SimGenHOI: Realistic Humanoid-Object Interaction

• The paper introduces SimGenHOI, a unified framework combining diffusion-based generative modeling and reinforcement learning to generate physically realistic humanoid-object interactions.
• It employs a three-stage pipeline—key action generation, trajectory interpolation, and a contact-aware RL torque policy—to ensure robust simulation performance.
• The framework’s mutual fine-tuning strategy bridges motion diversity with executable control, significantly reducing artifacts like penetration and foot sliding.

Searching arXiv for SimGenHOI and closely related HOI generation papers. SimGenHOI is a unified framework for physically realistic whole-body humanoid-object interaction that combines generative modeling and reinforcement learning to produce controllable and physically plausible HOI (Lin et al., 18 Aug 2025). In this framework, humanoid-object interaction denotes full-body motions of a simulated humanoid, based on SMPL-X, interacting with 3D objects in a physics simulator, with outputs that include humanoid motion, object motion, and contact patterns over long horizons (Lin et al., 18 Aug 2025). The system is conditioned on a text prompt, object geometry, sparse object waypoints, and the initial humanoid pose, and aims to bridge the gap between the diversity of diffusion-based motion generation and the executability of torque-controlled physical simulation (Lin et al., 18 Aug 2025).

1. Problem formulation and scope

SimGenHOI addresses the problem of generating physically executable HOI sequences in simulation, including key actions, interpolated dense trajectories, and a torque policy that realizes them in Isaac Gym (Lin et al., 18 Aug 2025). The target behaviors include daily manipulation and dynamic interactions such as picking up boxes, carrying, pushing, and placing, over many seconds (Lin et al., 18 Aug 2025). The central difficulty is that purely generative HOI models often exhibit implausible contacts, penetration, foot sliding, nonsensical support, and unrealistic whole-body motion, while physics-only controllers lack a high-level motion blueprint that is simultaneously controllable and semantically conditioned (Lin et al., 18 Aug 2025).

The framework is explicitly simulation-centric. Experiments are conducted in NVIDIA Isaac Gym with an SMPL-X-based humanoid and FullBodyManipulation objects, and the method is presented as a simulation-only system rather than a real-world deployment pipeline (Lin et al., 18 Aug 2025). This distinguishes it from humanoid HOI systems such as GenHOI, which reconstruct a robot-object scene in MuJoCo and execute optimized trajectories on a Unitree G1 platform (Bi et al., 11 Jun 2026). It also differs from PhyGenHOI, which models the human as a semantic agent and the object as a physical agent under a unified 3D Gaussian Splatting representation for 4D HOI generation (Benishu et al., 28 May 2026).

A plausible implication is that SimGenHOI occupies the intersection of text-conditioned generative motion synthesis and physically grounded control: it is neither a purely kinematic video-conditioned generator nor a purely task-specific policy learner. This positioning is central to its design.

2. System architecture

SimGenHOI consists of two tightly coupled components: an HOI generative model based on a Diffusion Transformer, and a contact-aware whole-body RL policy (Lin et al., 18 Aug 2025). The generative model predicts humanoid pose sequences, object pose sequences, and contact probabilities conditioned on text, geometry, masked motion, and diffusion timestep information (Lin et al., 18 Aug 2025). The control policy receives current humanoid and object states, reference future states from the generated trajectory, contact guidance, and the object geometry embedding, and outputs joint torques for the humanoid in Isaac Gym (Lin et al., 18 Aug 2025).

The overall inference pipeline proceeds in three stages (Lin et al., 18 Aug 2025). First, the diffusion model samples key actions autoregressively for long-horizon generation. Second, these key actions are interpolated into dense sequences and converted from a relative representation to a global simulation representation, with velocities computed by finite differences. Third, the HOI policy tracks the resulting motion in closed loop by applying torques at each simulation step (Lin et al., 18 Aug 2025).

This decomposition separates high-level motion planning from low-level physical realization. Related systems use different decompositions. PhyGenHOI separates a semantic human agent driven by a Motion Diffusion Model from a physical object agent simulated by the Material Point Method, and then couples them through windowed attraction, contact-driven re-simulation, and masked Video-SDS (Benishu et al., 28 May 2026). GenHOI uses a generated video as a high-level interaction prior and converts it into object-centric geometric constraints for humanoid trajectory optimization without task-specific training (Bi et al., 11 Jun 2026). SimGenHOI instead places the coupling between generation and control at the level of a learned torque policy and an iterative mutual fine-tuning loop (Lin et al., 18 Aug 2025).

3. Generative model and conditioning

The generative model uses a DDPM-style conditional diffusion process implemented as a DiT denoiser (Lin et al., 18 Aug 2025). The joint human-object-contact trajectory is denoted by $\mathbf{\tau}_0$, and the forward process is

$$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$

with reverse process

$$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$

and a training objective that predicts clean data with an $L_1$ loss (Lin et al., 18 Aug 2025):

$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$

The conditioning vector $\mathbf{C}$ comprises four elements (Lin et al., 18 Aug 2025). The first is an object geometry embedding derived from a Basis Point Set representation $\mathbf{G} \in \mathbb{R}^{1024 \times 3}$ and projected by an MLP to $\mathbf{\hat G} \in \mathbb{R}^{256 \times 3}$. The second is masked motion $\mathbf{S} \in \mathbb{R}^{T \times (12 + D)}$, which contains the initial humanoid pose and object pose, sparse object waypoints in $x,y$, the final target object position in $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$0, and masks for in-filling or refinement. The third is a text embedding $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$1 from a pre-trained CLIP model. The fourth is a noise-level embedding that is fused with the text embedding (Lin et al., 18 Aug 2025).

The model predicts three synchronized streams (Lin et al., 18 Aug 2025):

• Humanoid motion: $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$2.
• Object motion: $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$3.
• Contact probabilities: $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$4 for left and right hands and feet.

This conditioning design gives SimGenHOI a degree of controllability absent from methods that rely solely on free-form text. SimGraph provides an analogous argument in the image domain, where scene graphs serve as structured controls for generation and editing rather than unconstrained prompts (Vo et al., 29 Jan 2026). A plausible implication is that SimGenHOI’s masked motion and waypoint channels function as an HOI-specific control interface for object transport, timing, and endpoint specification.

4. Key actions, interpolation, and long-horizon generation

A defining property of SimGenHOI is its use of key actions rather than dense framewise prediction (Lin et al., 18 Aug 2025). A key action is a frame at which motion and/or contact changes significantly, such as the onset of grasp, lift, place, or regrasp (Lin et al., 18 Aug 2025). The diffusion model is trained to generate these sparse events, which are later interpolated into dense trajectories.

Key actions are extracted recursively from dense motion sequences (Lin et al., 18 Aug 2025). Starting from a dense sequence of joint positions $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$5, with $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$6, the method initializes the key-action set with the first and last frames, interpolates between currently selected key actions, computes a weighted reconstruction error, adds the frame with maximum error, and repeats until every segment falls below a threshold (Lin et al., 18 Aug 2025). The reconstruction error is

$$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$7

where $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$8 is Euclidean distance and $$q(\mathbf{\tau}_n \mid \mathbf{\tau}_{n-1}) = \mathcal{N}\big( \mathbf{\tau}_n; \sqrt{1 - \beta_n}\,\mathbf{\tau}_{n-1}, \beta_n \mathbf{I} \big),$$9 gives higher importance to critical joints such as hands and feet (Lin et al., 18 Aug 2025).

After key-action generation, dense whole-body trajectories are reconstructed by interpolation in joint position and rotation space (Lin et al., 18 Aug 2025). Positions are linearly interpolated; rotations are interpolated in a continuous 6D representation; and contact probabilities are interpolated or treated as per-frame guidance signals (Lin et al., 18 Aug 2025). Long-horizon generation is achieved by autoregressive sampling over overlapping windows, with the last frames of the previous segment provided through masked motion to enforce temporal coherence (Lin et al., 18 Aug 2025).

This architecture is important because it reduces the number of diffusion calls required for long-horizon HOI (Lin et al., 18 Aug 2025). It also creates a temporal abstraction distinct from the coupling strategies of related methods. PhyGenHOI performs timing-sensitive synchronization by identifying a contact joint and contact frame and then applying a Windowed Attraction Loss around that temporal neighborhood (Benishu et al., 28 May 2026). GenHOI, by contrast, recovers a single reference trajectory from a generated video and then applies terminal-state optimization plus smooth correction over the last $$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$0 frames (Bi et al., 11 Jun 2026). SimGenHOI’s key-action formalism is more native to long-horizon autoregressive generation than either of those alternatives.

5. Contact-aware whole-body control policy

The control component is a contact-aware whole-body RL policy that tracks the interpolated reference motion in simulation while correcting artifacts such as penetration, foot sliding, unstable support, and contact inconsistencies (Lin et al., 18 Aug 2025). Humanoid state is represented as

$$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$1

and object state as

$$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$2

(Lin et al., 18 Aug 2025). The policy input includes humanoid proprioceptive state, reference humanoid tracking state, current object observation, reference object state with future guidance, current contact state, predicted future contact state, and the object geometry embedding (Lin et al., 18 Aug 2025). The action is a vector of joint torques applied in Isaac Gym (Lin et al., 18 Aug 2025).

The control policy is built on the latent motion representation of PULSE and is trained with rewards that combine humanoid tracking, object tracking, and contact consistency (Lin et al., 18 Aug 2025). The humanoid tracking and contact reward is

$$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$3

while the object tracking reward is

$$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$4

The total reward is

$$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$5

(Lin et al., 18 Aug 2025).

The method also extends early termination to HOI-specific failure modes (Lin et al., 18 Aug 2025). Episodes terminate if average object keypoint deviations exceed $$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$6 m or if expected body-object contact is missing for more than 10 consecutive frames (Lin et al., 18 Aug 2025). This filters physically invalid interactions during training.

In comparison with GenHOI, which uses object-centric geometric constraints and terminal inverse kinematics to adapt trajectories, SimGenHOI incorporates contact guidance directly into the RL state and reward (Bi et al., 11 Jun 2026). Compared with PhyGenHOI, which explicitly simulates object dynamics and injects video priors near contact, SimGenHOI delegates physical correction to a torque policy trained in a rigid-body simulator (Benishu et al., 28 May 2026).

6. Mutual fine-tuning and physical realism

A distinctive element of SimGenHOI is its mutual fine-tuning strategy, in which the generator and controller iteratively improve one another (Lin et al., 18 Aug 2025). The process begins with a generative model trained on mocap data and a controller trained to track generated trajectories in simulation (Lin et al., 18 Aug 2025). Successful executions are then collected, key actions are re-extracted from these physically plausible trajectories, and the DiT is fine-tuned on them with the same $$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$7 objective (Lin et al., 18 Aug 2025). The improved generator is used to produce new trajectories that are easier to track, and the controller is correspondingly fine-tuned (Lin et al., 18 Aug 2025).

This procedure is intended to align the distribution of generated motion with the capabilities of the controller and the constraints of the simulator (Lin et al., 18 Aug 2025). Without such alignment, diffusion-generated trajectories may remain difficult to realize, even if they appear plausible in a kinematic sense. The method therefore treats physical validity not merely as a post hoc filtering criterion but as a source of new supervisory data.

The framework measures physical realism with foot height $$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$8, foot sliding FS, and hand penetration $$p_\theta(\mathbf{\tau}_{n-1} \mid \mathbf{\tau}_n, \mathbf{C}) = \mathcal{N}\big( \mathbf{\tau}_{n-1}; \mathbf{\mu}_\theta(\mathbf{\tau}_n, n, \mathbf{C}), \mathbf{\Sigma}_n \big),$$9 based on a Signed Distance Field around the object (Lin et al., 18 Aug 2025). The generation-only model improves condition matching and interaction quality relative to baselines, but tracked motions show the most pronounced gains in physical consistency (Lin et al., 18 Aug 2025). After tracking in simulation, SimGenHOI reports $L_1$0 mm, FS $L_1$1, and $L_1$2 (Lin et al., 18 Aug 2025). These numbers are presented as evidence that the combined generative-plus-control pipeline is necessary for eliminating penetration and reducing sliding.

A plausible interpretation is that mutual fine-tuning operationalizes a closed-loop form of dataset repair: generated motions are filtered through a physics-based controller, and successful realizations are fed back into the generator. This contrasts with GenHOI, which avoids task-specific training altogether and instead relies on one-shot video generation plus optimization (Bi et al., 11 Jun 2026). It also contrasts with PhyGenHOI, where the interaction refinement loop is realized through explicit physical re-simulation and masked score distillation rather than RL (Benishu et al., 28 May 2026).

7. Experimental evaluation, comparisons, and limitations

SimGenHOI is evaluated on the FullBodyManipulation (OMOMO) dataset, which contains 10 hours of high-quality human-object interaction with 15 objects, using data from 15 subjects for training and 2 held-out subjects for testing (Lin et al., 18 Aug 2025). The evaluation covers condition matching, interaction quality, deviation from ground truth, physical consistency after tracking, interaction tracking performance, and ablations on key actions, contact guidance, and mutual fine-tuning (Lin et al., 18 Aug 2025).

Quantitative results

Evaluation setting	Metric	Reported result
Generation-only	$L_1$3	$L_1$4–$L_1$5
Generation-only	$L_1$6	$L_1$7–$L_1$8
Generation-only	$L_1$9	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$0–$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$1
Generation-only	MPJPE	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$2–$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$3 mm
Tracked in simulation	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$4	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$5 mm
Tracked in simulation	FS	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$6
Tracked in simulation	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$7	$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$8
Interaction tracking	Succ$$\mathcal{L} = \mathbb{E}_{\mathbf{\tau}_0, n} \big\| \hat{\mathbf{\tau}}_\theta(\mathbf{x}_n, n, \mathbf{C}) - \mathbf{\tau}_0 \big\|_1.$$9	$\mathbf{C}$0
Interaction tracking	Succ$\mathbf{C}$1	$\mathbf{C}$2

In generation-only comparisons, SimGenHOI outperforms InterDiff, MDM, OMOMO variants, and CHOIS on condition matching, contact precision, recall, $\mathbf{C}$3, contact percentage, MPJPE, root translation error, object error, and object orientation error (Lin et al., 18 Aug 2025). In the tracked setting, it surpasses OMOMO ground truth tracked directly and CHOIS in foot height, foot sliding, and hand penetration (Lin et al., 18 Aug 2025). On 7 OMOMO motions selected as in OmniGrasp, it achieves perfect contact and target success, though OmniGrasp has a smaller object position error while SimGenHOI reports a smaller object velocity error (Lin et al., 18 Aug 2025).

The ablation study on 144 small-box sequences shows that removing key actions, removing contact guidance, or removing fine-tuning degrades success and tracking quality (Lin et al., 18 Aug 2025). The full model achieves the highest success rate at $\mathbf{C}$4 and the lowest object rotation error at $\mathbf{C}$5 rad among the listed ablations (Lin et al., 18 Aug 2025). These results are presented as evidence that all three components contribute materially to performance.

Relation to adjacent work

SimGenHOI sits within a broader set of recent HOI systems, but its technical profile is distinct. PhyGenHOI targets physically accurate and visually faithful 4D HOI generation from static 3D humans and objects represented as 3D Gaussian Splats, using MDM, MPM, and video diffusion guidance (Benishu et al., 28 May 2026). GenHOI focuses on zero-shot humanoid-object interaction by imitating a single generated video and refining it through contact-aware geometric optimization without task-specific policy training (Bi et al., 11 Jun 2026). SimGraph is not an HOI generator in the same sense, but it demonstrates a parallel trend toward structured control interfaces, using scene graphs to unify generation and editing in images (Vo et al., 29 Jan 2026). SimGenHOI differs from all three by combining text-conditioned key-action generation, whole-body torque control, and iterative generator-controller alignment in simulation (Lin et al., 18 Aug 2025).

Limitations

The reported limitations are explicit (Lin et al., 18 Aug 2025). The framework is simulation-only, and real-world transfer is not addressed. The generative model does not produce detailed finger articulation for complex grasps; instead, RL learns coarse contact behaviors suited mainly to non-dexterous manipulation. Diffusion sampling with DiT and autoregression is computationally heavy, and RL training remains expensive despite Isaac Gym acceleration. Generalization beyond the FullBodyManipulation dataset and its 15 objects is not established. Multi-object and more complex multi-contact scenarios are also outside the demonstrated scope (Lin et al., 18 Aug 2025).

These limits are informative when contrasted with related systems. GenHOI demonstrates real-world transfer to a Unitree G1 but uses local optimization and sparse point contacts rather than whole-body RL (Bi et al., 11 Jun 2026). PhyGenHOI models object deformation and causal momentum transfer but does not address torque-level execution on humanoid robots (Benishu et al., 28 May 2026). This suggests that “physical realism” remains method-dependent: in SimGenHOI it is primarily a property of simulation-consistent control and reduced artifacts under rigid-body dynamics, rather than photorealistic 4D rendering or hardware robustness.

SimGenHOI therefore represents a specific synthesis within the current HOI literature: a physically grounded, simulation-only, text-conditioned whole-body generation framework in which structured key-action diffusion and contact-aware RL are mutually optimized to improve executability and interaction fidelity (Lin et al., 18 Aug 2025).