Clipping Is a Timing Problem: Contact Between Decentralized Avatars Under Latency

When two people's 3D bodies are generated by two separate models, as in networked VR, telepresence, or two robots that each plan their own motion, each model sees its partner's motion only after a network delay. Close contact then breaks down. A handshake ends with fingers sunk into a palm, or with two hands hovering a centimeter apart. This note argues that much of this failure is a timing problem rather than a geometry problem, and that a classical result from distributed computing says exactly how large it must be.

That result says two agents can only coordinate an action to within the uncertainty of their message timing. Translated into geometry, a timing uncertainty of $\epsilon$ seconds between partners, multiplied by the speed at which surfaces approach, gives a floor on how much they must either overlap ("clip") or miss ("float") at the planned moment of contact. No collision loss, however well tuned, can beat that floor.

We propose to prove this floor, measure it on human-human interaction benchmarks, and build a decentralized generator that respects it. The generator has three parts. Partners agree on discrete contact events, which can be shared reliably at coarse time resolution. Each agent slows its final approach so that speed times timing uncertainty stays inside a soft-tissue tolerance. When an agreement is not confirmed, the agent defaults to a gap rather than an overlap.

Core Claim

The epistemic-logic literature shows that perfect common knowledge, meaning everyone knows, knows that everyone knows, and so on, is equivalent to perfectly simultaneous action. It is therefore impossible under any delivery jitter (Joseph Halpern et al.). Common knowledge revisited adds the positive half. A weaker notion, ε-common knowledge (everyone learns the fact within an $\epsilon$-second window), is necessary and sufficient for coordinating actions to within $\epsilon$. When delays have known bounds (Basil Tampakas et al.), $\epsilon$ is set by the jitter of the channel.

Our claim is a kinematic consequence of this. Let $v$ be the speed, along the contact normal, at which a partner's surface closes on the agent. That partner's commitment to touch at a planned instant $t^*$ is only known to within $\epsilon$. Let $p$ be the penetration depth and $g$ the floating gap measured when the agent believes contact occurs. Then in the worst case over admissible timings,

$$\max(p, g) \;\ge\; \tfrac{1}{2}\, v\, \epsilon ,$$

because the partner's surface position is uncertain over an interval of length $v\epsilon$, and the best a plan can do is aim at the interval's centre.

Take the roughly 2 mm soft-tissue tolerance $d$ that ContactOpt found necessary for stable hand-object contact. The bound then gives a design rule: the final approach must satisfy $v \le 2d/\epsilon$. At 50 ms jitter, that caps the closing speed at 8 cm/s in the last moments before contact.

The novelty is that the "contact-penetration duality" described in Layout-your-3D and ContactOpt stops being a tuning dilemma in the decentralized setting. It becomes a quantity with a provable lower bound and a prescribed remedy. Every existing interaction generator sidesteps the issue by sampling both people jointly, which implicitly assumes zero-latency common knowledge. That includes InterGen, InterMask and DHVAE.

How the two directions fit together

The two directions play different roles. Common-knowledge theory supplies the mechanism and the bound. Mesh interpenetration is the application regime and the measuring instrument.

Several epistemic results constrain the design.

  • Common Knowledge, Regained recovers common knowledge using local timestamps, but only under guaranteed delivery. With any packet loss, the Email Game impossibility returns. Contact is exactly when line-of-sight occlusion and packet drops are most likely, so we cannot rely on agreement eventually arriving. This is why we require that unconfirmed contact defaults to a gap: floating is safe, clipping is not.
  • The granularity argument in Common knowledge revisited also matters. Common knowledge holds if one models time coarsely. This motivates sharing contact events (which body parts touch, on which beat) at coarse resolution, while the fine-grained approach motion is handled locally.
  • MACKRL supplies a method precedent: condition joint decisions only on what a subgroup commonly knows. It works on StarCraft unit coordinates with a perfect step clock, and it has neither continuous surfaces nor timing jitter.
  • CAPI coordinates through a public-belief coordinator, but only on card and bidding games. We do not attempt to scale its lookahead to continuous motion. We borrow only its idea that decisions should be keyed to public information.

On the geometry side, our local contact terms reuse existing differentiable machinery. Grasping Field uses dual signed distance functions (a signed distance function gives, for any point in space, its distance to a surface, negative inside). ContactOpt contributes the soft tolerance. DHVAE's evaluation of penetration volume and contact ratio is our main scoreboard. InterMask's authors concede that their skeletal outputs clip when converted to body meshes. DHVAE reduces penetration volume on InterHuman to 0.390, against InterMask's 0.873. Both numbers are obtained with joint, latency-free sampling, and that is the gap we study.

Research Program

Theory. We formalize the setting as two agents in a bounded-delay interpreted system. Each agent outputs a continuous surface trajectory, and contact intents travel over a channel with jitter bounded by $\epsilon$ and loss rate $q$. We then prove three things:

  • the worst-case floor above;
  • a matching upper bound for the speed-governed protocol;
  • a loss-aware statement: with loss probability $q$ and $k$ acknowledgment rounds, the probability of entering an unconfirmed contact decays geometrically in $k$, and our default-to-gap rule converts every residual failure into floating rather than penetration.

This last result is the honest version of the Halpern-Moses message-round lower bound in a physical setting.

System. Each person is generated by its own copy of a reaction model, InterMask-style token unmasking with one partner fixed. Each copy is conditioned on the partner's motion stream as received, meaning delayed and lossy. The two copies share a narrow event channel carrying contact intents: a body-part pair and a target beat, plus acknowledgments. At sampling time, each agent applies three terms during the low-noise steps, following InterGen's damping schedule:

  • a local penetration term over SMPL meshes (SMPL is the standard parametric body mesh), computed from dual signed distance functions with tolerance $d$;
  • an attraction term toward the agreed contact;
  • a speed governor that limits normal closing speed to $2d/\hat\epsilon$, where $\hat\epsilon$ is the agent's running estimate of channel jitter.

The models are trained with randomized latency and loss so that they learn to read stale partner streams.

Measurement. We first measure the natural pre-contact closing speeds in InterHuman and Inter-X. This tells us how much the governor would distort typical human motion at realistic latencies. If real handshakes and partner dances already decelerate below $2d/\epsilon$ at 30–60 ms, the governor is nearly free. If they do not, the realism cost becomes a reported result in its own right.

Baselines

Naive reference. Two independent reaction generators each consume the delayed partner stream and apply a fixed-weight penetration loss, with no event channel and no governor. This is the do-nothing decentralized deployment. It shows how bad clipping and floating become as $\epsilon$ grows.

Centralized oracle. DHVAE and InterMask run with joint sampling at zero latency. On InterHuman, DHVAE reports penetration volume 0.390 versus InterMask's 0.873, as reported in DHVAE. InterMask reports FID 5.154 on InterHuman and 0.399 on Inter-X. No Contact Ratio value is reported on Lacuna for DHVAE on either dataset, so we will recompute it ourselves. This is not a fair competitor in the latency regime. It is the ceiling, and our target is to approach it as $\epsilon \to 0$ without degrading FID beyond its range.

Prediction-plus-repulsion (strongest decentralized competitor). Each agent forecasts the partner's current pose from the delayed stream using the same reaction backbone trained to predict ahead by $\epsilon$, then applies local signed-distance repulsion with ContactOpt's 2 mm tolerance. It is run on the same backbone, data, latency distributions and hyperparameter search budget as our method. Beating it means lower combined worst-case $\max(p,g)$ at matched FID and contact ratio, at every tested $\epsilon \ge 33$ ms, with the gap widening as $\epsilon$ grows.

Ablations. Three ablations isolate the key ingredients:

  • a constant margin matched to give the same average penetration volume as our method, which tests whether velocity scaling matters;
  • removing the event channel, so contact timing is inferred only from motion;
  • removing the default-to-gap rule under packet loss.

Why not just…?

Why not just predict where the partner is now and avoid it? This is the real threat, which is why it is our main baseline. Forecasting works for smooth motion. Contact onset, however, is where the partner's acceleration changes abruptly, and the forecast error there scales with delay in the same way as the bound. Forecasting reduces the effective $v$ only if the forecaster knows when the partner commits, and that is exactly the information the event channel carries. If prediction-plus-repulsion matches us across $\epsilon$, the honest conclusion is that the event channel is unnecessary, and our ablation without it would show the same thing.

Why not just use a fixed penetration tolerance, as ContactOpt does? A fixed 2 mm tolerance is calibrated for a centralized optimizer with exact poses. Under latency, a constant margin is either too small for fast approaches, producing clipping, or too large for slow ones, producing floating. Our constant-margin ablation at matched penetration volume tests this directly. If floating gap and peak depth come out the same, velocity scaling buys nothing.

Why not just run the coordination at a coarse time step so common knowledge holds? Common knowledge revisited permits this, but it gives no rule for choosing the granularity. Human contact unfolds over tens of milliseconds, so coarse physical steps miss collisions between frames. We adopt coarse granularity only for discrete contact events and keep motion at full frame rate. A coarse-step baseline is trivial to add, and we will report it.

Why not just centralize generation? In the interactive settings we target, centralization requires the very zero-latency channel that does not exist. It also inherits joint-model identity collapse, which the surveyed interaction work documents (3D Human Interaction Generation: A Survey). Centralized generation remains our ceiling, not our competitor.

Proposed Experiments

We evaluate on InterHuman, converting skeletons to SMPL meshes, and on Inter-X, which provides native SMPL-X meshes including hands. Latency is simulated with jitter bounds $\epsilon$ ∈ {0, 33, 66, 100, 200} ms and loss rates $q$ ∈ {0, 1, 5, 10}%.

Metrics include the standard ones: FID, R-precision, penetration volume, and contact ratio. We add per-contact peak penetration depth and floating gap at the agreed contact instant, measured in millimetres.

The first decisive check is whether measured worst-case $\max(p,g)$ for the naive and prediction baselines tracks $v\epsilon/2$ when plotted against empirical closing speed. If it does not, the timing account of clipping is wrong. The second check is whether our method stays near the 2 mm tolerance across all $\epsilon$ while the baselines grow linearly. The third check is whether, under packet loss, our failures shift from penetration to floating, as the default-to-gap rule predicts. The fourth check is a short perceptual study comparing governed and ungoverned approaches. It determines whether the speed cap looks natural, which decides whether the method is practical or only correct.