baselines in latency-sensitive scenarios. The manuscript also states that the reflector intervenes in roughly 60% of action steps without incurring critical latency.

Significance. The TCM idea—penalizing inference delay by converting it into environment frames—is a sensible and potentially useful complement to existing embodied-agent benchmarks, and the reflex-async-reflect architecture is an interesting direction for latency-sensitive decision making. However, the current manuscript does not validate these contributions: the empirical evidence is a single table with no error bars, seeds, or statistical tests, and the reported zero latency for RRARA is not justified under the paper's own TCM framework because the asynchronous LLM must consume wall-clock time. If the evaluation were corrected and extended, the paper could make a meaningful contribution; as it stands, the central performance claim is unsupported.

major comments (3)

[Section 4, Table 1] The reported RL=0.00 and LAR=0.00 for RRARA are inconsistent with the use of GPT-3.5 as the Reflector, since the same table lists GPT-3.5 with RL=2.35 s. The manuscript states that the reflector 'perpetually evaluates action outcomes' and intervenes in roughly 60% of action steps, so the LLM's inference time cannot be zero in aggregate. The paper never specifies whether TCM charges the reflector's wall-clock time when its feedback is adopted or when a new query is issued. If the reflector's latency is not charged, then RRARA is not being evaluated under the latency-aware protocol the paper advocates, and the comparison with baselines is ill-defined. If it is charged, the reported RL of 0.00 is incorrect.
[Section 3, 'without introducing additional inference latency'] The claim that the parallel reflect-and-feedback mechanism 'enhances decision quality without introducing additional inference latency' is misleading under the TCM framework. During GPT-3.5's 2.35 s reflection, the environment advances by approximately 70 frames at 30 FPS. The reflex continues moving toward its original target during this interval, and the feedback the LLM eventually provides is based on stale observations. Both the aborted reflex frames and the staleness of the feedback are real costs that TCM was designed to penalize. The paper must specify how these costs are included in RL and LAR; otherwise the zero reported for RRARA hides the very latency the paper claims to measure.
[Section 4] The empirical evaluation is a single table reporting point estimates for one scenario (fire), with no number of episodes, no random seeds, no variance measures, and no statistical significance tests. The claim that RRARA 'substantially outperforms' baseline methods is not supported by a single run without error bars. The abstract also mentions flood and wind scenarios, but no results for these scenarios are presented. The authors should provide multi-seed experiments across all three HAZARD scenarios and report standard deviations or confidence intervals.

minor comments (4)

[Section 4] The phrase 'Rule and Greedy poliy' contains a typo; it should be 'policy'.
[Section 4] The LAR definition is vague: 'the proportion of time spent reasoning relative to acting' could be formalized, e.g., as LAR = T_inf / (T_inf + T_exec), where T_exec is the total execution time. Please provide an explicit formula.
[Section 2] The TCM equation F_inf = T_inf × FPS is not numbered; numbering it would aid cross-referencing.
[Abstract] The phrase 'i.e., fire, flood, and wind scenarios' should be 'e.g.,' or 'such as', since these are examples of high-risk scenarios, not an exhaustive list.

Circularity Check

1 steps flagged · score 6.0 of 10

RRARA's reported RL=0.00/LAR=0.00 is by construction: the asynchronous GPT-3.5 reflector (2.35s standalone) is excluded from the TCM latency that the agent is designed to minimize.

self definitional [Section 3 (Rapid-Reflex Async-Reflect Agent); Table 1 in Section 4 (Experiments and Discussion)]
"By interleaving immediate reflexes with high-level reflection, RRARA achieves real-time responsiveness while integrating the high-level reasoning capabilities of the LLM, allowing for refined decision-making without sacrificing responsiveness. ... Empirical results show that the LLM-based evaluator intervenes in roughly60%of action steps, steering the agent toward better planning without incurring critical latency."

RL is defined as average inference time per decision step; RRARA's decision step is the rule-based reflex, so RL=0.00 is true by definition, not by measurement. The same GPT-3.5 model used as Reflector has RL=2.35s in Table 1, and it changes ~60% of actions; those interrupt-and-switch decisions are themselves LLM inferences. TCM's own rule (F_inf = T_inf × FPS) would charge those 2.35s, making RL=0.00 impossible unless the reflector's T_inf is excluded. The central claim 'substantially outperforms baselines in latency-sensitive scenarios' therefore rests on a metric that rewards exactly the design choice the paper made, rather than an independent measure of all deliberation that determines executed actions.

full rationale

The paper does not fit parameters or hide a self-citation chain, and the VR/DR comparisons against Rule, Greedy, MCTS, and GPT models are stated. The circularity is localized to the latency side of the evaluation: Table 1 reports RRARA RL=0.00/LAR=0.00 despite using GPT-3.5 (2.35s standalone) as the Reflector in ~60% of steps. Since 'without introducing additional inference latency' is asserted for an asynchronous process that must compute for seconds before interrupting, the zero-latency result is a consequence of how the decision step is defined, not an empirical finding. The paper would need to charge the reflector's wall-clock time (or show it does) for the central latency-aware claim to be supported. This is a partial, construction-based circularity, not a fully tautological derivation, so the score is 6.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted. The paper relies on a few domain assumptions about the HAZARD simulator and on the ad hoc assumption that asynchronous LLM feedback does not count toward decision latency.

assumptions (3)

domain assumption The HAZARD simulator at 30 FPS is a valid representation of dynamic disaster response for evaluation.
Used throughout Section 4 without validation.
domain assumption The environment advances by T_inf * FPS frames while the agent deliberates.
Stated in Section 2, a modeling choice that assumes no speculative execution.
ad hoc to paper The LLM reflector's wall-clock time can be ignored for evaluating the reflex agent.
Section 3 and Table 1 assign RL=0.00 to RRARA, effectively treating the asynchronous reflection as free in terms of decision latency.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LLM-Enhanced Rapid-Reflex Async-Reflect Embodied Agent for Real-Time Decision-Making in Dynamically Changing Environments." pith.science (2026). https://pith.science/paper/F77O63BV

@misc{pith2026250607223,
  author       = {Pith},
  title        = {Pith review of: LLM-Enhanced Rapid-Reflex Async-Reflect Embodied Agent for Real-Time Decision-Making in Dynamically Changing Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F77O63BV}},
  note         = {Machine review of arXiv:2506.07223}
}

read the original abstract

In the realm of embodied intelligence, the evolution of large language models (LLMs) has markedly enhanced agent decision making. Consequently, researchers have begun exploring agent performance in dynamically changing high-risk scenarios, i.e., fire, flood, and wind scenarios in the HAZARD benchmark. Under these extreme conditions, the delay in decision making emerges as a crucial yet insufficiently studied issue. We propose a Time Conversion Mechanism (TCM) that translates inference delays in decision-making into equivalent simulation frames, thus aligning cognitive and physical costs under a single FPS-based metric. By extending HAZARD with Respond Latency (RL) and Latency-to-Action Ratio (LAR), we deliver a fully latency-aware evaluation protocol. Moreover, we present the Rapid-Reflex Async-Reflect Agent (RRARA), which couples a lightweight LLM-guided feedback module with a rule-based agent to enable immediate reactive behaviors and asynchronous reflective refinements in situ. Experiments on HAZARD show that RRARA substantially outperforms existing baselines in latency-sensitive scenarios.

Figures

Figures reproduced from arXiv: 2506.07223 by the authors.

**Figure 2.** Illustration of the reflective process in RRARA. The [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 16 canonical work pages

[1]

ARNOLD: A benchmark for language- grounded task learning with continuous states in realistic 3d scenes

Ran Gong, Jiangyong Huang, Yizhou Zhao, Haoran Geng, Xiaofeng Gao, Qingyang Wu, Wensi Ai, Ziheng Zhou, Demetri Terzopoulos, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. ARNOLD: A benchmark for language- grounded task learning with continuous states in realistic 3d scenes. InICCV, pages 20426–20438, 2023. 1

work page 2023
[2]

Multiply: A multisensory object- centric embodied large language model in 3d world

Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Jun- yan Li, and Chuang Gan. Multiply: A multisensory object- centric embodied large language model in 3d world. In CVPR, pages 26406–26416, 2024. 1

work page 2024
[3]

Software-hardware co-design for embod- ied ai robots.arXiv preprint arXiv:2407.04292, 2024

Yiyang Huang, Yuhui Hao, Bo Yu, Feng Yan, Yuxin Yang, Feng Min, Yinhe Han, Lin Ma, Shaoshan Liu, Qiang Liu, and Yiming Gan. Software-hardware co-design for embod- ied ai robots.arXiv preprint arXiv:2407.04292, 2024. 1

arXiv 2024
[4]

Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, C

Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gok- men, Sanjana Srivastava, Roberto Mart ´ın-Mart´ın, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Ar- man Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R. Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Si...

work page 2022
[5]

Embodied agent inter- face: Benchmarking llms for embodied decision making

Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Er- ran Li, Ruohan Zhang, Weiyu Liu, Percy Liang, Li Fei- Fei, Jiayuan Mao, and Jiajun Wu. Embodied agent inter- face: Benchmarking llms for embodied decision making. In NeurIPS, 2024. 1

work page 2024
[6]

Manipllm: Embodied multimodal large language model for object-centric robotic manipulation

Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yux- ing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. InCVPR, pages 18061–18070, 2024. 1

work page 2024
[7]

Openeqa: Embodied question answering in the era of foundation models

Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Sil- wal, Paul McVay, Oleksandr Maksymets, Sergio Ar- naud, Karmesh Yadav, Qiyang Li, Ben Newman, Mo- hit Sharma, Vincent-Pierre Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Alexander Sax,...

work page
[8]

Towards generalisable audio representations for audio-visual navigation

Shunqi Mao, Chaoyi Zhang, Heng Wang, and Weidong Cai. Towards generalisable audio representations for audio-visual navigation. InCVPR-EAI, 2022. 1

work page 2022

Show all 20 references

[9]

Embodiedgpt: Vision-language pre-training via embodied chain of thought

Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. InNeurIPS, 2023. 1

2023
[10]

Mp5: A multi-modal open-ended embodied system in minecraft via active perception

Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in minecraft via active perception. InCVPR, pages 16307–16316, 2024. 1

2024
[11]

ALFRED: A benchmark for interpreting grounded instructions for everyday tasks

Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A benchmark for interpreting grounded instructions for everyday tasks. InCVPR, pages 10737–10746, 2020. 1

2020
[12]

Sadler, Jiaman Wu, Wei-Lun Chao, Clayton Washington, and Yu Su

Chan Hee Song, Brian M. Sadler, Jiaman Wu, Wei-Lun Chao, Clayton Washington, and Yu Su. Llm-planner: Few- shot grounded planning for embodied agents with large lan- guage models. InICCV, pages 2986–2997, 2023. 1

2023
[13]

Devon Hjelm, Zhe Gan, Zsolt Kira, and Alexander Toshev

Andrew Szot, Bogdan Mazoure, Omar Attia, Aleksei Timo- feev, Harsh Agrawal, R. Devon Hjelm, Zhe Gan, Zsolt Kira, and Alexander Toshev. From multimodal llms to general- ist embodied agents: Methods and lessons.arXiv preprint arXiv:2412.08442, 2024. 1

2024 arXiv
[14]

SPIN: simultaneous perception, interaction and navigation

Shagun Uppal, Ananye Agarwal, Haoyu Xiong, Kenneth Shaw, and Deepak Pathak. SPIN: simultaneous perception, interaction and navigation. InCVPR, pages 18133–18142,
[15]

Generative ai in embodied systems: System-level analysis of performance, efficiency and scala- bility.arXiv preprint arXiv:2504.18945, 2025

Zishen Wan, Jiayi Qian, Yuhang Du, Jason Jabbour, Yilun Du, Yang Katie Zhao, Arijit Raychowdhury, Tushar Krishna, and Vijay Janapa Reddi. Generative ai in embodied systems: System-level analysis of performance, efficiency and scala- bility.arXiv preprint arXiv:2504.18945, 2025. 1

2025 arXiv
[16]

EPO: hierarchical LLM agents with environment preference optimization

Qi Zhao, Haotian Fu, Chen Sun, and George Konidaris. EPO: hierarchical LLM agents with environment preference optimization. InEMNLP, pages 6401–6415, 2024. 1

2024
[17]

Towards learning a generalist model for embodied navigation

Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. Towards learning a generalist model for embodied navigation. InCVPR, pages 13624–13634, 2024. 1

2024
[18]

Tenenbaum, and Chuang Gan

Qinhong Zhou, Sunli Chen, Yisong Wang, Haozhe Xu, Wei- hua Du, Hongxin Zhang, Yilun Du, Joshua B. Tenenbaum, and Chuang Gan. HAZARD challenge: Embodied decision making in dynamically changing environments. InICLR,
[19]

TANGO: training-free embod- ied AI agents for open-world tasks.arXiv preprint arXiv:2412.10402, 2024

Filippo Ziliotto, Tommaso Campari, Luciano Serafini, and Lamberto Ballan. TANGO: training-free embod- ied AI agents for open-world tasks.arXiv preprint arXiv:2412.10402, 2024. 1

2024 arXiv
[20]

Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R

Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. ...

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

ARNOLD: A benchmark for language- grounded task learning with continuous states in realistic 3d scenes

Ran Gong, Jiangyong Huang, Yizhou Zhao, Haoran Geng, Xiaofeng Gao, Qingyang Wu, Wensi Ai, Ziheng Zhou, Demetri Terzopoulos, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. ARNOLD: A benchmark for language- grounded task learning with continuous states in realistic 3d scenes. InICCV, pages 20426–20438, 2023. 1

work page 2023

[2] [2]

Multiply: A multisensory object- centric embodied large language model in 3d world

Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Jun- yan Li, and Chuang Gan. Multiply: A multisensory object- centric embodied large language model in 3d world. In CVPR, pages 26406–26416, 2024. 1

work page 2024

[3] [3]

Software-hardware co-design for embod- ied ai robots.arXiv preprint arXiv:2407.04292, 2024

Yiyang Huang, Yuhui Hao, Bo Yu, Feng Yan, Yuxin Yang, Feng Min, Yinhe Han, Lin Ma, Shaoshan Liu, Qiang Liu, and Yiming Gan. Software-hardware co-design for embod- ied ai robots.arXiv preprint arXiv:2407.04292, 2024. 1

arXiv 2024

[4] [4]

Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, C

Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gok- men, Sanjana Srivastava, Roberto Mart ´ın-Mart´ın, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Ar- man Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R. Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Si...

work page 2022

[5] [5]

Embodied agent inter- face: Benchmarking llms for embodied decision making

Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Er- ran Li, Ruohan Zhang, Weiyu Liu, Percy Liang, Li Fei- Fei, Jiayuan Mao, and Jiajun Wu. Embodied agent inter- face: Benchmarking llms for embodied decision making. In NeurIPS, 2024. 1

work page 2024

[6] [6]

Manipllm: Embodied multimodal large language model for object-centric robotic manipulation

Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yux- ing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. InCVPR, pages 18061–18070, 2024. 1

work page 2024

[7] [7]

Openeqa: Embodied question answering in the era of foundation models

Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Sil- wal, Paul McVay, Oleksandr Maksymets, Sergio Ar- naud, Karmesh Yadav, Qiyang Li, Ben Newman, Mo- hit Sharma, Vincent-Pierre Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Alexander Sax,...

work page

[8] [8]

Towards generalisable audio representations for audio-visual navigation

Shunqi Mao, Chaoyi Zhang, Heng Wang, and Weidong Cai. Towards generalisable audio representations for audio-visual navigation. InCVPR-EAI, 2022. 1

work page 2022