1 Introduction↩︎

Human motion is fundamentally shaped by interaction with our surroundings. We engage with objects in diverse ways, simultaneously manipulating multiple objects, or choosing to handle a single object with one or both hands depending on factors such as its weight, size, or the task at hand. This kind of coordinated bimanual multi-object manipulation, although fundamental to human-like behavior, remains severely underexplored from current human-object interaction (HOI) generation research  [1]. Synthesizing such motions is more than a modelling curiosity: it underpins numerous applications, from realistic interactions in AR/VR systems, to embodied agents that assist in cluttered human environments and robotic systems that rely on manipulation priors  [2]–[4].

Recent progress in HOI generation has led to increasingly realistic motions on primarily single-object benchmarks, driven largely by diffusion-based models  [5]–[7], Zhang2025diffgrasp?. Despite this, most existing HOI methods are presented with three limitations. First, nearly all state-of-the-art full-body methods assume only single-object interactions. Although a few datasets  [8]–[10] include multi-object sequences, dedicated generative methods for this setting remain largely unexplored. Second, many existing methods focus only on a subset of the human body: whole-body approaches such as the OMOMO  [11] benchmark omit detailed hand articulation, while hand-centric methods like Text2HOI  [12] and LatentHOI  [13] partially address fine-grained grasping but sacrifice full-body generation. Third, current approaches in full-body HOI motion synthesis rely on auxiliary modules and optimization strategies to enforce accurate contacts. Some examples include predicting intermediate contact maps to guide generation  [14], conditioning the model on full object trajectories  Zhang2025diffgrasp? or iterative test-time refinement such as DNO  [15] with contact-aware constraints  [6], [7], resulting in inference overhead.

To address these limitations, we introduce a paradigm shift in both the representation of object motion and the architectural design of the generative model. We present Dex2HOI, a unified framework to tackle both single- and two-object full-body HOI synthesis from text. Our main objective is to approach HOI generation in a simple yet informed way: removing model redundancy, avoiding multi-stage pipelines, and producing accurate contacts in a single-shot diffusion sampling. At the representation level, we model object motion as a learnable decomposition of hand-relative branches fused through learnable weights. This decomposition encodes a strong inductive bias for hand-object contact and naturally guides the diffusion process toward smooth transitions of control between hands. At the architectural level, we propose the use of a Dual-Stream Diffusion decoder that processes each object in a dedicated interaction stream and models information across streams through bidirectional cross-attention. The body motions from each stream are then combined with our Motion Fusion Network to produce a single coherent motion. The same architecture handles both one- or two-object sequences without modification, supporting a unified treatment of single- and bimanual manipulation. Finally, we train an end-to-end model with contact-aware supervision, removing the need for test-time optimization or auxiliary contact-prediction modules. Together, these choices enable single-shot generation that is up to \(\times\)​540 faster than prior optimization-based state-of-the-art approaches. Overall, our contributions can be summarized as follows:

• We propose the first unified HOI generation model that synthesizes both single- and two-object interactions, advancing beyond existing single object approaches.

• We introduce a novel hand-relative object motion representation and train end-to-end denoising models that enable efficient single-shot generation of complex sequences while maintaining physical plausibility and contact consistency across two hands and two objects.

• We evaluate Dex2HOI generated motions across three benchmarks, GRAB  [16] for single-object HOI, and HIMO  [9] and HUMOTO  [10] for two-object HOI. We achieve state-of-the-art performance in both single- and two-object settings, which we further validate through VLM-based and user preference evaluation studies.

2 Related Work↩︎

2.0.0.1 Human Motion Generation.

Human motion generation is an open problem in computer vision, covering a broad set of tasks and applications. In recent years, diffusion-based models have emerged as the dominant paradigm  [17]–[19], with the Motion Diffusion Model (MDM)  [18] establishing a simple yet effective denoising framework for text-conditioned motion synthesis. Autoregressive formulations  [20], [21] extend diffusion-based Motion Synthesis to long-horizon generation by predicting motion segments conditioned on a temporal context window, recursively using the latter frames as the condition for the next segment. Subsequent work extends these paradigms along several axes, beyond isolated body motion. Human-Human interaction methods  [22]–[25] coordinate motion for multiple subjects, Human-Scene interaction methods  [26]–[29] condition motion in static 3D environments, while Human-Object interaction  [5]–[7], [30] targets plausible interaction sequences between a subject and usually one object at a time.

2.0.0.2 Human-Object Interaction Generation.

Progress in HOI generation has been particularly driven by the increasing presence of 4D HOI datasets, which span from small object manipulation datasets  [16], [31], [32], to interaction with larger everyday objects, including furniture and household devices  [11], [33], [34] and some with even articulated objects  [35]. These datasets include interactions with a single object of interest at the time, while datasets that address the manipulation of more than two objects are more limited  [8]–[10].

Early works on HOI generation  [30], [36], [37] rely on conditional variational autoencoders (CVAEs) to model the distribution of plausible whole-body interactions. Many recent works have adopted denoising diffusion to improve realism, diversity, and conditioning flexibility in HOI synthesis  [5]–[7], [14], [38]–[42], Zhang2025diffgrasp?. Among these, CoDA  [6] and HOIDiNi  [7] formulate HOI generation as a diffusion noise optimization problem, refining generated motion through contact-aware guidance to produce more plausible whole-body interactions, while CoDA further extends to articulated object manipulation. Additionally, DiffGrasp  Zhang2025diffgrasp? chooses to generate whole-body grasping motions guided by whole sequence object trajectories. Other works focus on controllability and structure, including CHOIS  [38] for controllable interaction synthesis, Contact-Guided HOI  [14] for leveraging contact priors. Recent large-scale frameworks such as InterAct  [40] further push toward generalizable single-object HOI animation.

Beyond purely generative formulations, a parallel line of work formulates HOI as a control problem and learns RL policies that produce physically plausible interactions  [4], [43]–[45]. The aforementioned methods primarily focus on single-object interactions and do not explicitly address the generation of more complex two-object manipulations, leaving a significant gap in modeling coordinated bimanual interactions.

2.0.0.3 Hand-Object Interaction Generation.

A complementary line of research targets hand-centric interaction synthesis, focusing on generating fine-grained grasps. Early approaches such as ManipNet  [46] predict dexterous finger motion using explicit hand-object spatial representations. More recent methods use denoising models to generate plausible hand-object interactions  [12], [13], [47]–[50]. LatentHOI  [13] learns a latent hand diffusion model for generalizable grasp generation. For more than one object, ManipTrans  [51] addresses bimanual two-object grasping by transferring dexterous manipulation priors across objects via residual learning. While these methods achieve high-fidelity hand-object contact, they disregard full-body motion synthesis. Dex2HOI bridges this gap by jointly modeling dexterous bimanual contact and full-body motion within a single framework.

3 Method↩︎

The overview of the proposed Dex2HOI is illustrated in Fig. 3. In Section 3.1 we introduce our hand-relative object representation, which disentangles object motion into wrist-local and global branches. Section 3.2 then describes our Dual-Stream HOI diffusion and Motion Fusion Network that coordinate per-object streams into a single coherent full-body motion. Finally, in Section 3.3 we introduce our geometry-informed interaction-aware training objective.

3.1 Human and Object Motion Representations↩︎

We represent each motion sequence as a temporally aligned multimodal feature vector containing human rotations, human root translation and per-object trajectory representations. For a sequence of length \(T\), the full motion feature at frame \(t\) can be defined as: \[\mathbf{f}_t = \left[ \mathbf{h}_t^{\{A,B\}},\; \mathbf{r}_t^{\{A,B\}},\; \mathbf{o}^{\{A,B\}}_t \right], \quad 0 \le t \le T\] where \(\mathbf{h}_t\) denotes the human 6D  [52] rotations, \(\mathbf{r}_t \in \mathbb{R}^3\) denotes the human root translation, and \(\mathbf{o}_t\) the hand-relative object representations for the two objects \({\{A,B\}}\).

3.1.0.1 Human Representation.

We represent human body poses at frame \(t\) as a set of 6D  [52] joint rotations \(\mathbf{h}_t \in \mathbb{R}^{6J}\), where \(J\) denotes the number of joints. We set \(J=52\) for datasets that utilize SMPL-X parametric human model  [53] and \(J=65\) for datasets that utilize Mixamo parametric avatars  [54]. To facilitate training, we also parameterize the 3D root translation \(\mathbf{r}_t\) as the displacement relative to the first frame position (global pelvis bone).

3.1.0.2 Hand-Relative Object Representation.

Modeling object motion in global coordinates alone is insufficient for predicting interaction-dependent trajectories: prior approaches consequently resort to multi-stage pipelines or test-time optimization to recover plausible contact. Relying purely on a global object trajectory provides a weak inductive bias and forces the network to implicitly infer the controlling hand at every frame. To address this, we introduce a hand-relative object representation that explicitly disentangles object motion into right-hand, left-hand, and global components. Specifically, for frame \(t\), we define the object feature vector as: \[\mathbf{o}_t = \left[ \mathbf{r}^{obj}_t,\; \Delta^{L}_t,\; \Delta^{R}_t,\; \mathbf{p}^{G}_t,\; \boldsymbol{\alpha}_t \right] \in \mathbb{R}^{18},\] where \(\mathbf{r}^{obj}_{t} \in \mathbb{R}^6\) denotes the global object rotation in 6D form  [52], \(\Delta^{L}_t \in \mathbb{R}^3\) and \(\Delta^{R}_t \in \mathbb{R}^3\) are the object translation expressed in the left and right wrist frame respectively, \(\mathbf{p}^{G}_t \in \mathbb{R}^3\) represents the object translation in global coordinates and \(\boldsymbol{\alpha}_t \in \mathbb{R}^3\) are learnable mixture weights that balance the contribution of wrist relative displacements and global-coordinate translation.

Letting \(\mathbf{p}^{L}_t,\mathbf{p}^{R}_t \in \mathbb{R}^3\) denote the global wrist positions, and \(\mathbf{R}^{L}_t,\mathbf{R}^{R}_t \in SO(3)\) the corresponding global wrist rotations obtained from forward kinematics, we define the wrist-relative offsets, relative to the right (\(R\)) and left (\(L\)) hand: \[\begin{align} \Delta^{\{R,L\}}_t &= (\mathbf{R}^{\{R,L\}}_t)^\top (\mathbf{p}^{obj}_t - \mathbf{p}^{\{R,L\}}_t) \label{eq:deltas} \end{align}\tag{1}\] Intuitively, these terms describe where the object lies in each wrist’s local frame. To ensure accurate object translation in cases where the object is not interacting with either hand, we include \(\mathbf{p}^{G}_t\) as a global coordinate fallback. Under this formulation, we define the global object translation through a soft mixture of the three relative representations as: \[\hat{\mathbf{p}}^{obj}_t = w^L_t \left(\mathbf{p}^{L}_t + \mathbf{R}^{L}_t \Delta^{L}_t \right) + w^R_t \left(\mathbf{p}^{R}_t + \mathbf{R}^{R}_t \Delta^{R}_t \right) + w^G_t \mathbf{p}^{G}_t. \label{eq:trajdecode}\tag{2}\] where \([w^L_t,w^R_t,w^G_t] = \mathrm{softmax}(\boldsymbol{\alpha}_t)\) are the learned mixing weights.

Using this formulation, the model learns to smoothly interpolate between left-hand control, right-hand control, and global motion through the learned weights \([w^L_t, w^R_t, w^G_t]\), which adapt at each frame depending on which hand is in contact with the object, instead of forcing all interaction regimes into a single translation branch.

3.2 Motion Modeling↩︎

We formulate human-object motion synthesis as auto-regressive conditional denoising over structured motion sequences, enabling the generation of arbitrarily long sequences  [7], [55]. Given an observed prefix of length \(T_p\), the AR model \(D_{\theta}\) predicts the future segment \(\mathbf{F}_{T_p:T_p+T_f-1}\), with a prediction length of \(T_f\) frames.

3.2.0.1 Dual-Stream HOI Diffusion \(\mathbf{D}_{\mathbf{\theta}}\).

To facilitate motion generation for multiple objects, we adopt a cooperative denoising paradigm  [22] with two coordinated diffusion streams \(A\) and \(B\) with shared weights, one for each object-conditioned interaction. Stream \(A\) carries motion representation \(\mathbf{f}^A_t = [\mathbf{h}_t,\,\mathbf{r}_t,\,\mathbf{o}^{A}_t]\) and stream \(B\) carries \(\mathbf{f}^B_t =[\mathbf{h}_t,\,\mathbf{r}_t,\,\mathbf{o}^{B}_t]\), where both share the human representation \([\mathbf{h}_t, \mathbf{r}_t]\) but differ object representation. At diffusion timestep \(\tau\), the shared denoiser \(D_\theta\) processes each stream conditioned on the other: \[\hat{\mathbf{f}}^A_0 = D_{\theta}\!\left(\mathbf{f}^A_\tau,\,\mathbf{f}^B_\tau,\,\mathbf{c}^A,\,\tau\right), \qquad \hat{\mathbf{f}}^B_0 = D_{\theta}\!\left(\mathbf{f}^B_\tau,\,\mathbf{f}^A_\tau,\,\mathbf{c}^B,\,\tau\right), \label{eq:denoise}\tag{3}\] where \(\mathbf{c}^A\) and \(\mathbf{c}^B\) are branch-specific conditioning signals containing: a frozen CLIP  [56] text embedding; object geometry encoded as Basis Point Set (BPS) features  [57]; and a lightweight MLP projection for \(\tau\). To enable information flow between the two HOI streams, we follow  [22] and apply bidirectional cross-attention \(\mathrm{CA}(\cdot)\) after each layer \(\ell\) of the shared HOI Diffusion \(D_{\theta}\), resulting in each stream effectively attending to the other.

3.2.0.2 Motion Fusion Network \(\mathbf{D}_{\mathbf{h}}\).

At the output of \(D_{\theta}\), both streams produce a complete prediction \(\hat{\mathbf{f}}^{A}_0\) and \(\hat{\mathbf{f}}^{B}_0\), each containing a full motion features with human representation and hand-relative object representation per object present. We decompose each prediction along the feature partition introduced in Section 3.1: \[\hat{\mathbf{f}}^{A}_0 = \big[\, \hat{\mathbf{h}}^A,\; \hat{\mathbf{r}}^A,\; \hat{\mathbf{o}}^{A} \,\big], \qquad \hat{\mathbf{f}}^{B}_0 = \big[\, \hat{\mathbf{h}}^B,\; \hat{\mathbf{r}}^B,\; \hat{\mathbf{o}}^{B} \,\big].\] To produce a single coherent human pose, derived from the separate HOI Diffusion streams \([\hat{\mathbf{h}}^A,\hat{\mathbf{r}}^A]\) and \([\hat{\mathbf{h}}^B,\hat{\mathbf{r}}^B]\), we introduce a dedicated Motion Fusion Network \(D_{\mathbf{h}}\), implemented as a stack of self-attention layers that operates only on the human representations (Figure 3): \[\big[\hat{\mathbf{h}},\; \hat{\mathbf{r}}\big] = D_{\mathbf{h}}\!\left(\big[\hat{\mathbf{h}}^A,\hat{\mathbf{r}}^A\big],\; \big[\hat{\mathbf{h}}^B,\hat{\mathbf{r}}^B\big]\right).\] The fused body prediction \([\hat{\mathbf{h}}, \hat{\mathbf{r}}]\) replaces the per-stream body outputs, while the object predictions \(\hat{\mathbf{o}}^{A}\) and \(\hat{\mathbf{o}}^{B}\) are taken directly from their originating streams without modification. This design enforces a clean factorization of the generation process: each object’s dynamics are denoised under its own object-specific conditioning, while the body is reconciled in a shared latent that accesses both interactions. As we demonstrate through ablations in Section 4.3, the Motion Fusion Network outperforms simpler alternatives such as averaging or selecting one of the two human predictions, which often lead to inconsistent motions.

3.3 Training Objective↩︎

Since our HOI diffusion model solely predicts rotation parameters for the human body, we apply a differentiable forward kinematics layer \(\mathrm{FK}(\cdot)\) to map predicted poses to world-space joint positions, thus enabling velocity and contact losses that are otherwise inaccessible in the rotation domain. Along with the main diffusion objective \(\mathcal{L}_{diff}\) and the well established geometry regularization losses [6], [7], [22], which cover joint and position velocity and foot skating, and interpenetrations, we also enhance our training objective by introducing the following contact-aware conditioning:

3.3.0.1 Hand-relative object representation supervision.

Our hand-relative object representation \(\mathbf{o}_t\) induces a structured decomposition of object motion into wrist-relative and global branches, according to Eq. 2 . We explicitly supervise the different components. First, we supervise the mixture weights by comparing the predicted weights \([\hat{w}^L_t,\hat{w}^R_t,\hat{w}^G_t] = \mathrm{softmax}(\hat{\boldsymbol{\alpha}}_t)\), extracted from the predicted representation \(\hat{\mathbf{o}}t\), against hand proximity-derived ground-truth targets \([w^L_t,w^R_t,w^G_t]\):

\[\mathcal{L}_{mix} = \left\|\hat{w}^{ s}_t - w^{s}_t \right\|_2^2, \quad s \in \{L, R, G\}.\] Second, we supervise the decoded object translation \(\hat{\mathbf{p}}^{obj}_t\), obtained via the mixture decoder of Eq. 2 , against the ground-truth world-space position \(\mathbf{p}^{obj}_t\): \[\mathcal{L}_{trans}= \left\| \hat{\mathbf{p}}^{obj}_t - \mathbf{p}^{obj}_t \right\|_2^2,\] enforcing correct trajectory reconstruction while ensuring gradients propagate through all branches and mixture weights. Finally, we regularize the predicted wrist-relative offsets \(\hat{\Delta}^{L}_t\) and \(\hat{\Delta}^{R}_t\), as defined in Eq. 1 , against the ground-truth \({\Delta}^{L}_t\) and \({\Delta}^{R}_t\), ensuring further encouraging geometrical consistency: \[\mathcal{L}_{off}= \frac{1}{2} \left( \left\|\hat{\Delta}^{L}_t - \Delta^{L}_t\right\|_2^2 + \left\|\hat{\Delta}^{R}_t - \Delta^{R}_t\right\|_2^2 \right).\]

3.3.0.2 Hand-Object contact regression.

We enforce fine-grained interaction geometry via a unified finger-object distance loss. Letting \(d_{t,j}\) denote the ground-truth signed nearest-surface distance from hand joint \(j\) at frame \(t\) to the object mesh, and \(\hat{d}_{t,j}\) the predicted counterpart, we regress nearest-surface distances collectively via Huber loss \(\rho_\delta\): \[\mathcal{L}_{dist}= \frac{1}{|\mathcal{A}|} \sum_{(t,j)\in\mathcal{A}} \rho_\delta\left(\hat{d}_{t,j} - d_{t,j}\right),\] where \(\mathcal{A} = {(t,j) : d_{t,j} < d_{min}}\) selects frames where hand joint \(j\) lies within approach radius \(d_{min}=0.10\) of the object surface.

Overall, the full training objective is defined as: \[\mathcal{L} = \mathcal{L}_{diff} + \mathcal{L}_{vel} + \underbrace{ \lambda_{w}\mathcal{L}_{mix} + \lambda_{t}\mathcal{L}_{trans} + \lambda_{o}\mathcal{L}_{off} }_{\text{\emph{object representation supervision}}} + \lambda_{p}\mathcal{L}_{dist}.\]

4 Evaluation↩︎

4.1 Datasets↩︎

To train and evaluate Dex2HOI we use the following datasets:

GRAB. A dataset [16] with 10 subjects interacting with 51 everyday objects. We follow the typical split  [6], [7] using subjects 1 through 9 for training and subject 10 (145 sequences) as the test set.

HUMOTO. A mo-cap dataset [10] that contains both single and multi-object manipulation tasks in Mixamo representations [54]. We retain clips that contain up to two moving objects, splitting according to the provided annotations at a frame level. The resulting benchmark yields 337 train and 45 test sequences, mostly from tabletop manipulation scenes; the broader 1/2-object subset used for one-object training and evaluator training contains 1828 train and 322 test sequences.

HIMO. A multi-object interaction dataset  [9] covering simultaneous manipulation of 2 or 3 objects. The 2-object split contains 2578 train and 522 test sequences, and the 3-object split contains 1045 train and 176 test sequences, following the official splits  [9]. In this work, we use the 2-object split to evaluate our work. More information about the datasets can be found in the Appendix.

4.2 Motion Evaluators and Metrics↩︎

To evaluate the generated motions on the HUMOTO interactions, we trained a dedicated transformer-based motion evaluator on ground-truth motions of the human joints and of the two objects, following the protocol in [30]. The motion encoder is used to compute FID\(\downarrow\), Diversity\(\rightarrow\), and MultiModality (MM \(\uparrow\)) based on extracted motion embeddings and assess distributional quality, while for text-alignment metrics MMdist (MMd \(\downarrow\)) and R-Precision (R@3 \(\uparrow\)), we train an additional motion-text aligner using a contrastive objective. Finally, the penetration depth metric measures the mean hand and body–object interpenetration distance across frames.

To evaluate on GRAB we follow the existing protocol [7], [30]. In contrast to the HUMOTO evaluator that is trained using a contrastive learning objective, the GRAB evaluator serves as an action-recognition network, since GRAB comes with discrete action labels [16]. We report the same six metrics across all GRAB experiments, with the difference that MultiModality is calculated between generations of the same action labels instead of instruction prompts, following [30].

4.3 Baselines and Ablations↩︎

4.3.0.1 Baselines.

On the two-object benchmarks HUMOTO and HIMO, we compare Dex2HOI against the HIMO generative baseline [9] and an MDM baseline [18], which we adapt to the HOI task by extending its representation to jointly predict object and human motion. Given the limited availability of multi-object evaluation benchmarks and to also evaluate our separate HOI Diffusion Models \(D_{\theta}\), we perform additional evaluation on the single-object against previous state-of-the-art full-body HOI methods, namely HOIDiNi [7], CoDA [6] and IMoS [30] and MDM [18].

The aforementioned methods, besides MDM, contain some type of hand-object contact informed optimization at inference, which generates plausible results, yet it introduces an inference time overhead. Additionally, adapting these test-time optimization frameworks to two-object sequences is a non trivial task; Their contact objectives are formulated around a single object–hand pair, with no mechanism to resolve which hand should contact which object at each timestep. Naively running per-object optimization in parallel compounds without providing any coordination signal risks yielding inter-object penetrations and incoherent motion. As shown in Table 1, Dex2HOI achieves the best FID, Diversity and R@3 on GRAB while running up to \(\times\)​540 times faster on average than CoDA and \(\times\)​490 times than HOIDiNi, depending on the selected duration of the iterative noise optimization (we use 200 opt. steps on both methods over 60-frame predictions). Notably, our single-shot generation surpasses or matches optimization-based methods across standardized metrics, without the exhaustive runtime requirements. Similarly, on HUMOTO dataset (Table 2), Dex2HOI achieves the best FID, penetration, and MModality across both single- and two-object settings. The largest gain over prior methods is observed on the single-object split, with substantial FID improvement.

4.3.0.2 Ablation Study.

We design our ablation study around the core components of Dex2HOI and we provide the following ablation variants for two-obj evaluation: (i) w/o hand-relative representation, where we replace the proposed representation with global object trajectory prediction. (ii) w/o dual-stream architecture, where we use a single denoising stream that jointly predicts human motion together with both objects. (iii) w/o Motion Fusion Network, where we replace the learned \(D_{\mathbf{h}}(\cdot)\) model with an element-wise average of the two stream body predictions. (iv) w/o mixture supervision , where we remove the \(\mathcal{L}_{\text{mix}}\), \(\mathcal{L}_{\text{trans}}\), and \(\mathcal{L}_{\text{off}}\) terms, leaving only the diffusion MSE loss on the hand-relative object representation. This tests whether the hand-relative representation can be learned implicitly, or whether weight supervision is required. (v) w/o finger-contact supervision. where we remove the \(\mathcal{L}_{\text{dist}}\), and finally (vi) GRAB single prior. As a final sanity check, we perform transfer learning by initializing our HOI model \(D_{\theta}\) with diffusion layers pretrained exclusively on single-object GRAB sequences, which are kept frozen. We retrain only the bidirectional cross-attention and Motion Fusion Network \(D_{h}\) using two-object data. At test time, the model generates two-object interactions, allowing us to assess whether bimanual coordination can be achieved by relying on fixed single-object priors. Additionally, we evaluate on the GRAB dataset on (i) w/o hand-relative representation and (ii) w/o finger-contact supervision settings. Ablation results for all variants are reported in Table 3, and qualitative hand-representation and \(L_{Dist}\) comparisons in Figure 6. Qualitative results are also shown in Figures 4 and 5 for GRAB and HUMOTO respectively. For further evaluation analysis, please refer to the Appendix.

4.4 VLM assessment and User Study↩︎

We complement our quantitative assessment with two perceptual evaluations. Inspired by the VLM evaluation protocol of [58], we query InternVL3-8B [59] with three targeted prompts and the rendered videos, rating motion naturalness, hand–object contact quality and text prompt alignment on a scale of 1 to 5. We report per-question mean scores across all test sequences, with error bars denoting std over 10 runs. For the user study, 30 participants are shown anonymized side-by-side renderings of the methods and rank them using the same criteria. As shown in Figure 7, Dex2HOI consistently achieves superior interaction scores, in both single- and two-object benchmarks, demonstrating more natural motions and better contacts than competing baselines, further validated by our user study.

5 Discussion and Conclusion↩︎

We presented Dex2HOI, a unified framework for human-object interaction generation that extends beyond conventional single-object HOI to coordinated bimanual dexterous manipulation of up to two objects. We model multi-object interactions through dual denoising streams, coupled via bidirectional cross-attention and fused into a single coherent full-body motion. By modeling object motion through our hand-relative representation, we unify single-hand and bimanual manipulation within the same formulation, producing plausible hand-object interactions in single-shot inference, with no additional contact optimization. Our extensive evaluation shows that Dex2HOI generalizes across both single-object and multi-object settings, achieving SOTA results and improved motion realism.

Limitations. Despite Dex2HOI’s promising results, certain challenges remain. While our architecture supports long sequence synthesis through autoregressive sampling, the diversity and complexity of the motions remain bounded by the bimanual interactions present in current 4D HOI datasets, which are limited. Future work should address long-horizon action planning and the development of standardized benchmarks for scaling multi-object bimanual manipulation.

References↩︎