Collisions as Coordination: Resampling Far-Range LiDAR Object Completions That Physically Conflict

At long range a LiDAR sensor sees almost nothing. A pedestrian at 70 meters may return fewer than five laser points (Mitigating Long Range Sparsity in 3D Object Detection). Recovering the full shape of such an object is a one-to-many problem: the same handful of points fits a sedan, an SUV, or a hedge. Generative completion models are the natural tool here. A diffusion model is a generator that learns to turn random noise into samples by gradual denoising, and diffusion models avoid the "blurry average shape" that deterministic completion produces (3D Shape Generation and Completion through Point-Voxel Diffusion). To run fast enough for driving, though, such a model must complete many objects at once and independently. That reintroduces a failure first documented in parallel machine translation. When output slots are sampled without seeing each other, each slot picks a plausible mode on its own, and the combination is incoherent (A Study of Syntactic Multi-Modality in Non-Autoregressive Machine Translation).

This proposal starts from one observation: in 3D scenes this incoherence is physically visible. Two independently completed neighbors that each chose a long-vehicle mode will intersect. A completed car that claims space a laser beam visibly passed through is wrong. An object that ends up hovering above the road is wrong. We propose to treat these physical violations, meaning interpenetration, free-space violation and missing ground support, as the signal that says which parallel slots chose incompatible modes. We then fix only those slots by partially re-noising and resampling them, conditioned on their now-fixed neighbors. This borrows the mask-and-repredict loop of Mask-Predict, except that the confidence used to decide what to re-mask is replaced by a physical-plausibility energy.

Core Claim

The claim has two parts, and both are testable. The diagnostic part says that among independently sampled completions of sparse far-range objects, the amount of physical conflict an object has with its scene predicts its geometric error against ground truth. In other words, collisions mark wrong-mode choices rather than being harmless cosmetic artifacts. The method part says that resampling conflicting objects from an intermediate noise level beats the two obvious alternatives at equal compute: pushing colliding shapes apart by gradient optimization, as in DM-CFO, Layout-your-3D and ContactOpt, and drawing many full samples and rejecting colliding ones. The reasoning is that gradient repair deforms a wrong mode into a distorted or floating shape, whereas resampling can jump to a different mode.

The closest existing work falls into two groups, and each lacks half of this. Collision-aware generators such as DM-CFO and Layout-your-3D add a collision loss to dense, per-scene optimization that takes minutes; DM-CFO reports roughly 5 minutes per tooth arrangement. They treat collisions as something to minimize, not as a signal for choosing a mode. LiDAR completion diffusion (LiDiff) works at metric outdoor scale but imposes no physical constraint at all and stops at about 50 meters. Our proposal applies to the far field, which LiDiff does not cover, and uses physical violations to choose between modes rather than to clean up geometry.

Research Program

Per-object completion model. We build a point-cloud completion diffusion model in the style of PVD. It conditions on the sparse points inside a detected far-range object region and keeps the observed points fixed while generating the rest, following PVD's completion setup. Following LiDiff's finding that normalizing coordinates to a unit box distorts metric geometry, it operates on unnormalized local offsets in the object frame. Because objects are processed as a batch of independent slots, this stage has the same conditional-independence structure as a non-autoregressive decoder. That independence is intended; it buys speed. We distill the model to a small number of denoising steps. Whether a few steps preserve the multiple modes is something we measure rather than assume.

Plausibility energy. For a scene with completed objects $S_1, \dots, S_N$, we define an energy for each object:

$$E_i = \sum_{j \neq i} \mathrm{Pen}(S_i, S_j; \tau_i) + \lambda_f \, \mathrm{Free}(S_i) + \lambda_g \, \mathrm{Float}(S_i).$$

Here $\mathrm{Pen}$ measures how deeply points of $S_i$ lie inside the signed-distance field of $S_j$, beyond a tolerance $\tau_i$. $\mathrm{Free}$ counts completed points that lie on segments the laser traversed before hitting something farther away, which is space known to be empty. $\mathrm{Float}$ measures the gap between the object's lowest surface and the estimated ground. The weights $\lambda_f$ and $\lambda_g$ balance the three terms.

The free-space term matters most at long range. A distant object returns few points, but the beams that grazed past it and hit the background still outline its silhouette. That information is free, and existing completion models ignore it.

Range-scaled tolerance. The tolerance $\tau_i$ addresses the contact-versus-penetration trade-off, which is the open problem named in Resolving Mesh Interpenetration in 3D Reconstruction and Generation. ContactOpt showed that allowing a fixed 2 mm of interpenetration stabilizes optimization for hand–object pairs, but a fixed margin makes no sense when uncertainty grows with distance. We set $\tau_i = \alpha \, r_i \, \Delta\theta$, where $r_i$ is the object's range, $\Delta\theta$ is the sensor's angular beam spacing, and $\alpha$ is a single tuned constant. This form follows directly from the fact that at a fixed angular resolution, the spacing between neighboring beams grows linearly with range. Overlap smaller than the local sampling gap therefore cannot be distinguished from contact and is not penalized.

Collision-triggered resampling. After the first parallel pass, every object whose $E_i$ exceeds a threshold is re-noised to an intermediate diffusion time $t^\ast$. It is then denoised again with the gradient of $E_i$ added as guidance, while its neighbors are held fixed. The loop repeats for a small, fixed number of rounds. This is Mask-Predict's re-mask-the-least-confident rule with physics in place of model confidence. It also mirrors how Non-Autoregressive Neural Machine Translation used an explicit plan (fertility) to coordinate parallel slots. In our case the shared plan is the scene's occupancy.

Analysis. The analysis comes before the method claims. We will measure how often independent completions of far objects collide, and whether collision energy correlates with per-object error. If the correlation is weak, the method's premise fails, and we would report that.

Baselines

The naive reference point is independent parallel completion with no coordination: the same per-object diffusion model, one pass, no energy. Every gain is measured against it, and it shows directly how often the multimodality failure produces physical violations.

The strongest existing method for LiDAR completion is LiDiff (Scaling Diffusion Models to Real-World 3D LiDAR Scene Completion). It denoises the whole scene jointly, so it already lets objects condition on each other. That makes it the most serious competitor to our explicit coordination step. LiDiff reports a Chamfer Distance of 0.375 m on SemanticKITTI, against 1.256 m for PVD. We will rerun the released configuration on our splits with its 50 sampling steps and report its interpenetration, free-space and floating rates, none of which it reports. To beat it, we must match or improve far-field per-object geometric error while producing lower violation rates at equal or lower wall-clock time. On the extended 50–150 m range, LiDiff is untested; we found no reported result on Lacuna for diffusion completion beyond 50 m.

The physics-repair baseline takes the uncoordinated samples and applies gradient descent on the same energy $E_i$ with the same range-scaled tolerance, with no re-noising. This is the DM-CFO and ContactOpt recipe transplanted to LiDAR. It isolates the key ingredient of our method: whether resampling the mode beats deforming the shape.

The rejection-sampling baseline draws $K$ independent samples per object and searches for the lowest-energy combination. $K$ is chosen so that its denoising compute equals that of our resampling loop. A further ablation fixes the energy but replaces energy-triggered re-masking with confidence-triggered or random re-masking at the same budget. This separates "physics picks what to resample" from "resampling more helps."

For downstream relevance we compare against detectors without generative completion. On the Waymo Open Dataset, SWFormer reports 73.36 L2 mAPH at 43 ms on a T4 GPU, using deterministic max-pooling "voxel diffusion" to fill hollow object centers. On Argoverse 2, Far3D reports 24.4% mAP out to 150 m. All methods that share the completion network are trained on the same data with the same tuning budget.

Why not just…?

Why not just complete the whole scene jointly, as LiDiff does? Joint denoising does let objects see each other, and it may already avoid most collisions near the sensor. This is a real threat. However, LiDiff's training targets and evaluation stop around 50 m, it imposes no free-space or contact constraint, and it takes several seconds per scene. Our head-to-head measures violation rates for LiDiff and for our method separately for each range bin. If joint denoising already matches our violation and error numbers in the far field at similar cost, explicit coordination is unnecessary and we will say so.

Why not just push colliding shapes apart with a collision loss? That is what DM-CFO and Layout-your-3D do, and the mesh-interpenetration literature reports the cost: hard repulsion produces floating objects, and soft repulsion leaves visible clipping. DM-CFO also reports adjacent instances sticking together. A shape completed in the wrong mode stays in the wrong mode when deformed. The physics-repair ablation tests this directly. Our hypothesis predicts that it lowers violation rates about as well as resampling but leaves per-object error, and especially the error on the object's unseen portion, largely unchanged.

Why not just draw many samples and reject the colliding ones? At equal compute this could work, and we include it as a baseline for exactly that reason. Its weakness is combinatorial: neighboring objects' choices interact, so independent rejection either ignores those interactions or must search over joint configurations. Re-masking only the conflicting slots spends compute where conflicts occur. If rejection matches us at equal budget, the contribution reduces to the diagnostic finding and the energy design, and we will report it that way.

Why not just accumulate points over time, or fill hollow centers deterministically? Accumulating frames fills static surfaces. But in SemanticKITTI, baseline accuracy on moving classes often falls below 20% mIoU (SemanticKITTI), consistent with moving objects smearing into trails, and surfaces that stay occluded are never observed. SWFormer's deterministic center filling is fast and strong for box regression, but by construction it outputs a single answer where the data allow several. We include multi-frame input as an additional condition and report whether our method's gains persist on top of it.

Proposed Experiments

The central benchmark is a range-transplant protocol that supplies ground truth where none exists. We take densely observed near-range objects from SemanticKITTI, using the accumulated scans that LiDiff also uses for completion targets. We then re-simulate them at 50–150 m by ray-casting the sensor's beam pattern, keeping their true neighbors, the ground, and the background that defines free space. The full near-range shape becomes the target.

We also plant pairs of objects with known close spacing, such as parked cars and pedestrians beside vehicles. This creates controlled situations where neighboring objects compete for the same space.

Because Chamfer Distance rewards blurry averages, we report it for comparability but rely on two other geometric measures. The first is minimum-matching error over samples, which says whether the right mode is among the samples. The second is error on the unobserved part of each object. Physical plausibility is measured by penetration volume beyond tolerance, free-space violation rate and floating rate. All metrics are reported per 25 m range bin, along with latency per scene.

The first decisive check is the diagnostic correlation between $E_i$ and per-object error in the independent pass. The second compares resampling with physics repair on unobserved-part error at matched violation rates. The third places resampling against rejection sampling and LiDiff at matched wall-clock time.

A secondary test uses Waymo and Argoverse 2. Completed far objects are fed to a fixed SWFormer-style detector, and we ask whether 3D detection accuracy improves at 50 m and beyond relative to the uncompleted input. That would show the plausibility gains carry into perception rather than staying cosmetic.