acceptodds
Under review as a conference paper at ICLR 2027

Close encounters: decentralized multi-agent policy learning for contact-rich embodied interaction

Abstract

Contact-rich interaction is central to physical assistance and close cooperation, yet poses a challenging coordination problem: agents must make intentional contact, avoid interpenetration, and continually adapt to one another while using only partial, egocentric observations. We formulate this problem as a decentralized multi-agent control task involving two virtual human agents. Each agent has access to partial egocentric observations and fine-grained contact goals specified by body-surface marker pairs. Collaboration emerges through online multi-agent reinforcement learning (MARL) of a single policy that serves both agents across all tasks. We learn this coordination without captured two-person interaction data by introducing a shared latent action space for planning and control. Both agents operate in this space, which is learned from single-human motion data. The same representation enables us to generate synthetic interaction demonstrations through centralized trajectory optimization in the joint latent action space of the two agents, and use these demonstrations to bootstrap online RL. Across nine tasks with variation in body shape and initial pose, this combination increases success over online RL from 36% to 83% on two-contact tasks and from 23% to 75% on four-contact tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Human-human interactions

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_x93bOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates close-contact two-person interaction (touches, multi-contact hugs) as cooperative decentralized control. Two SMPL virtual humans act from partial egocentric marker observations (a 90 degree head-mounted frustum with normal-facing visibility) and are conditioned on vertex-level contact goals. A single shared actor outputs actions in the latent space of a causal 1D-conv motion VAE trained on single-person HumanML3D data. The same latent space is used to generate synthetic two-person demonstrations through centralized BFGS trajectory optimization of contact/penetration costs, which bootstrap the actor and critic before PPO fine-tuning. Across nine kinematic tasks with randomized body shape and initial pose, BC+PPO raises success over PPO from scratch from 36% to 83% on 2-contact and from 23% to 75% on 4-contact tasks, with ablations on parameter sharing, visibility, initialization factors and success tolerances.

Critical strengths
  • Using a single-human motion prior to regularize the policy's action space and to synthesize two-agent demonstrations by differentiable trajectory optimization is an elegant way to sidestep the scarcity of mesh-level close-interaction datasets. The resulting large gains on multi-contact tasks (23% to 75%) are the paper's main evidence that this pipeline works.
  • Vertex-pair contact goals, signed-distance penetration penalties and egocentric visibility make a more demanding and more realistic coordination benchmark than skeleton-level HHI generation. The visibility and initialization ablations give useful insight into what makes multi-contact tasks hard.
Critical weaknesses
  • The bodies are kinematic SMPL meshes with no forces, balance or dynamics, so "contact-rich" here means geometric proximity without interpenetration. Both agents are the same jointly trained policy, so the agent is never tested against a partner it was not co-trained with: a passive or scripted partner, a recorded human motion, or a differently trained policy. Adapting to a partner whose actions you don't control is the core of the assistive-care motivation, and self-play conventions may not transfer.
  • The comparison is limited. No adapted interaction-synthesis baselines (InterControl, Ready-to-React). Results come from 500 evaluation episodes, from a single training seed, and planner demos require manual rejection of failures, so the reliability of the reported gaps is unclear.
Questions for authors
  1. How does the policy perform when paired with a partner it was not trained with: a frozen or passive partner, a scripted motion, a recorded HumanML3D/Inter-X motion, or a separately trained policy?
  2. How many training seeds were run, and what is the variance in success for PPO vs. BC+PPO on 2- and 4-contact tasks?
  3. Why does the independent (non-shared) policy collapse on 4-contact up (0.10) while matching elsewhere? Is this an optimization instability that also affects the unified policy across seeds?
6 comments
AuthorsOpenReview ↗

Response to reviewer x93b (1/2)

Dear Reviewer x93b, Thank you for your careful review, and for recognizing the role of the single-human motion prior in demonstration synthesis, the gains on multi-contact tasks, and the design of the benchmark and ablations.

Statistical reporting (W2/Q2)

As also noted by Reviewers 9SXj and p3g7, we have updated all quantitative results in the paper (shown in blue in Tables 1, 2, 3, 4 and Figure 3) to report the mean and standard deviation over 3 training seeds, each evaluated on the same 500 episodes. The main BC+PPO advantage remains. The ordering in Table 4 changes, in the expected direction: orientation-only randomization now gives higher four-contact success than full randomization (0.74 vs. 0.69). A single 100-epoch run takes about 35 hours on one A100 80GB with 32 workers, and the same compute also had to cover the additional experiments requested in the reviews, so running more than three seeds has not been possible yet.

Kinematic-only setting (W1)

We agree, and state it in the introduction (lines 72–80) and in Section 5 (lines 467–468): bodies are kinematic, with no contact forces or balance. Physical assistance is our motivation, not a claim we evaluate. Geometric coordination, i.e. reaching the requested contacts without interpenetration and from partial observations, is a prerequisite for physical interaction, and it is far from trivial: PPO from scratch reaches only 20% on four-contact tasks, against 69% for our method. Physics-based work such as AssistMimic tracks reference demonstrations; our policies reach specified contact goals with no reference motion to track, so the two are complementary.

Planner demo rejection (W2)

As Reviewer p3g7 also asked, we clarify the rejection step. Rejection is an automatic success check, not hand curation: we execute each optimized trajectory and keep it only if it meets the evaluation success criterion, until we have 500 per task. The rejection rate averages about 33%, from 2% (H2H) to 69% (2-contact Low). “Manual rejection” in Section 5 was a poor choice of words, now corrected. Many rejected trajectories are near-successes that methods learning from imperfect demonstrations could use.

Results clarification (Q3)

We thank the reviewer for raising this question and giving us the opportunity to clarify these results.

The low success of the independent policies on 4-contact Low and Cross comes from under-training within our fixed budget of 100 epochs. Each agent has its own network, so each network only receives its own agent’s transitions, which is about half the samples a shared network gets from the same environment steps. As a result, learning a specific task can start late on the harder tasks, and depending on the seed, the model finds a solution more or less quickly.

This is illustrated by two training seeds: training success first exceeds 5% at epochs 95 and 38 for Low, and at epochs 97 and 59 for Cross, compared with epochs 30 and 27 for Up. Thus, the two seeds show substantial variation in convergence speed. The seed reported in the submission only began solving Low and Cross during its final five epochs, explaining the low reported success rates of 0.11 and 0.10. We therefore interpret these results as reflecting the fixed training budget and slower convergence of the independent policies on the harder 4-contact tasks, rather than an inability of the policies to learn these tasks.

Partner adaptation (W1/Q1)

We thank Reviewers x93b and 9SXj for this question. We added a cross-play evaluation (Appendix C.3, Table 9, in blue). We train policies with different seeds and, at test time, control the two agents with policies from two different seeds, so each agent faces a partner never seen during training. We compare with self-play, where a single policy controls both agents (our main method).

Success rate1-contact2-contacts4-contacts
Self-play0.940.820.69
Cross-play0.840.430.33

A drop is expected: independently trained policies settle on different but equally valid conventions, and a multi-contact task succeeds only when the two are compatible. Cross-play therefore retains 84% on one-contact tasks and falls to 43% and 33% on two- and four-contact tasks. This is consistent with the known zero-shot coordination gap of self-play (Hu et al., ICML 2020). Training against a population of diverse partners improves coordination with unseen partners in other cooperative domains (Strouse et al., NeurIPS 2021; Lupu et al., ICML 2021) and is our next step. Passive, scripted or recorded partners remain untested.

AuthorsOpenReview ↗

Response to reviewer x93b (2/2)

Missing comparisons (W2)

A comparison on exactly the same task is hard for all three methods, because these methods do not share our interface (goals as pairs of body-surface points, random initial states, per-agent partial observations):

  • Text-to-interaction methods (e.g., InterGen): text alone cannot specify which body-surface points are required to make contact, applying this method to our setting would fail entirely.

  • Ready-to-React: no goal input, and its released model covers boxing only. Training it on our tasks would require paired interaction motions; we have not evaluated such an adaptation.

  • InterControl: the closest, but its constraints are between skeleton joints, not surface points, and it generates both motions jointly and offline from the full state.

We still ran InterControl on our nine tasks, using its released model without retraining. Each target pair becomes a distance constraint between the nearest joints, clips start from our initial states, and the generated meshes are scored with our success criterion (500 clips per task).

InterControl is zero-shot, is constrained on joints while success is measured on the surface, and uses a mean body shape (ours: random shapes). Conversely, it is centralized with full state access, an advantage over our decentralized policy.

1-contact2-contact4-contact
Ours (BC+PPO): success0.940.820.69
InterControl: success0.040.000.00
Clips reaching all contacts (within 10 cm)0.400.010.00
Clips within 2.6 cm penetration0.140.070.10

Success requires both conditions at once: all contacts reached and penetration within 2.6 cm. InterControl fails on both. On 1-contact tasks, the hand reaches its target in 40% of the clips, but mostly by entering the partner's body: only 14% of the clips stay within the penetration limit. On hugs, it almost never reaches all targets (1% of the clips with 2 contacts, none with 4), and penetration remains as frequent. We stress that this comparison is not fair to InterControl: it is evaluated zero-shot on a task it was not designed or trained for.

Reviewer_x93bOpenReview ↗

Thank you for your response

I have few things that I want your clarification:

  1. InterControl is conditioned on text, which text prompt was used?
  2. When you say InterControl generates both motions jointly and offline from the full state, the proposed method also uses one shared network for both agents thus I think each agent knows its partners policy. So I think the proposed method is not very different from InterControl other than using skeleton vs surface in essence.
Reviewer_p3g7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks at how two virtual humans can make specific body contacts, like touching a head or back or hugging, without passing through each other. It treats this as decentralized multi-agent control in a kinematic setting, where each agent sees only its partner’s visible markers and the target contacts. Both agents share one policy that acts in the latent space of a causal motion VAE trained on single-person motion. The same latent space is used to jointly optimize both agents’ trajectories into synthetic demonstrations, which pretrain the policy before PPO fine-tuning. Across nine tasks, this raises success over PPO from scratch from 36% to 83% on 2-contact tasks and from 23% to 75% on 4-contact tasks. Ablations cover parameter sharing, visibility, initial randomization, and contact tolerances.

Critical strengths
  1. A clean idea. One motion latent space serves both offline joint planning and online decentralized control, so demonstrations are generated directly in the policy’s action space, with no captured two-person data needed.

  2. A large, consistent gain. BC+PPO clearly beats both BC and PPO, especially on 4-contact tasks (75% vs 23%), and the ablations (shared vs. separate policies, visibility, initialization, tolerance sensitivity) are informative.

Critical weaknesses
  1. Kinematic-only setting. There are no physics, contact forces, or balance. “Penetration-free contact” is a geometric check, so the physical-assistance motivation (carer–patient) isn’t really tested.

  2. Only nine fixed tasks, with no generalization test. Training and testing use the same nine contact specifications, differing only in initial randomization.

  3. Didn’t provide video demos.

  4. Weak partial observability. Visibility uses only frustum and normal culling, with no self-occlusion and perfect marker positions without noise. Full vs. partial visibility differs by only 2–5 points (Tab. 3), so the “egocentric partial observation” setting is quite easy.

  5. Motion quality isn’t measured. Only success and the training return are reported. There’s no measure of motion naturalness (FID against real motion, jitter, foot skating) and no user study, and the “planner exploits the cost” claim rests only on returns.

  6. Inconsistencies and missing variance.

  7. Missing related work. Physics-based multi-character interaction (Zhang et al., “Simulation and Retargeting of Complex Multi-Character Interactions,” SIGGRAPH 2023) and latent action spaces for humanoid control (ASE, SIGGRAPH 2022; PULSE, ICLR 2024) are not discussed.

Questions for authors
  1. How natural is the motion: FID against real motion, jitter, foot skating, or a user study?

  2. Can you report results over multiple training seeds, and explain the Tab. 4 4-contact result and the gap between Fig. 3 and Tab. 1?

  3. How many planner trajectories were rejected, and how? How sensitive is BC+PPO to the number and quality of demonstrations?

  4. Could you provide some video demos?

4 comments
AuthorsOpenReview ↗

Response to reviewer p3g7 (1/3)

Dear Reviewer p3g7, Thank you for your review and for highlighting the clean idea of a shared latent space for planning and control, the large gains of BC+PPO, and the informativeness of our ablations.

Statistical reporting (W6/Q2)

As also noted by Reviewers 9SXj and x93b, we have updated all quantitative results in the paper (shown in blue in Tables 1, 2, 3, 4 and Figure 3) to report the mean and standard deviation over 3 training seeds, each evaluated on the same 500 episodes. The main BC+PPO advantage remains. The ordering in Table 4 changes, in the expected direction: orientation-only randomization now gives higher four-contact success than full randomization (0.74 vs. 0.69). A single 100-epoch run takes about 35 hours on one A100 80GB with 32 workers, and the same compute also had to cover the additional experiments requested in the reviews, so running more than three seeds has not been possible yet.

Example videos (W3/Q4)

Videos were part of the original submission. The anonymous video gallery shows success and failure cases for every task and is referenced three times in the manuscript: a footnote on page 2 (line 94), the qualitative results on page 8 (lines 401–402, with the URL), and Appendix C.3 (line 909). We believe this addresses W3 and Q4, and we would be grateful if the reviewer could consult the gallery, in particular for motion quality (W5).

Kinematic-only setting (W1)

We agree, and state it in the introduction (lines 72–80) and in Section 5 (lines 467–468): bodies are kinematic, with no contact forces or balance. Physical assistance is our motivation, not a claim we evaluate. Geometric coordination, i.e. reaching the requested contacts without interpenetration and from partial observations, is a prerequisite for physical interaction, and it is far from trivial: PPO from scratch reaches only 20% on four-contact tasks, against 69% for our method. Physics-based work such as AssistMimic tracks reference demonstrations; our policies reach specified contact goals with no reference motion to track, so the two are complementary.

Missing related work (W7)

We thank the reviewer. The revised manuscript (changes in blue) now discusses these works: Zhang et al. (2023) and ASE (Peng et al., 2022) in Sec. 2 (lines 143-146), and ASE and PULSE (Luo et al., 2024) in Sec. 3.2 (lines 244–250). In short:

  • Zhang et al. (2023) train physically simulated characters to reproduce a given reference interaction and retarget it to other body shapes. They need a two-person reference motion for every interaction; we use none, and specify only which surface points should touch.
  • ASE and PULSE are related, but address a different problem: a single physically simulated humanoid, with no second agent and no interaction. They learn a latent action space from single-person motion to make downstream RL tractable, the role our latent space plays for the policy. Their rollouts, however, pass through a physics simulator that is not differentiable. Ours is kinematic and differentiable from latent actions to posed meshes and contact and penetration costs, which is what lets us reuse the latent space for centralized trajectory optimization and generate two-person demonstrations.

Results inconsistencies (W6/Q2)

We thank the reviewer for raising these points and clarify them below.

  • Consistency between Table 1 and Figure 3. The two results are consistent but use different settings. Table 1 reports evaluation success over 500 episodes per task and Figure 3 reports training success. Averaged over the 9 tasks, Table 1 gives 0.44 (PPO) and 0.80 (BC+PPO), which matches the final values in Figure 3: ~0.43 and ~0.80.
  • Consistency of the 4-contact results in Table 4. In the original version, randomizing only the orientation gave lower 4-contact success than full random initialization. This came from using a single seed. Over 3 seeds, the results follow the expected trend: 0.74 for orientation-only against 0.69 for full randomization.
AuthorsOpenReview ↗

Response to reviewer p3g7 (2/3)

Planner demo rejection (Q3)

Rejection is an automatic success check, not hand curation: we execute each optimized trajectory and keep it only if it meets the evaluation success criterion, until we have 500 per task. The rejection rate averages about 33%, from 2% (H2H) to 69% (2-contact Low). “Manual rejection” in Section 5 was a poor choice of words, now corrected. Many rejected trajectories are near-successes that methods learning from imperfect demonstrations could use.

Influence of the number of demos (Q3)

We thank the reviewer for this question. We added an ablation on the number of demonstrations used for BC pretraining. We report the success rate over all tasks just after BC (before any PPO update) and at the end of training.

Pretraining demosSuccess after BC (%)Final success (%)
None (PPO only)0.043.6
1002.680.2
5001.977.1
9004.080.5
27004.978.0
4500 (default)4.780.3

The results show that pretraining is important, but the number of demonstrations is not. Without pretraining, PPO reaches only 43.6%. With pretraining, all runs reach 77–81%, even with only 100 demonstrations (about 11 per task). BC alone is weak (2–5% success), so it does not solve the task. It only gives PPO a good starting point, which is enough for RL to learn.

Motion quality (W5/Q1)

We thank the reviewer for raising this concern and allowing us to clarify this point.

Our policy does not predict body poses directly. It acts in the latent space of a VAE trained on HumanML3D, and the VAE decodes each action into a full-body motion. So the quality of our motions depends on the quality of the VAE. In the revision, we added an evaluation of the VAE (Section C.3, Table 7). We use standard metrics on the HumanML3D test set and compare with the VQ-VAE of T2M-GPT. The results show that our VAE decodes natural, high-quality motions, so our policy has a good action space.

We also directly evaluate the motions produced by our policy (Section C.3, Table 8) using two standard metrics: jitter (mean joint jerk) and foot skating rate. We compare against real motion data (HumanML3D, InterHuman, and Inter-X) and OmniControl. The results are consistent with real motion: foot skating is only 1.9%, lower than both the real datasets (4.4–12.0%) and OmniControl (5.5–6.1%), while our jitter (63.9 m/s³) lies within the range of real data. These results confirm that our policy produces natural motions without excessive jitter or foot sliding, which can also be directly verified in our video gallery.

Unseen tasks (W2)

We evaluate the trained policy zero-shot on 33 held-out contact specifications built from the same body markers as the nine training tasks. None of these 33 specifications is used as a training goal. The table groups all 33 specifications into four categories (7 + 13 + 7 + 6); the nine trained tasks are a separate reference row. This tests transfer to new combinations and pairings within the existing marker vocabulary.

Unseen specification, relative to trainingExampleSetsSuccess
Trained tasks (reference)—90.80 ± 0.03
Single pair, trained only inside a larger setone arm of a trained hug70.23 ± 0.04
Half of a trained 4-contact taskthe waist contacts of a trained full hug130.17 ± 0.01
Pairs from two different trained tasks, combinedleft arm of one trained hug + right arm of another70.14 ± 0.03
Trained markers, in a pairing never trainedleft palm→head (only right palm→head was trained)60.00 ± 0.00
(mean ± std over the 3 seeds)

We observe that:

  1. Partial transfer within trained pairings. When the request is a subset or recombination of trained contact sets (the three middle categories), success is 23%, 17% and 14%: well below the 80% on trained tasks, but not zero.
  2. No extrapolation to new pairings. When the request needs a marker pairing never seen in training (last category), success is zero. With only nine training specifications this is expected: the policy learns the pairings it was trained on rather than a general contact skill, and training on a broader, sampled set of specifications is the natural way to close this gap.
AuthorsOpenReview ↗

Response to reviewer p3g7 (3/3)

Weak partial observability (W4)

We agree that our visibility model omits self-occlusion and sensing noise; Section 5 notes that we rely on structured marker observations without image-based perception.

Two clarifications:

  1. The setting is egocentric, and hard even with full visibility. Each agent sees its partner only through a 90° head-mounted camera, in its own frame, and only markers facing the camera; the partner's far side is never observed. With all markers visible, success is still only 97%, 87% and 74% on one-, two- and four-contact tasks, so the main difficulty is decentralized coordination. Partial visibility costs a further 3–5 points: the policy copes with hidden markers. By contrast, InterControl and Ready-to-React assume the fully observed state of both characters, and Reviewer x93b notes that our vertex-pair goals, penetration penalties and egocentric visibility make a more demanding benchmark than skeleton-level interaction generation.

  2. Toward realistic sensing. Our focus is the control and coordination problem, not perception, so we use structured marker observations. A natural extension is to estimate marker positions from the egocentric image with an off-the-shelf depth or human mesh recovery model, which would introduce realistic noise. Our actor already takes a variable subset of visible markers, so its architecture would not change. However, rendering and running such a model at every step of online RL would add substantial computational overhead, and we leave it to future work.

Dear Reviewer p3g7,

Thank you again for your time and your helpful feedback on our paper.

Since the discussion period ends today, we wanted to check if our response and the revised paper answer your concerns. If something is still not clear, please tell us and we will be happy to explain more.

If your concerns are solved, we would be grateful if you could consider updating your score.

Reviewer_9SXjOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies decentralized coordination between two kinematic human avatars performing contact-rich interactions. A causal motion VAE provides a shared action space for centralized trajectory optimization and decentralized policy execution. Synthetic demonstrations initialize a shared actor and privileged critic, followed by online PPO. Across nine tasks, this improves success over PPO from 36% to 83% on two-contact tasks and from 23% to 75% on four-contact tasks.

Critical strengths
  1. The approach coherently connects centralized planning and decentralized control through a shared differentiable motion representation.
  2. It avoids reliance on captured two-person interaction data, addressing a practical data limitation.
  3. Improvements over PPO are large and consistent across the nine tasks. The learned policy also exceeds the planner’s average success.
  4. A single policy supports both agents and all tasks. Ablations examine parameter sharing, visibility, and initialization variation.
Critical weaknesses
  1. Potential errors in the success criterion. Appendix Eq. 9 checks the maximum signed distance, whereas avoiding excessive penetration appears to require the minimum. Its aggregate contact-distance norm also differs from the main text’s requirement that every pair satisfy its threshold. These may be notation errors, but require verification.
  2. Permissive contact tolerances. Success allows 10 cm target separation and 2.6 cm penetration. At 1 mm penetration tolerance, success falls to 31% and 14% for two- and four-contact tasks. The results support approximate geometric coordination more strongly than precise, penetration-free contact.
  3. Limited evidence of partner adaptation. Evaluation uses co-adapted agents sharing one policy. It does not test unfamiliar partners, scripted behavior, or mid-episode perturbations. Timestep conditioning also leaves open how much success depends on synchronized behavior.
  4. Incomplete attribution of gains. BC+PPO establishes the value of demonstration-based initialization, but the experiments do not isolate the contribution of the shared latent representation, critic pretraining, or penetration continuation strategy. MSE behavior cloning is also a limited imitation baseline.
  5. Limited generalization and statistical reporting. Evaluation covers the nine training tasks under bounded initialization variations. Independently trained seeds, uncertainty, unseen contact combinations, and total computational cost are insufficiently characterized.
Questions for authors
  1. Is Eq. 9 implemented using the minimum signed distance? Does contact success require every target pair to be within 10 cm, or an aggregate norm below 10 cm?
  2. Are penetration violations tracked across the entire episode? Does evaluation use sparse markers, decimated vertices, or the full-resolution mesh?
  3. What are success rates under jointly tighter contact and penetration thresholds, validated on the full mesh?
3 comments
AuthorsOpenReview ↗

Response to reviewer 9SXj (1/2)

Dear Reviewer 9SXj, Thank you for your review. The strengths you list (a shared latent space linking centralized planning to decentralized control, no need for captured two-person interaction data, and large, consistent gains over PPO) are the central contributions of the paper.

Statistical reporting and computational cost (W5)

As also noted by Reviewers p3g7 and x93b, we have updated all quantitative results in the paper (shown in blue in Tables 1, 2, 3, 4 and Figure 3) to report the mean and standard deviation over 3 training seeds, each evaluated on the same 500 episodes. The main BC+PPO advantage remains. The ordering in Table 4 changes, in the expected direction: orientation-only randomization now gives higher four-contact success than full randomization (0.74 vs. 0.69). A single 100-epoch run takes about 35 hours on one A100 80GB with 32 workers, and the same compute also had to cover the additional experiments requested in the reviews, so running more than three seeds has not been possible yet. At test time our policy completes an episode in 0.8 s, while the trajectory-optimization planner baseline takes about 53 s, roughly 66x slower, because it solves a new optimization for every episode (we will add it to the final version).

Correction of Eq. 9 (W1/Q1)

We thank the reviewer for spotting these inconsistencies. Both are notation errors in Appendix Eq. 9. The implementation follows the definition given in the main text, and every reported result was obtained with it. The corrected equation reads (changes in blue in the revised appendix):

rgoal,t=I[max⁡1≤n≤N∥mn,t1−mn,t2∥2<τ    ∧    min⁡t′≤t min⁡1≤k≤VPk,t′>δ]r_{\text{goal},t}=\mathbb{I}\Big[\max_{1\le n\le N}\big\|m^1_{n,t}-m^2_{n,t}\big\|_2<\tau \;\;\land\;\; \min_{t'\le t}\,\min_{1\le k\le V}\mathbf{P}_{k,t'}>\delta\Big]

Clarifications on success criterion and evaluation (W2/Q1/Q2)

We thank the reviewer for requesting clarification on our success criterion and evaluation protocol.

  • Contact. Each task specifies target pairs among the 69 sparse body-surface markers that also form the policy observations (lines 192–194, 204–205). Success requires every pair to be within τ=10\tau = 10 cm at the same frame (lines 313–314): a conjunction of per-pair conditions, not an aggregate norm. Markers are vertices of the full-resolution SMPL mesh.
  • Penetration. We compute the signed distance from each vertex of one agent to the surface of the other agent, in both directions. This distance must stay above δ=−2.6\delta = -2.6 cm (negative is inside) at every frame of the episode. If it goes below this value even once, the episode is a failure. To reduce the computational cost, we compute it on a decimated SMPL mesh (431 vertices, 858 faces; Appendix B.2, lines 799–803). So, to Q2: violations are tracked over the entire episode; contact uses sparse markers, penetration uses decimated vertices, and a full-mesh re-evaluation follows.

Why 2.6 cm (W2). We chose the 2.6 cm margin to allow shallow overlap as a proxy for tissue deformation, while aiming to prevent shortcuts through the partner's wrist, which on a real body would mean breaking a bone. Under tighter margins our advantage persists: at 1 mm, BC+PPO retains 31% and 14% success on two- and four-contact tasks, against 6% and 2% for PPO (Fig. 5).

Tighter evaluation (Q3)

Full mesh. For efficiency, penetration is computed on a decimated SMPL mesh (431 vertices). We re-evaluated the same BC+PPO policies with the success test on the full 6,890-vertex mesh, with unchanged thresholds (τ=10\tau = 10 cm, ∣δ∣=2.6|\delta| = 2.6 cm), 3 seeds and 500 episodes per task.

TasksDecimated meshFull mesh
1 contact94.5 ± 0.585.7 ± 0.9
2 contacts81.9 ± 2.061.4 ± 2.7
4 contacts68.7 ± 5.845.0 ± 5.1
All tasks80.3 ± 2.861.4 ± 2.6

The decimated surface smooths thin parts (fingers, elbows, shoulders) and under-estimates penetration depth by about 1–1.5 cm.

Joint thresholds. We added Fig. 8 in Section C.3, it varies τ\tau and ∣δ∣|\delta| jointly on the same rollouts (with decimated mesh). Success degrades smoothly and reaches 100% at loose thresholds. Contact distance is the more binding of the two: at τ=10\tau = 10 cm, tightening ∣δ∣|\delta| from 2.6 to 1 cm gives 77/50/29% (1/2/4 contacts), while tightening τ\tau alone to 7.5 cm gives 74/47/21%. This matches training, which rewards and terminates as soon as all pairs are within 10 cm.

The full mesh SDF computation and thresholds are training hyperparameters, so a policy could be trained directly with stricter ones, e.g. a smaller τ\tau and a margin measured on the full mesh.

AuthorsOpenReview ↗

Response to reviewer 9SXj (2/2)

Partner adaptation (W3)

We thank Reviewers 9SXj and x93b for this question. We added a cross-play evaluation (Appendix C.3, Table 9, in blue). We train policies with different seeds and, at test time, control the two agents with policies from two different seeds, so each agent faces a partner never seen during training. We compare with self-play, where a single policy controls both agents (our main method).

Success rate1-contact2-contacts4-contacts
Self-play0.940.820.69
Cross-play0.840.430.33

A drop is expected: independently trained policies settle on different but equally valid conventions, and a multi-contact task succeeds only when the two are compatible. Cross-play therefore retains 84% on one-contact tasks and falls to 43% and 33% on two- and four-contact tasks. This is consistent with the known zero-shot coordination gap of self-play (Hu et al., ICML 2020). Training against a population of diverse partners improves coordination with unseen partners in other cooperative domains (Strouse et al., NeurIPS 2021; Lupu et al., ICML 2021) and is our next step. Passive, scripted or recorded partners remain untested.

Unseen tasks (W5)

We evaluate the trained policy zero-shot on 33 held-out contact specifications built from the same body markers as the nine training tasks. None of these 33 specifications is used as a training goal. The table groups all 33 specifications into four categories (7 + 13 + 7 + 6); the nine trained tasks are a separate reference row. This tests transfer to new combinations and pairings within the existing marker vocabulary.

Unseen specification, relative to trainingExampleSetsSuccess
Trained tasks (reference)—90.80 ± 0.03
Single pair, trained only inside a larger setone arm of a trained hug70.23 ± 0.04
Half of a trained 4-contact taskthe waist contacts of a trained full hug130.17 ± 0.01
Pairs from two different trained tasks, combinedleft arm of one trained hug + right arm of another70.14 ± 0.03
Trained markers, in a pairing never trainedleft palm→head (only right palm→head was trained)60.00 ± 0.00
(mean ± std over the 3 seeds)

We observe that:

  1. Partial transfer within trained pairings. When the request is a subset or recombination of trained contact sets (the three middle categories), success is 23%, 17% and 14%: well below the 80% on trained tasks, but not zero.
  2. No extrapolation to new pairings. When the request needs a marker pairing never seen in training (last category), success is zero. With only nine training specifications this is expected: the policy learns the pairings it was trained on rather than a general contact skill, and training on a broader, sampled set of specifications is the natural way to close this gap.

Attribution of gains (W4)

We thank the reviewer for this comment and clarify the role of each component below.

  • Shared latent representation. The latent action space is a prerequisite for learning rather than an optional component. The two natural alternatives are not viable in our setting:

    • Raw SMPL parameter space: acting directly in raw SMPL parameter space is too unconstrained, as the policy must then discover plausible human motion from scratch, and exploration does not produce valid motion.
    • Other VAE: for instance, the EgoGen (Li et al., CVPR 2024) motion VAE has limited motion diversity and only supports walking, so it cannot express the close-contact interactions we target.
  • Critic pretraining. In our experiments, critic pretraining mainly accelerates early training and has little effect on final performance. We therefore regard it as an implementation detail for efficiency rather than a source of the reported gains.

  • Penetration continuation. This strategy is used only for the planner; it is not used to train the BC+PPO policies. It can therefore affect BC+PPO only indirectly, through the demonstrations it helps the planner find.

  • MSE behavior cloning. A simple objective, deliberately. BC alone performs poorly in closed loop because errors compound, yet it gives PPO an initialization good enough to lift four-contact success from 20% to 69%. We therefore do not see better BC as the way forward: demonstrations only need to be good enough to start online learning, which supplies the gain.

Dear Reviewer 9SXj,

Thank you again for your time and your helpful feedback on our paper.

Since the discussion period ends today, we wanted to check if our response and the revised paper answer your concerns. If something is still not clear, please tell us and we will be happy to explain more.

If your concerns are solved, we would be grateful if you could consider updating your score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal summary

Dear Area Chair and Reviewers,

We thank all reviewers for their feedback. We ran every experiment that was requested and that fit in our compute budget. All changes are in blue in the revised paper.

New experiments

  • Seeds. All main results (Tables 1–4, Figure 3) now report mean ± std over 3 training seeds. The BC+PPO advantage remains: 0.80 vs. 0.44 over all tasks.
  • Stricter evaluation. We re-evaluated on the full 6,890-vertex mesh, and we added a joint sweep of the contact and penetration thresholds (Fig. 8).
  • Unseen partners. We added a cross-play evaluation, where the two agents are controlled by policies from different seeds (Table 9).
  • Unseen tasks. We tested the policy zero-shot on 33 held-out contact specifications.
  • Number of demonstrations. We added an ablation from 100 to 4,500 demonstrations. Pretraining matters, but 100 demonstrations are already enough.
  • Motion quality. We evaluated the VAE (Table 7) and the motions of our policy, with jitter and foot skating (Table 8).
  • Baseline. We ran InterControl on our nine tasks.

We also clarify several other specific points in the reply to each related reviewer.

We hope these additions answer the reviewers’ concerns, and we are happy to discuss further.

Area_Chair_z5CXOpenReview ↗

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!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.