PAPER: IntroductionHand-object interaction (HOI) involves jointly modeling hand articulation and object dynamics to generate and interpret realistic manipulation sequences [9,17,25,26,15,47,5]. This reflects one of the most pervasive human behaviors, deeply embedded in daily activities. Generating and understanding 3D HOI sequences is critical for advancing machine capabilities in human-centric applications. In augmented/virtual reality (AR/VR), realistic HOI modeling enables immersive digital experiences, allowing users to manipulate virtual objects naturally. For robotics, it provides the foundation for dexterous, feedback-driven manipulation in unstructured environments.Generating HOI sequences from natural language instructions remains a significant challenge in 3D interaction research. While traditional methods [19,40] rely on handcrafted motion priors, recent diffusion-based approaches [2,3] directly map text to action sequences, eliminating the need for manual motion design. However, due to limited data and modeling capacity, these models only deal with closed datasets and struggle to generalize to unseen objects and open-vocabulary instructions.Recent advances in Large Language Models (LLMs) have significantly enhanced joint vision-language understanding. Building on this progress, concurrent work like HOIGPT [13] proposes an LLM-based architecture for aligning textual instructions with HOI sequences. However, these approaches face :"I'm feeling thirsty, could you find a water bottle and take a sip?" :"I'm a bit dehydrated, please open the water bottle cap with both hands, then drink the bottle water using your right hand.""I want to relax and read, go ahead and hold the book with both hands to read it.""We need the cocoa, can you take it out with your left hand, please?" "I'm feeling a bit thirsty, could you pour some milk using your right hand?" fundamental generalization challenges in unseen shapes and complex 3D interaction scenarios due to the lack of 3D knowledge in LLM. Fig. 1 illustrates the comparison between conventional Closed-Set HOI and Open-World HOI.The growth of 3D datasets has advanced 3D multimodal large language models (MLLMs) with strong capabilities in understanding and grounding 3D geometric, semantic, and functional relationships. Critically, fine-grained part-level grounding and affordance grounding [34] [4] can identify actionable object regions and their interaction possibilities, providing strong priors for synthesizing physically consistent hand-object interactions (HOI). Building on this insight, we propose OpenHOI, the first open-world HOI synthesis framework capable of generating long-horizon manipulation sequences for unseen objects guided by open-vocabulary instructions. Our approach fine-tunes a 3D MLLM endowed with comprehensive affordance reasoning priors, enabling two core capabilities: 1) Generalizable Affordance Prediction: Precise localization of interaction regions (e.g., handles, buttons) for both known and novel objects. 2) Instruction Decomposition: Breaking down open-vocabulary commands (e.g., "I'm feeling thirsty, could you find a water bottle and take a sip") into executable sub-tasks grounded in object affordances. By integrating these priors with a diffusion-based interaction generation and physics-aware refinement, OpenHOI achieves unprecedented generalization across object categories, task horizons, and linguistic complexity.In summary, our contributions are as follows:• We introduce OpenHOI, the first open-world hand-object interaction synthesis framework capable of generating long-horizon manipulation sequences for unseen objects guided by open-vocabulary instructions. Unlike prior closed-set HOI methods, OpenHOI generalizes to novel objects and complex, linguistically diverse commands.• OpenHOI involves fine-tuning a 3D MLLM that jointly learns geometric affordance priors and semantic task decomposition. Subsequently, affordance-driven HOI Diffusion with Physical Refinement is developed to generate realistic HOI sequences for each task.• Our experiments demonstrate state-of-the-art performance, outperforming existing methods with a large margin. Notably, OpenHOI generalizes robustly to unseen objects and open-vocabulary instructions, achieving strong compositional generalization across diverse scenarios.Table 1: Comparative analysis of our method versus existing HOI synthesis approaches.
================================================================================

REVIEW
--------------------------------------------------------------------------------
# Summary Of The Paper

The paper introduces **OpenHOI**, a novel open-world hand-object interaction (HOI) synthesis framework that leverages a 3D multimodal large language model (MLLM) to generate long-horizon HOI sequences from open-vocabulary instructions. The framework operates in two stages: (1) instruction decomposition and affordance reasoning using a 3D MLLM, and (2) HOI sequence generation using a diffusion model conditioned on affordance maps and subtasks. The model incorporates physical refinements to ensure realistic and physically plausible interactions. The paper reports strong performance on the GRAB and ARCTIC benchmarks, particularly on unseen objects and open-vocabulary instructions.

---

# Strengths

- **Novel Framework Design**: OpenHOI introduces a compelling open-world HOI synthesis framework that generalizes to unseen objects and open-vocabulary instructions. This is a significant departure from prior closed-set methods, which are constrained to predefined object classes and instructions.
- **Integration of MLLM and Diffusion-Based Generation**: The combination of a 3D MLLM for affordance reasoning and diffusion-based HOI generation is innovative. This integration allows for both high-level semantic understanding and low-level motion planning, which is a notable advancement in the field.
- **Comprehensive Evaluation**: The paper conducts extensive experiments on established benchmarks (GRAB and ARCTIC), and includes ablation studies that isolate the effects of key components such as affordance reasoning, classifier-free guidance, and physical refinement. The results show substantial improvements over existing methods, particularly on unseen objects and open-vocabulary instructions.
- **Architectural Details Provided**: The paper provides detailed explanations of the MLLM fine-tuning process, the use of classifier-free guidance, and the incorporation of physical constraints into the diffusion process. These details are critical for reproducing the work and understanding the inner workings of the model.

---

# Weaknesses

- **Insufficient Clarification of Training Regimen**: The paper describes the use of a two-stage fine-tuning process involving coarse-grained and fine-grained datasets, but does not clarify the exact nature of these datasets, their sources, or the methodology for transitioning between stages. This raises questions about the reproducibility and validity of the fine-tuning process.
- **Ambiguity Around Hyperparameter Selection**: The loss function used during MLLM training combines task and affordance losses with coefficients $\lambda_{\text{task}}$ and $\lambda_{\text{aff}}$, but the paper does not report the values of these hyperparameters or the process used to select them. This omission undermines the replicability of the model.
- **Limited Discussion of Failures and Edge Cases**: While the paper notes that the model struggles with tasks such as targeting specific objects ("open the second cabinet") and degrading after multiple consecutive actions, it does not thoroughly analyze the root causes of these failures or propose solutions. This weakens the argument for the model’s robustness and practical utility.
- **Lack of Comparison With Relevant Baselines**: The paper excludes HOIGPT [13], a closely related method, from the comparison. This omission reduces the credibility of the claim that OpenHOI is the "first open-world" HOI synthesis framework, as it does not adequately address how it compares to existing alternatives.
- **Vague Description of Physical Refinements**: The physical refinements incorporated into the diffusion process are described in a high-level fashion without concrete details on how the constraints are encoded or enforced. This makes it difficult to assess the true impact of these refinements on the realism and plausibility of the generated sequences.

---

# Questions For The Authors

1. **Clarify the Nature of the Datasets Used for Fine-Tuning**: Could the authors provide more detailed information about the datasets used for coarse-grained and fine-grained tuning, including their origins, sizes, and the criteria used to transition between stages of training?

2. **Provide Values and Rationale for $\lambda_{\text{task}}$ and $\lambda_{\text{aff}}$**: What were the exact values of $\lambda_{\text{task}}$ and $\lambda_{\text{aff}}$ used during training, and what was the process for selecting these values?

3. **Address the Handling of Specific Object Targeting**: The paper notes that the model struggles with tasks like "open the second cabinet". Could the authors elaborate on the reasons behind this limitation and propose potential improvements to address such scenarios?

4. **Include HOIGPT in the Baseline Comparison**: Why was HOIGPT [13] excluded from the comparison, and could the authors provide a revised comparison that includes this relevant baseline?

5. **Elaborate on the Implementation of Physical Refinements**: Could the authors provide more detailed information on how the physical constraints (such as penetration and motion in-between) are encoded into the loss functions, and whether these constraints are learned or manually defined?

---

# Limitations Not Addressed By The Authors

- **Generalization Beyond Single-Object Tasks**: The paper does not evaluate the model’s performance on multi-object interaction tasks or complex scenes involving multiple objects. This limits the understanding of the model’s scalability and applicability to real-world scenarios.
- **Computational Efficiency and Real-Time Performance**: There is little discussion of the computational cost or latency of the model, particularly in relation to real-time applications in robotics or VR. This is a concern for practical deployment.
- **Ethical and Societal Implications**: The paper does not address potential ethical concerns or broader societal impacts of deploying such a model in real-world applications, such as misuse in surveillance or deception.

---

# Soundness

**Score: 3 (Good)**  
While the paper presents a well-designed framework and demonstrates strong performance on benchmark datasets, the lack of detailed explanations regarding training methodologies, hyperparameter selection, and the exclusion of relevant baselines reduce the soundness of the claims. The ablation studies and empirical results are solid, but the absence of deeper technical insights limits the overall reliability of the conclusions.

---

# Contribution

**Score: 3 (Good)**  
The paper contributes a novel open-world HOI synthesis framework that integrates MLLMs with diffusion-based generation and physical refinement. However, the novelty is partially undermined by the reliance on existing techniques and the omission of key comparisons with prior work. The contribution is valuable but not groundbreaking.

---

# Confidence

**Score: 4 (High Confidence)**  
Despite the identified shortcomings, the paper presents a well-documented and technically sound framework. The results are convincing, and the methodology is largely transparent. The confidence in the results is high, although some ambiguities persist.

---

# Rating

**Score: 8 (Accept)**  
The paper deserves acceptance due to its novel framework, strong empirical results, and comprehensive evaluation. However, the authors are advised to address the outstanding questions and limitations in future revisions to strengthen the contribution and clarity of the work.

---

# Brief Justification For Rating

The paper introduces a novel and well-executed framework for open-world HOI synthesis that demonstrates strong performance on established benchmarks. While the technical details surrounding training and hyperparameter selection are insufficient, the empirical results are robust, and the framework is clearly described. The exclusion of relevant baselines and lack of detailed discussions on failure modes limit the impact, but the overall contribution is significant enough to warrant acceptance.

================================================================================
