acceptodds
Under review as a conference paper at ICLR 2027

From Digital Human Interactions to Physics-Based Humanoid Skills: Physics-Grounded Post-Training of Interaction Generators

Abstract

Recent methods have made promising progress in generating interactions between two humanoids, largely relying on physics-based tracking policies to convert digital reference motions into executable trajectories. However, limited tracking capabilities restrict the range of reference motions that can be successfully executed, reducing data utilization. Moreover, even successful tracking does not guarantee physically plausible responses or faithful realization of the intended interactions. In this paper, we introduce DIGHT, a co-adaptive framework that couples a Digital human Interaction Generator with a Humanoid Tracking policy. Our DIGHT first executes multiple text-conditioned interaction candidates in simulation using a fixed tracker. It then constructs physics-grounded preferences from the resulting rollouts, covering both general executability and interaction fidelity. Rather than collapsing these signals into a single scalar reward for candidate ranking, we align the pretrained generator using physics-decoupled diffusion direct preference optimization (DPO), preserving criterion-specific supervision without differentiating through the simulator. To improve executability, preference pairs are derived from tracking error, friction, and floating. Additionally, to improve interaction fidelity, we propose to incorporate force feedback from simulator as a measure of contact fidelity and construct preferences over contact occurrence, location, duration, and force magnitude. The aligned generator then supplies reference motions for fine-tuning the tracker, improving compatibility between generation and physical execution. Extensive experiments demonstrate that our approach not only improves the physical plausibility of generated motions but also enables more reliable and faithful humanoid interactions in simulation.

Primary area:
generative models
Keywords:
motion generation, motion interaction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AQfJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

When executing generated human motions in a physics simulator, physical constraints can make it challenging to faithfully reproduce the intended motions. This paper proposes DIGHT, a framework that encourages a human motion generator to produce motions that a humanoid tracking policy can execute while preserving the intended interactions.

Specifically, DIGHT fine-tunes the generator using direct preference optimization (DPO) to produce motions that are easier for the tracker to execute while maintaining interaction fidelity. The tracker is subsequently adapted to better follow the motions produced by the aligned generator. The authors claim that their approach generates more physically executable motions than the compared methods considered in the paper.

Critical strengths

The paper presents a framework for generating two-person interactions that can be more reliably tracked and executed in a physics simulator.

The proposed preference design explicitly considers contact occurrence, topology, duration, and force between two humanoids.

Critical weaknesses
  1. The novel contributions of this paper beyond existing methods are not sufficiently clear.

Several key components of the proposed framework appear to have already been explored in prior work. Specifically, adapting a human motion generator to the capabilities of a humanoid tracker has been investigated in PhysMoDPO, while physics-decoupled diffusion DPO and the use of multiple physics-grounded preferences appear closely related to PhysMoDPO and BideDPO. Furthermore, the idea of jointly improving a generator and a tracker has already been explored in works such as PARC.

As it stands, the main contribution of this paper appears to be the integration of these existing techniques into a framework for two-humanoid interaction generation, along with the introduction of contact-related preferences.

  1. The experimental results do not sufficiently substantiate the contributions specific to this paper.

First, the comparison between InterGen + DPO and DIGHT is difficult to interpret because the paper does not clearly explain the specific differences between the two methods. Without a precise description of how the standard DPO baseline is implemented and how it differs from DIGHT, it is unclear what conclusions can be drawn from this comparison. At a minimum, the authors should explicitly clarify the differences between the DPO baseline and the proposed approach.

Second, Tables 1, 2, 3, and 4 lack metrics that directly evaluate interaction and contact fidelity, such as contact topology precision and recall. Given that the primary contribution specific to this paper appears to lie in modeling two-person interactions and contact fidelity, these metrics are essential for validating the effectiveness of the proposed approach.

Although Force Error is reported, this metric only measures the discrepancy in mean contact force between generated and ground-truth executions. It does not capture whether contacts occur at the correct body regions or whether their durations are accurately reproduced. Therefore, the current evaluation metrics are insufficient to demonstrate improvements in the interaction-specific aspects of the proposed method.

In particular, the results in Tables 1–4 mainly demonstrate improvements in general motion quality and physical executability, which have already been addressed in prior work such as PhysMoDPO. They do not provide sufficient evidence that the contributions specific to this paper lead to better interaction fidelity.

Third, Table 5 reports contact-related metrics, including contact topology precision and recall, and compares InterGen against DIGHT. However, this comparison evaluates the combined effects of the entire framework, including components derived from existing methods, rather than isolating the effects of the proposed contact-related preferences. Consequently, it is difficult to determine how much of the improvement can be attributed to the contributions specific to this paper.

Questions for authors

Please refer to the Critical Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer AQfJ

We thank the reviewer for recognizing the more reliable execution of generated two-person interactions in simulation and the value of our contact-aware preference design. We address the specific concerns below.

W1: Positioning and contribution.

Our central contribution is introducing inter-person contact forces for text-to-interaction generation. Making plausible interaction and contact as intended is one of the most important aspects of two-person motion. However, existing methods such as PhysMoDPO and PARC lack mechanisms to ensure the fidelity of complex interaction. Considering only whether the single motion is physically executable can lead to the failures shown in Fig. 1 of the main paper. The intended contact may fail to occur, or it may occur with excessive force destabilizing the humanoids.

Our key insight is the inter-person contact force can directly reveal whether contact occurs, which body parts make contact, and what forces arise. These signals are essential for improving interaction fidelity. Using DPO to incorporate non-differentiable feedback into training is common practice. Our main different with existing methods (e.g., PhysMoDPO) is to use inter-person contact forces to construct contact preference pairs that directly reflect how the two people interact, thereby improving interaction fidelity.

Through introducing the contact preferences, our method improves contact topology precision from 38.29% to 46.77% and recall from 52.16% to 63.94% on contact-annotated benchmark. Qualitative comparisons in Fig.1, Fig. 3, Appendix A, and the videos in supplementary material also illustrate the improvement in interaction fidelity.

W2-1: DPO baseline and additional comparisons.

In Table 1, InterGen + DPO normalizes each of the five physics scores across the K candidates for each prompt, then sums them to rank candidates and form preference pairs.

We additionally compare DIGHT with two baselines to evaluate our optimization design: a) DIGHT (one group). We retain criterion-specific preference objectives but jointly normalizes all five loss weights without grouping; b) PhysMoDPO. We adapt PhysMoDPO to two-person interactions using the same reward it used. The results are as follow:

MethodFID↓Skate↓Float↓Interpen↓Jerk↓Force Error↓
InterGen1.5397.2722.4010.69148.96577.41
InterGen + DPO1.5016.0420.969.85134.82491.40
DIGHT (one group)1.4866.2819.419.03126.56455.22
PhysMoDPO1.5165.6119.6710.32128.86523.33
DIGHT1.4745.6718.588.42111.94427.98

Compared with the other optimization designs, DIGHT achieves the best performance.

W2-2: Contact fidelity.

We additionally evaluate these methods on our contact-annotated subset, and report the contact topology precision and recall. For InterAgent, Succ. reports only the non-fall rate, as it has no reference trajectory for computing body deviation.

For Table 1:

MethodTopo. PrecTopo. RecallSucc.
InterGen38.2952.1684.70
InterGen + SFT35.6357.1185.49
InterGen + DPO43.9555.0686.68
DIGHT48.3160.8388.47

For Table 2:

MethodTopo. PrecTopo. RecallSucc.
InterAgent27.6358.4577.14
DIGHT46.7763.9489.23

For Table 3:

RewardTopo. PrecTopo. RecallSucc.
-38.2952.1684.70
trk41.7253.8786.88
+fric42.5351.7687.48
+float42.1253.2587.48
+contact48.3160.8388.47

For Table 4:

MethodTopo. PrecTopo. RecallSucc.
InterGen38.2952.1684.70
DIGHT*48.3160.8388.47
DIGHT46.7763.9489.23
AuthorsOpenReview ↗

Response to Reviewer AQfJ

W2-3: Metrics beyond mean contact force for evaluating interaction fidelity.

To demonstrate the interaction fidelity improvement, we introduce contact topology precsion and recall in main paper (Tab. 5), which evaluates the correctness of contact regions. Topology precision measures the fraction of realized contact pairs that are intended, while recall measures the fraction of intended contact pairs that are realized. More evaluating results of additional baselines are shown in W2-2.

MethodTopo. PrecTopo. RecallSucc.
InterGen38.2952.1684.70
DIGHT46.7763.9489.23

Additionally, to assess whether the durations are accurately reproduced and how accurately generated motions reproduce contact forces at individual joints, we introduce two additional metrics Duration and Joint Force Error. Specifically, for duration metric, we compute the absolute difference between the longest inter-person contact episode in the generated and reference executions. For joint force error metric, we compare the mean contact force at each joint between the generated and ground-truth executions and averaging the absolute differences across joints. The result is provided below:

MethodFIDMM-DistR-Precision (Top-1)SkateFloatInterpenJerkForce ErrorJoint Force ErrorDuration
InterGen1.5390.9540.4747.2722.4010.69148.96577.4138.130.24
DIGHT1.4740.9420.4865.6718.588.42111.94427.9827.570.17

DIGHT improves on all metrics, showing that our approach improves interaction fidelity.

W2-4: Evidence for improved interaction fidelity and comparison with PhysMoDPO

To evaluate whether our method improves interaction fidelity, we introduce Force Error, Topology Contact Precision, and Topology Contact Recall in the main paper. In W2-3, we introduce two additional metrics related to interaction fidelity. More quantitative evidences are provided in W2-2 and W2-3.

We also adapt PhysMoDPO to two-humanoid interactions and evaluate it under matched settings:

MethodTopo. PrecTopo. RecallSucc.Force Error
PhysMoDPO42.9450.5087.08523.33
DIGHT w/o tracker adapted48.3160.8388.47427.47
DIGHT46.7763.9489.23413.47

Under matched settings, both DIGHT variants achieve better interaction fidelity than PhysMoDPO.

W2-5: Effect of contact preferences.

To isolate the contribution of contact preferences, we report the topology precision and recall for different reward combinations:

RewardTopo. PrecTopo. RecallSucc.
-38.2952.1684.70
trk41.7253.8786.88
+fric42.5351.7687.48
+float42.1253.2587.48
+contact48.3160.8388.47
+tracker adapted46.7763.9489.23

Adding contact preferences significantly improves topology precision and recall.

Dear Reviewer AQfJ,

We would like to follow up and ask whether our responses have addressed your concerns. We would be happy to discuss any remaining issues. If our responses have resolved your concerns, we would be grateful if you could update your assessment accordingly.

Thank you for your time!

Best regards,

Authors of Submission #5253

Reviewer_vyEGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets the limited tracking capabilities and physically plausible responses in generated interactions between two humanoids. The proposed method includes generating multiple text-conditioned interaction candidates and constructing physics-grouded preferences with physics-decoupled diffusion direct preference optimization to align the pretrained generator. In addition, processing of preference pair selection and force feedback incorporation are introduced to improve executability and interaction fidelity. The reference motions from the aligned generator is then used for fine-tuning the tracker. Experiments are conducted to validate the proposed method.

Critical strengths
  • A decoupled preference optimization is proposed to finetune digital interaction generator with evidence of performance improvement to demonstrate the effectiveness.
Critical weaknesses
  • The proposed work is finetuning existing models like InterGen with their generated results and DPO based optimization borrowed from other similar work. The overall contribution is limited in this aspect.
  • The ablation study does not well justify the need of the proposed method. The idea of generating candidates and then selected preferred ones to finetune the generator and tracking policy is acknowledged and straightforward. The improvement over the original model is expected. However, it would be critical to demonstrate that which design of the proposed method contribute to the improvement and to which extent. That said, the comparison AND discussion against other improvements on InterGen, including the SFT and DPO in Table 1 and other methods using multiple candidates to finetune the models.
Questions for authors
  • see the second weakness
  • How does the proposed method performs compared with the best performance with K motions generated by InterGen? I understand the proposed finetuned model only need to generate one motion in inference. But it would be good to have such a comparison.
  • Since the K candidates are generated by existing models, the limitations of using such candidates to finetune the model should be discussed.
2 comments
AuthorsOpenReview ↗

Response to Reviewer vyEG

We thank the reviewer for recognizing the effectiveness of our preference optimization. We address the specific concerns below.

W1: Positioning and contribution.

Our central contribution is introducing inter-person contact forces for text-to-interaction generation. Making plausible interaction and contact as intended is one of the most important aspects of two-person motion. However, existing methods such as PhysMoDPO lack mechanisms to ensure the fidelity of complex interaction. Considering only whether the single motion is physically executable can lead to the failures shown in Fig. 1 of the main paper. The intended contact may fail to occur, or it may occur with excessive force destabilizing the humanoids.

Our key insight is the inter-person contact force can directly reveal whether contact occurs, which body parts make contact, and what forces arise. These signals are essential for improving interaction fidelity. Using DPO to incorporate non-differentiable feedback into training is common practice. Our main different with existing methods is to use inter-person contact forces to construct contact preference pairs that directly reflect how the two people interact, thereby improving interaction fidelity.

Through introducing the contact preferences, our method improves contact topology precision from 38.29% to 46.77% and recall from 52.16% to 63.94% on contact-annotated benchmark. Qualitative comparisons in Fig.1, Fig. 3, Appendix A, and the videos in supplementary material also illustrate the improvement in interaction fidelity.

W2: Ablation of DIGHT’s design choices

To better justify the need for our proposed method, we add additional ablation experiments that quantify the contribution of each design choice.

First, we compare DIGHT with three alternatives to evaluate our optimization design: a) InterGen+SFT. We fine-tune InterGen on the winners from each DIGHT preference set; b) InterGen+DPO. We normalize each of the five physics scores across the KK candidates for each prompt, then sums them to rank candidates and construct preference pairs; c) DIGHT (one group). We retain criterion-specific preference objectives but jointly normalizes all five loss weights without grouping. The results are as follows:

MethodSkate↓Float↓Interpen↓Jerk↓Force Error↓
InterGen7.2722.4010.69148.96577.41
InterGen + SFT6.5721.5413.12129.13543.62
InterGen + DPO6.0420.969.85134.82491.40
DIGHT (one group)6.2819.419.03126.56455.22
DIGHT5.6718.588.42111.94427.98

Leveraging the same KK candidates, our method achieves the best performance across all metrics.

Second, to assess whether contact preferences improve interaction fidelity, we evaluate different preference combinations on our contact-annotated benchmark. The results are shown below:

RewardTopo. PrecTopo. RecallSucc.
-38.2952.1684.70
trk41.7253.8786.88
+fric42.5351.7687.48
+float42.1253.2587.48
+contact48.3160.8388.47

Removing the contact preferences, the contact topology precision and recall reduced heavily. These drops demonstrate that introducing contact preferences improves the interaction fidelity.

Q1: Comparison with best-of-K sampling

We construct a best-of-KK sampling baseline with InterGen. For each prompt, we generate K=12K=12 candidates and rescale each physical metric across the candidates to [0,1][0,1], where higher score indicates better performance. We select the candidate with the highest total score. The results are as follows:

MethodFIDR-Precision (Top-1)Skate↓Float↓Interpen↓Jerk↓Force Error↓
InterGen1.5390.4747.2722.4010.69148.96577.41
Best of K (K=12)1.9750.4595.8819.799.10125.52526.35
DIGHT1.4740.4865.6718.588.42111.94427.98

DIGHT achieves better results across all metrics while requiring only one generated candidate at inference.

Q2: Discussion of Limitations.

Thank you for this suggestion. We have added a limitations section in Appendix M to discuss it.

Dear Reviewer vyEG,

We would like to follow up and ask whether our responses have addressed your concerns. We would be happy to discuss any remaining issues. If our responses have resolved your concerns, we would be grateful if you could update your assessment accordingly.

Thank you for your time!

Best regards,

Authors of Submission #5253

Reviewer_h3V5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a system to improve text-conditioned two-person humanoid motion controller. They key is the post-train the kinematic motion generator with physics feed back from tracking policy. Using the finetuned generator to create motion for tracking policy finetuning is a plus.

For kinematic motion generator post-training, it constructs prefereces fro tracking, friction, floating, contact occurrence and contact quality, and then uses a balanced diffusion-DPO objective, borrowed from BideDPO, to update the generator. Results across two generators and two datasets are promising.

Critical strengths
  1. The paper raises an important distinction that successful tracking does not necessarily responsible for interaction fidelity. It's propose solution of using the contact occurrence and quality preference as criterion, though straightforward, is well-motivated design. The quantitative ablation experiments show some level of improvement, especially for interperson penetration and force error, which is promising.
  2. It demonstrate its efficacy for post-training on two pretrained generator and benchmark.
  3. The human-reviewed contact annotation subset provides an additional evaluation of intended interpersonal contacts, which is useful.
Critical weaknesses
  1. I will view this work mainly as system-level contribution, because using DPO with preference from tracker for tuning motion generator (PhysMoDPO) is not new, and the "decoupled" DPO (BideDPO) are borrowed. The integration with tracker adaptation seems to be subtle. The contribution is in principle the contact-aware extension to two-person interactions. More explicit positioning and matched baseline details would clarify the advance.
  2. For contact-related preference annotation, the reference forces come from GT motions tracking trajectories. However, the implementation of tracking policy is unclear. Contact force from GT tracking trajectory does not necessarily reflect to the actual human interaction force. For example, a controller may reproduce a reactive motion through its own actuation even without the expected contact force from the other participant. This raises concerns about the reliability of tracker-derived forces as supervision for interaction fidelity.
  3. If I understand correctly, Eq. 5 divides five losses into general-physics and contact groups, normalizes weights within each group, and introduces a contact-group coefficient. The experiments do not establish why this hierarchy is preferable to jointly normalizing all five losses, equal weighting, or other groupings.
  4. The terminology appears stronger than the method warrants. “Co-adaptation” suggests a coupled adaptation process, whereas the method sequentially fine-tunes the generator and then the tracker, holding the other fixed at each stage. Likewise, “physics-decoupled” does not describe an explicit mechanism that disentangles physical factors or resolves conflicts between objectives; it constructs criterion-specific preferences and computes separate losses, which still update the same generator. Terms such as “sequential generator–tracker fine-tuning” and “criterion-specific preference optimization” would describe the method more precisely.
Questions for authors
  1. How does DIGHT compare with PhysMoDPO with matched setting?
  2. Please clarify the implementation and training of the tracking policy. How do you ensure that contact forces obtained by tracking ground-truth motions reliably reflect the intended interaction, rather than artifacts of the controller?
7 comments
AuthorsOpenReview ↗

Response to Reviewer h3V5

We thank the reviewer for recognizing our motivation, the effectiveness of contact-aware post-training, and the value of our contact annotations. We address the specific concerns below.

W1: Positioning and contribution.

Our central contribution is introducing inter-person contact forces for text-to-interaction generation. Making plausible interaction and contact as intended is one of the most important aspects of two-person motion. However, existing methods such as PhysMoDPO lack mechanisms to ensure the fidelity of complex interaction. Considering only whether the single motion is physically executable can lead to the failures shown in Fig. 1 of the main paper. The intended contact may fail to occur, or it may occur with excessive force destabilizing the humanoids.

Our key insight is the inter-person contact force can directly reveal whether contact occurs, which body parts make contact, and what forces arise. These signals are essential for improving interaction fidelity. Using DPO to incorporate non-differentiable feedback into training is common practice. Our main different with existing methods (e.g., PhysMoDPO) is to use inter-person contact forces to construct contact preference pairs that directly reflect how the two people interact, thereby improving interaction fidelity.

Through introducing the contact preferences, our method improves contact topology precision from 38.29% to 46.77% and recall from 52.16% to 63.94% on contact-annotated benchmark. Qualitative comparisons in Fig.1, Fig. 3, Appendix A, and the videos in supplementary material also illustrate the improvement in interaction fidelity.

W2-1: Implementation of tracking policy.

The implementation details and training rewards are presented in Appendix I. Tracking policy observes both participants through cross-attention and outputs joint targets for PD control, with interpersonal collisions enabled.

W2-2&Q2: Reliability of contact supervision

In our implementation, we disable self-collisions and retain only PhysX contact records that pair a rigid body from one humanoid with a rigid body from the other. A reaction reproduced through actuation alone does not produce an interpersonal contact signal unless the simulator reports contact between the two humanoids.

To further evaluate our force supervision can improve interaction fidelity, we remove the contact preferences and test the resulting model on our contact-annotated benchmark:

MethodTopo. Prec. ↑Topo. Recall ↑Exec. Succ. ↑
InterGen38.2952.1684.70
DIGHT w/o contact42.1253.2587.48
DIGHT46.7763.9489.23

With force supervison, DIGHT achieves the best performance on interaction-related metrics.

Furthermore, our visualization in Fig. 1 and Fig. 3 also illustrate this improvement of interaction fidelity. For example, on shoulder pat visualization (Fig. 1), DIGHT establishes the intended shoulder contact while preserving the receiving humanoid’s balance, whereas the baseline produces excessive forces that destabilize the receiver.

W3: Effectiveness of hierarchical preference grouping.

To demonstrate that our hierarchy grouping is preferable, we compare it with two baselines: a) InterGen + DPO. We normalize each of the five physics scores across the KK candidates for each prompt, then sums them to rank candidates and construct preference pairs; b) DIGHT (one group). We retain criterion-specific preference objectives but jointly normalizes all five loss weights without grouping. The results are as following:

MethodFID↓Skate↓Float↓Interpen↓Jerk↓Force Error↓
InterGen1.5397.2722.4010.69148.96577.41
InterGen + DPO1.5016.0420.969.85134.82491.40
DIGHT (one group)1.4866.2819.419.03126.56455.22
DIGHT1.4745.6718.588.42111.94427.98

DIGHT achieves the best results across all metrics

W4: Terminology.

Thank you for the suggestion. In the newly uploaded manuscript, we revised the relevant terminology throughout the paper to describe our method more precisely.

Q1: Comparison with PhysMoDPO

We adapt PhysMoDPO to two-humanoid interactions and evaluate it under matched settings:

MethodFID↓Skate↓Float↓Interpen↓Jerk↓Force Error↓
InterGen1.5397.2722.4010.69148.96577.41
PhysMoDPO1.5165.6119.6710.32128.86523.33
DIGHT1.4745.6718.588.42111.94427.98
MethodTopo. PrecTopo. RecallSucc.
InterGen38.2952.1684.70
PhysMoDPO42.9450.5087.08
DIGHT46.7763.9489.23

As shown in tables, our method substantially improves performance on interaction-related metrics.

Thanks for the response. I appreciated the experiment for comparing with PhysMoDPO. It shows that the proposed method achieves better performance compared with PhysMoDPO. The experiments on preference grouping also almost solve my concern.

As mentioned by the author that the key contribution is using contact force to detect inter-person contact, my main concern remains that training a motion tracking policy does not necessary reproduce reliable contact force. As I mentioned, since the tracking policy is trained to reach the target state at the target frame, its training objective does not enforce natural motor behaviors. A controller can drive a character to stand up or fall down even without external forces from the other character. The contact precision mainly account for reference motion precision. So when we look at how the reference motions were obtained, it uses GMR with additional interaction preserving objective to retarget the original motion. It seems that a simple heuristic-based contact detection algorithm is off the shelf, and the sim force-based contact detection accuracy is bounded by the heuristic-based contact detection algorithm used in retargeting (or bounded by the quality of mocap data). It would be important to measure the contact detection accuracy and compare kinematic heuristic algo with simulation force-based algo. The resolution of the provided text-derived LLM annotated contact labels is too low to provide informative ground truth labels.

AuthorsOpenReview ↗

Response to Reviewer h3V5

Dear Reviewer h3V5,

Thank you for your response. We address your remaining concern below.


Comparison with kinematic algorithm

We use simulated contact forces as a proxy for task-relevant physical interaction rather than to recover the exact forces in the original motion. Both resting a hand gently on a partner’s shoulder and pushing against it exert a contact force, but tracking the hand’s faster approach and continued forward motion after contact can produce a larger force during a push. A kinematic contact detector can only label both motions as contact, whereas the simulated force can help distinguish them.

To further verify the force-based contact preferences is the better design choice, we replaced the force-based contact preferences with a heuristic-based contact detection preference while keeping other preferences and DPO optimization strategy. We used a contact distance threshold of 20 cm, which yielded the highest topology F1 on our contact-annotated benchmark, and construct the kinematic contact preference using contact topology error and contact duration.

Contact PreferenceSkateFloatInterpenJerkTopo. PrecTopo. Recall
None5.9318.819.51124.5742.1253.25
Kinematic contact6.3119.5611.37135.4634.8856.75
Force contact5.6718.588.42111.9448.3160.83
Contact ThresholdTopo. PrecisionTopo. RecallTopo. F1
10cm66.5040.5750.39
20cm39.3579.2452.59
30cm30.1987.1844.85
40cm23.5693.0437.60
50cm18.3096.4230.76
60cm11.4999.1220.59
Force58.2783.5268.65

Our force-based preference achieved the best results on all reported metrics. Compared with using no contact preference, the kinematic contact preference even increased interpenetration from 9.51 to 11.37.

Additionally, on our contact-annotated benchmark, increasing the threshold from 10 to 20 cm raised topology recall from 40.49% to 79.24%, while precision fell from 66.46% to 39.35%. This shows that the kinematic contact detector is sensitive to the threshold even at the sequence-level topology granularity. Together, this threshold analysis and the ablation results suggest that simulated contact forces are a better source of contact supervision for improving interaction fidelity.


Using caption-level annotation for text-to-interaction

Finally, human motion and interaction are inherently diverse. The same prompt, such as “shake hands” or “pat someone on the shoulder,” can be realized in many valid ways that differ in when contact begins and how long it lasts. Requiring generated motions to match the frame-level contact timing of a single reference would penalize these valid variations. We therefore evaluate whether the body regions specified in the caption make contact at any point in the sequence. This choice is consistent with the clip-level text annotations used by benchmarks such as HumanML3D, KIT, InterHuman, Inter-X and SocialX.


We hope the additional experiments and clarifications have addressed the points you raised. If our responses have resolved your concerns, we would be grateful if you could update your assessment accordingly.

Thank you for your time!

Best regards,

Authors of Submission #5253

Comments on the paper

3 comments
~Reyhaneh_Hosseininejad1OpenReview ↗

Questions about support and intended falls

Dear Author(s,

I came across this paper and had a few questions after reading it. I would like to use this discussion to ask the authors for clarification.

  1. Missing objects. How do you handle recordings where someone sits on a chair or leans on an object missing from the motion data? Do you add that support in the simulation, or exclude those sequences?

  2. When falling is part of the action. If the recorded motion shows someone falling and then being helped up, we want the tracker to reproduce the fall. Given the height rule in Appendix E.2 and the fall filtering in D.2, do you make an exception when falling is intended? Can the simulation continue through the fall and recovery, and how is that sequence scored?

  3. Contacts in the recorded motion. Section 4.2 uses the simulated recording to determine which contacts should occur. Have you checked whether that simulation misses contacts shown in the original motion? I saw the text-based evaluation in Appendix G; here I mean errors in the simulated recordings used to train the generator. How do you handle these errors?

  4. How contact is modeled. Human bodies deform during contact. Could changing the body shapes used for collisions or making the contact softer change which motion gets a better force score? Have you checked how much these settings affect the result?

If the code is available, it would also help me check these details myself. Thank you for your time.

Area_Chair_KzCgOpenReview ↗

Rebuttal Discussion

Dear Reviewers,

The authors have now submitted their rebuttal and responded to the concerns raised in the reviews. We would greatly appreciate it if you could take some time to review their responses and participate in the discussion, particularly regarding whether your concerns have been adequately addressed.

Thank you very much for your time and valuable contributions!

AuthorsOpenReview ↗

Thank you for your interest in our work

Thank you for your interest in our work and for taking the time to share your questions. We’d be happy to discuss these points with you.

Q1: Missing objects. In line with prior work such as InterAgent and Human-X, our study focuses on interactions between two people. Simulating object support, such as sitting on a chair, also requires knowing the object’s shape and where it was during the recording. However, datasets that pair human–human motion with the geometry and poses of supporting objects remain scarce, so our simulations did not include support objects. Reconstructing scene geometry and tracking motions that depend on it are directions for future work.

Q2: Intended Fall. Thank you for your valuable suggestion. The height threshold may flag a tracking failure when the fall is part of the reference motion. Motions involving a fall followed by assisted recovery account for less than 0.5% of all frames in InterHuman. These rare cases do not affect our main conclusions.

Q3: Contacts in the recorded motion. Our main contribution is to use contact forces to improve generated motions. In our experiments, optimizing for contact forces made the motions more physically realistic (Tab. 5). Finding a more accurate way to identify those contacts for training would be an interesting direction to explore.

Q4: Contact modeling. For comparison fairness, we retargeted all motions to the same skeleton and used a common rigid-body collision model. The reported force scores are therefore specific to that setup. Related physics-based methods for human–human interaction, such as InterAgent and Human-X, also use rigid-body humanoids. To the best of our knowledge, comparable work has yet to model body deformation during contact. Study of different body shapes and with deformable contact models would be a useful direction for future work.

We will release the code once the paper is accepted.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.