EmbodiedGPT: Vision-Language Pre-Training via Embodied Chain of Thought

Embodied AI is a crucial frontier in robotics, capable of planning and executing action sequences for robots to accomplish long-horizon tasks in physical environments. In this work, we introduce EmbodiedGPT, an end-to-end multi-modal foundation model for embodied AI, empowering embodied agents with multi-modal understanding and execution capabilities. To achieve this, we have made the following efforts: (i) We craft a large-scale embodied planning dataset, termed EgoCOT. The dataset consists of carefully selected videos from the Ego4D dataset, along with corresponding high-quality language instructions. Specifically, we generate a sequence of sub-goals with the "Chain of Thoughts" mode for effective embodied planning. (ii) We introduce an efficient training approach to EmbodiedGPT for high-quality plan generation, by adapting a 7B large language model (LLM) to the EgoCOT dataset via prefix tuning. (iii) We introduce a paradigm for extracting task-related features from LLM-generated planning queries to form a closed loop between high-level planning and low-level control. Extensive experiments show the effectiveness of EmbodiedGPT on embodied tasks, including embodied planning, embodied control, visual captioning, and visual question answering. Notably, EmbodiedGPT significantly enhances the success rate of the embodied control task by extracting more effective features. It has achieved a remarkable 1.6 times increase in success rate on the Franka Kitchen benchmark and a 1.3 times increase on the Meta-World benchmark, compared to the BLIP-2 baseline fine-tuned with the Ego4D dataset.

Paper

References (73)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer eM5s8/10 · confidence 4/52023-07-01

Summary

The given text discusses the field of embodied AI and its applications in robotics. It highlights the potential of large language models (LLMs) in enabling robots to process natural language instructions, perform multi-modal tasks, and plan actions in physical environments. However, the collection of high-quality embodied multi-modal data poses challenges due to the need for egocentric data and structured language instructions. To address these challenges, the authors propose the creation of two datasets: EgoCOT and EgoVQA. EgoCOT is a large-scale embodied planning dataset that contains egocentric videos from the Ego4D dataset, along with carefully generated, filtered, and human-verified step-by-step language instructions. EgoVQA is an extension of the Ego4D dataset, focusing on egocentric human-object interaction video question answering tasks. Based on these datasets, the authors present an end-to-end multi-modal embodied foundation model called EmbodiedGPT. It consists of four integrated modules: a frozen vision model for encoding visual features, a frozen language model for natural language processing, an embodied-former for aligning visual and embodied instructions, and a policy network for low-level action generation. The model incorporates "chain-of-thought" planning capabilities and aims to enhance task performance and adaptability. The contributions of this work include the development of the EmbodiedGPT model, the creation of the EgoCOT and EgoVQA datasets, and the demonstration of the model's effectiveness in various embodied tasks such as embodied control, embodied planning, video captioning, and video question-answering. EmbodiedGPT outperforms other models on benchmark tasks, showcasing its strong performance and advantages in generating reliable planning sequences and integrating high-level planning with low-level control.

Strengths

Addressing a key challenge in embodied AI by creating new datasets: One major strength of the paper lies in recognizing a critical issue in embodied AI - the lack of good quality embodied multimodal data, especially those accompanied by structured language instructions. To tackle this issue, the authors created two new datasets - EgoCOT and EgoVQA which have the potential to significantly improve research in this area. Proposing an effective solution with state-of-the-art results: Building upon these newly developed datasets, the authors introduced the EmbodiedGPT model. This model achieved impressive performance across multiple embodied tasks including embodied control, embodied planning, video captioning, and video question answering. Focus on improving the interpretability of LLMs: Another notable aspect of the paper is its focus on improving the transparency and explainability of LLMs used in embodied AI. The authors designed the EmbodiedGPT architecture so that each module can be interpreted separately or jointly analyzed together with attention weights. Encouraging open source collaboration through code release: Finally, the authors released their code publicly, fostering further research collaborations within the community.

Weaknesses

Potential overfitting to specific domains: Despite being trained on diverse tasks, EmbodedGPT might have become too specialized in certain areas where the training datasets were biased towards, leading to suboptimal generalization abilities when facing unseen situations or domains. Besides, the interpretability and explainability of the EmbodiedGPT may remain problematic, griven transformer-based training on language, vision, and action policy, which lacks explicit memory components to ensure accurate responses based on past experiences. Training EmbodedGPT requires substantial computational resources, making it less accessible to smaller organizations or individual researchers who cannot afford expensive hardware.

Questions

What is the computational cost of training such a large model? How it the long-term behavior of EMbodiedGPT? How can embodiedGPT benefit real-world robotic applications?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations

Reviewer 5cDX3/10 · confidence 5/52023-07-02

Summary

This paper proposes an end-to-end multi-modal foundation model for embodied AI. Specifically, the authors create a dataset termed EgoCOT from the Ego4D dataset and introduce a paradigm for both the vision and embodied tasks using large language models. However, I have some concerns about this paper. My detailed comments are as follows.

Strengths

1、The paper well combines the large language models with embodied planning. 2、The proposed EgoCOT dataset and EgoVQA dataset are helpful for downstream tasks such as embodied planning and VQA.

Weaknesses

To me, this paper is close to being interesting but I, unfortunately, do feel there are some weaknesses that, currently, prevent me from recommending acceptance. 1、The main idea has limited technical contribution: The entire paper reads like a technical report and many parts of it are not coherent. The network architecture level in this regard can be said to have no innovative contribution. The frozen modules in the paradigm have been proposed and used in many papers related to large language models. Essentially, embodied former is also actually a cross-attention network used for visual and language modality interaction. As for the policy network, it simply encodes vision and combines it with multi-modal interaction output as input. I understand this paper has the purpose to use large language models for embodied tasks. However, at the method level, I see more about how to interact and align the image or video modality with the text modality, and in the experimental part, it also focuses on reflecting the ability of the method in image caption and object grounding, not even in some important experiment on embedded tasks (such as embedded navigation, embedded rearrangement, Instruction Following and Dialogue, etc.) 2、As a side note, I would like to let the authors know about some flaws in the presentation of the paper: (1) In Section 3.1, in the second stage of the training, how are the language projection and prefix language adapter updating? I suggest that the authors explain this in detail. (2) In Section 3.1, in the third stage of the training, does the Conv3D module need to be learned? And how to decide the time offset to be 2? I didn’t see any explanation or ablation studies of these hyperparameters. (3) In Section 3.4, how to decide the start and end time points of a clip? Please explain the relationship of \beta and \alpha and why the \beta/2*\alpha becomes the bound. (4)I suggest adding more descriptions of the tasks conducted in the experiments. For example, I am confused about that why are the indicators Plan Reasonable and Plan Executable related to image caption tasks from MS-COCO.

Questions

See the weaknesses.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

See the weaknesses.

Reviewer 5cDX2023-08-19

Thanks for the responses of the authors. So far, the authors' reply has not satisfied me. Here are my reasons: 1. As for the novelty of the paper, I still believe that this paper does not have enough contributions and novelty. The author claims that “…The core focus is exploring how chain-of-thought training can impart understanding of object interactability...while other multimodal models tend to concentrate on the detection of objects…”, which is extremely incorrect and has not been rigorously investigated. In fact, several previous works have explored how to use LLM to embodied tasks. Specifically, they also all use the prompt engineer and chain of thought, such as [1][2]. Moreover, please **DO NOT** try to avoid and refute the novelty of the architecture. Review BD8Q is also concerned about the novelty of the architecture. As far as the above is concerned, I really don't see the novelty and value of this paper, but it makes me feel that the author is overclaiming. 2. Regarding the newly added embodied task experiments, the authors disappoint me even more. I have read the pdf of the rebuttal. I think these experiments are very tricky, making no comparison with the SOTA methods and it is difficult for me to see any in-depth analysis, especially how to reflect the role and novelty of the method proposed in this paper in the tasks. Take embedded navigation as an example, the benchmarks in this community are very mature now. The authors should conduct their method in the strong benchmarks (such as R2R, RXR, VLNCE, etc.) of VLN to compare and analyze with existing SOTA methods and give the quantified data, similar to NavGPT [2]. However, the authors gave a so-called "navigation" demo containing a trajectory composed of photographed pictures. As a reviewer, I don't even know what the setting is and whether the setting is reasonable, where the data comes from, and how the method of the paper works in it. There is also another reviewer (i.e., Reviewer P2YP) who questions the setting of this paper for the embodied tasks. I strongly recommend the authors make a comprehensive survey and understanding of the embodied community. And give a comparison of the **quantitative results** of the **SOTA methods** for each embodied task I mentioned. Otherwise, I think the authors cannot simply put the "embodied" hat on his method. Thanks again for your time for reading my questions. [1] Wu J, Antonova R, Kan A, et al. Tidybot: Personalized robot assistance with large language models[J]. arXiv preprint arXiv:2305.05658, 2023. [2] Zhou G, Hong Y, Wu Q. NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models[J]. arXiv preprint arXiv:2305.16986, 2023.

Authorsrebuttal2023-08-20

Dear reviewer 5cDX, Thank you for your response and suggestions. We respectfully push back your premise that novelty of the architecture should be the main consideration. We believe our multi-modal chain-of-thought training, end-to-end embodied control, and our EgoCoT dataset are strong contributions. Just as LLama2 [1] saw significant performance gains over LLama [2] by incorporating higher quality data and thereby greatly benefited the community, we believe our EgoCOT dataset offers a valuable contribution to the field of embodied AI. Moreover, the papers [3, 4] you mentioned were either published on arXiv after the NeurIPS submission timeline or overlapped with it. We believe these should not be used as grounds to diminish our contribution. We agree with you that more experiments on strong VLN benchmarks would be beneficial to our work, and we will surely include them as our future work. However, the focus of our work is pioneering research on how to build a general multimodal embodied AI foundation model that can serve as a basic tool for multiple tasks, rather than achieving SoTA performance on all benchmarks, which is out of our current scope. In this work, we referred to the most representative embodied large model work Palm-e [5] from Google, and adopted diverse manipulation tasks to validate the model's effectiveness. We also provided ample feasibility studies of applying EmbodiedGPT to real-world navigation, rearrangement and other tasks, to demonstrate its versatility. Thank you for your time and consideration. [1] Touvron H, Martin L, Stone K, et al. Llama 2: Open foundation and fine-tuned chat models[J]. arXiv preprint arXiv:2307.09288, 2023. [2] Touvron H, Lavril T, Izacard G, et al. Llama: Open and efficient foundation language models[J]. arXiv preprint arXiv:2302.13971, 2023. [3] Wu J, Antonova R, Kan A, et al. Tidybot: Personalized robot assistance with large language models[J]. arXiv preprint arXiv:2305.05658, 2023. [4] Zhou G, Hong Y, Wu Q. NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models[J]. arXiv preprint arXiv:2305.16986, 2023. [5] Driess D, Xia F, Sajjadi M S M, et al. Palm-e: An embodied multi-modal language model[J]. arXiv preprint arXiv:2303.03378, 2023. Sincerely, \ Authors

Reviewer P2YP8/10 · confidence 5/52023-07-05

Summary

In this work, the authors propose an end-to-end multi-modal model for Embodied AI field. Based on the Ego4D dataset, the authors build a large-scale Embodied planning dataset, i.e., EgoCOT, which contains machine-generated language instruction with chain-of-thoughts mode. Besides, an EmbodiedGPT based on LLM is proposed to achieve the plan generation task, where the corresponding training approach is designed for prefix tuning on EgoCOT. Experimental results demonstrate that EmbodiedGPT can achieve the promising ability of plan generation.

Strengths

- This work provides the EgoCOT dataset with EgoVQA as an extension, which can promote the development of LLM in the Embodied AI field. - This work makes the pioneering attempt to build a foundation model with the chain-of-thought capability. Building the framework based on the Vision Transformer and LLaMA, and designing the Embodied-Former as the bridge between the visual and language modality is reasonable. - Extensive quantitative and qualitative results are provided for examining the model's effectiveness. The results show that EmbodiedGPT can produce reasonable answers to describe what happened in the input video and generate the step-by-step action sequence. - The paper is well written and easy to follow with detailed instructions about how to construct and train the model.

Weaknesses

- In the field of Embodied AI, an important task is \*agent navigation*, but this work does not cover this part. I understand that the authors focus on the multi-modal foundation model, and there is a vision-language navigation (VLN) task [1][2] that requires the agent to conduct navigation according to the given language instruction. Thus I believe the VLN is a promising task for authors to test the model's navigation ability, e.g., verifying if the model can generate step-by-step navigation instructions with chain-of-thoughts. - I am confused about the setting of EmbodiedVQA in this paper. In my view, EmbodiedVQA [3] is actually the task that embodied agents navigate in rich 3D environments in order to answer the given question. - Currently, the model can generate detailed descriptions of a given video and convert them into planning steps with chain-of-thoughts. However, can the model generate executable steps for completing a specific task based solely on a static image of an object, rather than a demonstration video? - The authors should provide more details on how to use the model for Embodied control, specifically how to convert the text output into step-by-step control signals, as this is a crucial aspect for implementing embodied agents. - Minor: the colors of Figure 2 are a bit chaotic, and if there is no special meaning, I suggest it can be simplified. [1] Vision-and-Language Navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR 2018. [2] Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR 2020. [3] Embodied Question Answering. In CVPR 2018.

Questions

See Weaknesses.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

I do not foresee obvious undesirable ethical or social impacts at this moment.

Reviewer BD8Q7/10 · confidence 4/52023-07-07

Summary

This paper introduces EmbodiedGPT, a new multi-modal foundation model designed for application to embodied (robotic) tasks. EmbodiedGPT combines existing large pretrained (and frozen) vision and language models (namely the ViT-G/14 EVA-CLIP vision model and LLaMA-7B) through the use of an "Embodied-Former" which converts input text tokens and vision features into prefix tokens that can be fed directly to the LLaMA model. To pretrain this Embodied-Former model, this work introduces two new datasets (EgoCOT and EgoVQA), which contain video clips taken from the Ego4D dataset annotated with text captions generated by ChatGPT that, for EgoCOT, describe chain-of-thought-style plans corresponding to the actions taken within the clip, and, for EgoVQ, include questions and answers related to the clip. Using these new datasets and a 3-stage training pipeline, this paper trains the EmbodiedGPT model and, via human evaluations, show that humans generally prefer the captions and embodied plans generated by EmbodiedGPT even when compared to a LLaVA-13B baseline which has many more parameters. Finally, the paper shows that, by attaching an MLP policy head to the output of the Embodied-Former, it is possible to use imitation learning from a small set of demonstrations to train performant physical-manipulation agents in simulation. Experiments show that EmbodiedGPT outperforms ablations of itself and a R3M-based SoTA model for these tasks.

Strengths

This paper studies a very popular, important, and challenging area of work: building foundational models for robotic control and embodied world understanding. The proposed EmbodiedGPT model represents a modest but non-trivial step towards this grand goal. In my view, the major contributions of this are the substantial efforts taken to create the EgoCOT/EgoVQA datasets and execute the complex and computationally expensive 3-stage strategy required for training the EmbodiedGPT model. Generating these datasets and training the model clearly required significant effort and the experimental results appear quite strong.

Weaknesses

While I lean positive for this paper, I will highlight some weaknesses below that temper my excitement. ## Statistical significance of the human results The only evaluation of EmbodiedGPT against other existing large vision-and-language models comes in a 30-participant human study using 20 tasks (10 captioning, 5 simulated planning, 5 real planning). As this sample size is quite small, and the performance difference sometimes relatively minor, it's challenging to draw meaningful conclusions from the results in Table 1. Moreover, it is hard to know how representative the 20 tasks are of general captioning/planning tasks, how were these chosen? There are ways to provide a statistical analysis of this data but they are not trivial: you'd need to account for the correlated nature of the data (e.g. that individual participants rated multiple examples and that performance may be correlated with the task type and within each question). While not a perfect treatment, at the very least a paired t-test could be used to compare EmbodiedGPT with the other models. ## Really "foundational"? While EmbodiedGPT is touted as a "foundation model for embodied AI", the results do not quite support this in my view. In particular, I feel that the EmbodiedGPT misses the critical capacity to handle partial observability and I was surprised not to see any real-world examples of robotic control, more details below. ### Lack of tasks involving partial observability The only embodied tasks evaluated in this paper involve static robotic arms acting in fully observed (or almost fully observed) environments. There are many embodied tasks (e.g. visual navigation) for which partial observability plays a critical role. From my understanding, the policy component of the EmbodiedGPT model makes decisions from a single visual observation in isolation from all previously seen states and thus would be expected to fail in all such tasks. ### Lack of application to real-world embodied tasks While it is great to see that EmbodiedGPT outperforms the R3M model in the simulated environments, one of the strengths of the R3M paper was demonstrating that it also functioned well when applied to real robots. While not necessary for acceptance, it's difficult to call EmbodiedGPT a robotic foundation model without at least some real-world results. ## Architectural novelty This is a very minor point as I consider the training of large models to be a significant contribution in itself, but the proposed EmbodiedGPT architecture is not especially novel: it combines popular pretrained vision/language models with well-known prefix tuning ideas. While you cite Tsimpoukelli et al. (2021), I was a bit surprised to see more discussion of how their work relates to the proposed approach. Moreover, simply searching for "embodied" in the citations of the above paper (i.e. https://scholar.google.com/scholar?hl=en&as_sdt=5%2C48&sciodt=0%2C48&cites=16154696122208258147&scipsc=1&q=embodied) appears to surface several additional citations that should be included in this work.

Questions

I have listed some of my major questions and weaknesses above. If some of these weaknesses can be addressed (especially the point about statistical significance) I expect to increase my score. In the below I include, more minor, line-by-line comments and questions. Line 60 - "EmbodiedGPT ’s" should be "EmbodiedGPT’s". I suspect this is a problem due to latex macros, try adding {} after your macro inline, i.e. \mymodel{}. Line 77 - This "200M annotated videos" statement seems like it must be wrong. Later on, you note that there are 2927 hours of video in total, this is approximately 10M seconds of video which would correspond to ~200M frames depending on the framerate, is that what you mean? Figure 2 - This figure could use some work, for instance, it's very unclear how videos are being input into the model. Lines 146-147 - What is the value of the "learnable embodied queries"? What makes them "embodied"? Lines 172-174 - I found this description of the video encoding procedure to be too vague, please add more details. Lines 181-204 - This entire discussion could be simplified and better presented, the idea here is quite simple and much of the formalism is extraneous. For instance, you introduce a "mapping function (...) M: z -> z'" but this mapping function is just a linear projection and the "M" notation seems to be never used again. You could just say "z is then mapped to z' by via a linear projection." Line 241 - Extra space after the comma in 9,645. Lines 225-228 - How is \beta_i "adjustable" if it's defined in this way? It seems to be a fixed function of other fixed values. - Why not just set [t^{start}_{i}, t^{end}_{i}] to be [(t_{i-1}+t_{i})/2, (t_{i+1} + t_{i})/2]? Figure 3 - This seems like a surprising qualitative example to choose as the EmbodiedGPT model says many wrong things. E.g. (1) the clock is not actually above the door, (2) the coat hanger is not on the floor, and (3) the coat hangers can be seen towards the top left, not the "near the right side of the doorway". Figure 4 - Again, this is a surprising example as the video shows it lowering the ring over the cylinder twice whereas EmbodiedGPT seems to only suggest this happens once.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

I would have liked to see some discussion of the limitations I discussed in the _Really "foundational"?_ section of the weakness. Otherwise, limitations are reasonably well addressed.

Authorsrebuttal2023-08-08

Real robot demo of EmbodiedGPT

Dear Area Chair and reviewers, Thanks for your efforts in handling our manuscript and we would like to express our gratitude for the constructive suggestions provided by the reviewers. In response to the request from reviewer [BD8Q,5cDX], we have conducted an experiment on a real robot, focusing on a typical embodied rearrangement task. The experiment was performed using a Franka Emika robot arm. Please find a demonstration video at the following link: https://www.dropbox.com/scl/fi/qrc2514rgwoalzg0419gc/rearrangement_demo_EmbodiedGPT.mp4?rlkey=1d2wea2g26tccy478vvwsnu5z&dl=0. Best, \ The authors.

Reviewer eM5s2023-08-10

Thanks for the additional details provided. I will take them into consideration.

Authorsrebuttal2023-08-13

Sincerely looking forward to more discussion with you

Dear Reviewer 5cDX, We deeply appreciate the time you dedicated to reviewing our work and the insightful feedback you provided. We have crafted detailed responses and updated our results based on your comments, aiming to address all your concerns. We genuinely hope to engage in further dialogue to ensure that we have fully met your expectations. If any aspects of our work remain unclear, or if you have additional concerns, please don't hesitate to communicate them. We remain committed to providing clarity and addressing all feedback. We are always ready to address your further concerns and improve our paper. Best regards, The Authors.

Authorsrebuttal2023-08-13

Thanks for your great efforts and look forward to your reply

Dear reviewer BD8Q, Firstly, we'd like to express our profound gratitude for your previous suggestions. Your recommendations were meticulous and constructive, prompting deeper reflections on many issues. We truly value the effort and thought you've invested in our work. We deeply appreciate the time you dedicated to reviewing our work and the insightful feedback you provided. Following your comments, we've crafted detailed responses and made necessary updates, aiming to address all your concerns. We sincerely hope to further discuss with you and ensure that we've completely met your expectations. If any aspects of our work remain unclear, or if further concerns arise, please don't hesitate to share them. Our commitment remains steadfast in providing clarity and responding to all feedback. Best regards, The Authors.

Reviewer BD8Q2023-08-15

Dear Authors, I very much appreciate the substantial effort you have undertaken in this rebuttal period: the addition of a real-world experiment is especially impressive! While I do believe there are some ways in which the paper could be further improved (e.g. the real-world and partially-observed experiments are still much more "qualitative" than they are "quantitative") I am generally satisfied with the responses and plan to increase my score. Best, BD8Q

Authorsrebuttal2023-08-16

Sincerely thank you for your support and recognition of our paper

Dear Reviewer BD8Q, We sincerely thank you for your constructive suggestions and effort. We will add more quantitative results to our final version according to your recommendations. Thank you for your kind suggestions and warm help. Please don’t hesitate to let us know if there are any additional clarifications we can offer, looking forward to your post-rebuttal rating! Best, The Authors.

Authorsrebuttal2023-08-18

Would you kindly consider raising the score?

Dear reviewer 5cDX, Thank you again for your insightful suggestions to improve this work. We appreciate you taking the time to provide detailed and constructive feedback. Regarding your concern about the limited novelty of the architecture, we wish to clarify that the core innovations of our approach lie in the embodied chain of thought training methodology and the EgoCOT dataset, rather than the model architecture itself. Based on your recommendation, we have added experiments on embodied navigation and rearrangement tasks. These help showcase our model's ability to leverage its knowledge to infer sensible directions and manipulate objects properly. As for dialogue and instruction following, our original experiments already incorporated human language inputs to accomplish goals across multiple turns. We have also provided more implementation details around the training process and hyperparameters based on your questions. We will include them in our final paper. As the rebuttal period is ending soon, we wonder if our response answers your questions and addresses your concerns. If yes, would you kindly consider raising the score? Thanks again for your very constructive and insightful feedback! Best, \ Authors

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC