Stance Is Staticness: Contact-Locked Gaussian Avatars on Reconstructed Terrain

When a person walks across a room, each planted foot stops moving relative to the floor for a few hundred milliseconds. Current methods that rebuild a moving person and their surroundings from video ignore this fact. They describe the person as a cloud of small colored blobs (3D Gaussian splatting, a representation that renders very fast by projecting the blobs straight onto the screen) and drive the blobs with an estimated body skeleton. Any drift in the skeleton's global path shows up as feet that glide over the floor, sink into it, or hover above it. Methods that fix foot sliding, in turn, work only on skeletons and assume a perfectly flat floor at height zero. They never touch the rendered geometry or the actual reconstructed ground.

We propose to treat a planted foot as a temporarily static part of the scene. During stance, the Gaussians that form the sole of the foot should behave like background: they hold still in world coordinates and rest on the reconstructed ground surface, whatever shape that surface has. We enforce this with a training-time loss. The loss measures the ground under each foot by probing the scene's own static Gaussians, which is the same kind of surface finding that surface-guided sampling methods use to locate interfaces along a ray. Because the constraint acts only during optimization, rendering speed is unchanged. Because the foot pushes on the ground, it also acts as a sensor that fills in floor regions the camera never saw clearly.

Core Claim

Foot sliding in reconstructed dynamic avatars is a violation of world-frame staticness at the foot–ground interface. The violation lives in the rendered primitives, not only in the skeleton. Our claim has three parts. First, a contact-gated loss on sole Gaussians removes visible sliding and penetration on non-planar terrain where flat-floor and skeleton-only fixes fail. The loss combines zero tangential velocity with zero distance to a ground surface queried from the static Gaussian field. Second, it does so with no loss of render frame rate, and novel-view quality stays within noise of the base avatar model. Third, letting the ground respond to the foot recovers occluded support geometry better than a static-scene reconstruction alone.

Research Program

The base system is deliberately standard. The avatar is SAGA. SAGA binds Gaussians to an SMPL body mesh, then lets them detach to model shoes and clothing, and renders at 60+ FPS. The scene is a static 3D Gaussian field. Transient pixels are masked in the manner of RobustSplat, which delays densification and bootstraps distractor masks so that moving people do not create floaters. One part of RobustSplat's design hurts this task: it labels the whole person as a distractor to be thrown away. That is wrong exactly at the foot–ground boundary, where the static map needs evidence that the foot provides. We therefore exempt contact-labelled foot pixels from the transient mask.

Contact probabilities $c_t \in [0,1]$ come from the temporal convolutional detector of Zou et al.. That detector reads a 7-frame window of 2D keypoints, predicts toe and heel contact, and reports 0.9292 mAP. We use it unchanged as the source of contact labels. We drop its flat-ground ($y=0$) enforcement.

The ground query borrows from FRoG. FRoG computes the median depth along rays through the Gaussians to find a reliable surface for anchor injection. We cast a short ray from each sole Gaussian along the local up direction through the static field only. The alpha-composited median hit gives a signed height $d_S(x)$ above the reconstructed surface. This is a surface-guided query in the sense of Vox-Surf and DINER, which concentrate computation near a located interface. Those methods use the interface to save samples. We use it to define the contact substrate.

Let $\mathcal{F}$ be the set of sole Gaussians, those bound to SMPL sole faces or detached from them. Let $x_i(t)$ be the world position of Gaussian $i$ at frame $t$, and let $P_\parallel$ project a displacement onto the tangent plane of the queried surface. The added loss is

$$ \mathcal{L}_{\text{contact}} = \sum_t \sum_{i\in\mathcal{F}} \Big[ c_t \big( \lVert P_\parallel (x_i(t{+}1) - x_i(t)) \rVert^2 + \lambda\, d_S(x_i(t))^2 \big) + \mu \max(0, -d_S(x_i(t)))^2 \Big]. $$

The first term stops tangential sliding during stance. The second term removes floating and sinking during stance. The third term, a hinge, forbids penetration at all times.

Two design choices matter. The first is that the loss acts on Gaussian positions, including detached ones, rather than on joints. It therefore constrains what is actually rendered, including the shoe surface that SAGA's second stage frees from the mesh. The second is a gradient schedule. Early in training, gradients from $d_S$ flow only to the foot. Later, a small fraction flows to the static Gaussians, so that repeated stance evidence can raise or complete a floor the camera saw poorly. This is a continuous, non-planar counterpart to the contact-guided hallucination in CRISP, which infers hidden support surfaces from body pose but must then fit planar cuboids for its simulator.

Rendering latency enters as a constraint, not a goal. Everything above is training-time, and the deployed renderer is unchanged SAGA plus static splatting. We will report frame rate to confirm this. We will also test one modest acceleration that the staticness view suggests: during labelled stance, sole Gaussians reuse their previous-frame deformation instead of re-querying the body model. We expect the gain to be small, because feet are a small share of the blobs, and we will report it honestly rather than claim it as a contribution.

Baselines

The naive reference point is SAGA with the static scene, trained with no contact handling. This shows how much sliding and penetration current fast avatars produce in scenes. The file lists no exact SAGA PSNR on ZJU-MoCap. It reports only 60+ FPS and 12-minute training, so we will rerun it ourselves.

The strongest competitor on this specific question is a skeleton-level fix: Zou et al.'s two-stage zero-velocity optimization applied to the SMPL trajectory, with the corrected poses then fed to SAGA. This competitor directly targets footskate. Its reported number is on Human3.6M, where MPJPE falls from 71.05 mm to 69.63 mm and sliding is described as nearly removed (Zou et al.). To make the comparison fair, we will use the same contact detector, the same initial body-mesh fits, the same SAGA training schedule, and the same number of iterations. We will also run a variant that replaces its $y=0$ floor with our queried surface, so that we separate where the constraint acts from what ground it uses.

Two further reference points come from the file. For quality on clothed humans, 4DAnyone reports 24.15 PSNR on DNA-Rendering. For speed in dynamic Gaussian rendering, FRoG reports 91.13 FPS on N3DV. Neither has a contact model. No reported result was found on Lacuna for any foot-sliding metric on a Gaussian or radiance-field avatar, so the sliding numbers for every baseline will come from our own reruns.

Two ablations isolate the key ingredients. The first is skeleton-only versus Gaussian-level constraint. The second is a fitted plane versus the queried static-field surface. We count it as beating the strongest baseline if stance-phase world-frame drift of sole Gaussians and penetration depth both drop significantly on non-flat terrain, with PSNR and LPIPS within 0.3 dB and 0.005 of naive SAGA and render frame rate unchanged.

Why not just…?

Why not just correct the skeleton with zero-velocity IK and render with an unchanged avatar? This is the real threat, and it may suffice on flat floors with tight-fitting shoes. It has two weaknesses. In SAGA's detached stage, shoe Gaussians move independently of the mesh under a soft alignment loss, so a locked skeleton does not guarantee locked rendered soles. And the existing zero-velocity method enforces height zero, so on stairs or slopes it either floats or sinks the foot. Our skeleton-only ablation, run with the queried surface, settles the question. If it matches the full method on drift and penetration, the Gaussian-level loss is unnecessary, and we will report that the contribution reduces to terrain-aware skeleton correction.

Why not just fit a ground plane and penalize distance to it? On flat floors this probably works as well as our method, and we expect it to. The proposal differs only where terrain is non-planar. CRISP's own limitation is that planar primitives cannot represent curved or irregular ground. Our plane-versus-surface ablation, stratified into flat sequences and sequences with stairs or furniture contact, makes this explicit.

Why not just run the motion through a physics simulator? CRISP shows this cuts tracking failures from 55.2% to 6.9%, but only after simplifying the scene to cuboids, and it produces no images or novel views. EnliveningGS adds hard contact physics to Gaussians, but needs a pre-rigged tetrahedral mesh and assumes known flat ground. Neither reconstructs from casual video. We will use CRISP-style simulation as an evaluation probe on our outputs, not as a competing pipeline.

Why not just constrain the feet with monocular depth? 4DAnyone reports that sparse, accurate skeletons beat dense, error-prone monocular depth for reconstruction. FRoG notes that mean-depth surface estimates are vulnerable to floaters. Per-frame depth also flickers, and a flickering depth target would itself jitter the stance foot. We will include a depth-prior variant to test this directly.

Proposed Experiments

The primary data are PROX and EMDB. PROX supplies monocular indoor capture with registered scene scans, so penetration depth and floating can be measured against true geometry, including sitting and stepping on furniture. EMDB supplies outdoor sequences with global trajectories, so sliding can be measured in world coordinates. ZJU-MoCap and DNA-Rendering test whether rendering quality survives the constraint, measured by PSNR, SSIM, LPIPS, and FPS on the same GPU class. Human3.6M checks that body accuracy (MPJPE) does not degrade relative to Zou et al.

The sliding metric is the mean world-frame tangential displacement of sole Gaussians per stance frame. We will also report penetration and float distance against scan geometry and Chamfer distance of the recovered floor under occlusion. Finally, we will train a CRISP-style humanoid tracking controller on our recovered trajectories and terrain and report its failure rate.

Three checks are decisive. The first: on flat-floor sequences, the plane baseline should tie our method. If it does not, our gains come from something other than terrain. The second: on non-planar contact sequences, the skeleton-only ablation should leave measurable sole drift, and the plane baseline should leave measurable penetration. If either matches us there, our central claim fails. The third: under occluded floor regions, enabling ground gradients should reduce floor Chamfer error without new floaters, and render FPS must stay equal to the naive SAGA reference.