Foot Sliding as a Grazing Artifact

Motion generators that turn text or music into 3D character animation still make feet skate across the floor while the character is supposed to be standing on them. The standard fixes all take a contact label ("this foot is planted now") and then force the foot's horizontal speed to zero, through a loss, a post-hoc edit, or a physics simulator. This proposal starts from a different diagnosis, borrowed from the mathematics of systems that switch between regimes at a boundary. A real foot hits the ground transversally: it crosses the floor with a clear downward speed, and its sideways speed is then cut off almost instantly. A smoothed neural trajectory instead approaches the floor tangentially, skimming it with near-zero downward speed and leftover sideways speed. Dynamicists call this a grazing contact, as opposed to a passable, or crossing, one. Our claim is that foot sliding is mostly a grazing artifact. The generator turns crossing touchdowns into grazing ones, and the skid is the residue of that smoothing.

This diagnosis makes predictions that the usual framing does not. Sliding should concentrate at touchdowns with a shallow approach angle. Height-threshold contact labels should be least reliable at exactly those touchdowns. The high-frequency jitter that physics simulators introduce when they clean up generated motion should also cluster there, because grazing contacts against rigid boundaries are where impact "chatter" arises. If these predictions hold, the fix is to restore transversal touchdowns, not to freeze feet after the fact. We propose to test the diagnosis first and then build a touchdown-level training signal and event-aligned foot parameterization around it.

Core Claim

We define the touchdown angle of a foot event as $\theta = \arctan(|\dot z| / \|\dot{\mathbf{x}}_h\|)$. Here $\dot z$ is the foot's vertical velocity and $\dot{\mathbf{x}}_h$ its horizontal velocity, both measured at the first frame the foot reaches the floor. $\theta$ near zero means grazing and large $\theta$ means a clean crossing.

The claim has three parts. First, real motion-capture touchdowns have $\theta$ bounded away from zero, while generated touchdowns pile up near zero. Second, the skid distance per stance phase falls steeply as $\theta$ grows. Third, supervising touchdowns to be transversal, with rapid tangential arrest after crossing, reduces sliding with less of the freezing and jitter side effects than zero-velocity methods.

The mechanism is simple. If a network approximates an instantaneous horizontal-velocity reset with a ramp of temporal width $w$, the foot travels roughly $\tfrac12 v_h w$ after touchdown, where $v_h$ is the pre-contact horizontal speed. That figure is just the area under a linear ramp from $v_h$ to zero. The same smoothing also flattens the vertical approach, which is what drives $\theta$ toward zero.

Impact-oscillator theory supplies the reason labels break down. Intermittency in impact oscillators close to resonance shows that at grazing, meaning contact at exactly zero normal velocity, whether an impact happens at all is maximally sensitive to perturbation, with impact velocity scaling as the square root of the distance from threshold. The motion analogue is that a thresholded contact label on a grazing foot flips under tiny perturbations. Methods that key everything off such a label are therefore least reliable where sliding actually happens.

What Exists and How This Differs

The non-smooth dynamics side provides vocabulary and constraints, not a method. AN ANALYTICAL PREDICTION OF PERIODIC FLOWS IN THE CHUA CIRCUIT SYSTEM gives explicit conditions that distinguish grazing from passable flows at a switching boundary. It also argues that eigenvalue-based analysis misses transitions that happen at the boundary. We reuse its grazing/passable distinction as a per-event measurement.

Chattering and related behaviour in impact oscillators shows that low-velocity impacts on rigid boundaries cascade into many rapid impacts before coming to rest. We use it to predict where simulator jitter should concentrate.

Both papers work in one- or two-dimensional idealized oscillators. We do not attempt analytical bifurcation theory for a full articulated body, which the file's own open-problems list flags as intractable. We use the theory only at the level of single foot events, where the one-dimensional reduction is honest.

On the motion side, three lines of work attack sliding. The first is contact labels plus zero-velocity losses, as in Combining Recurrent Neural Networks and Adversarial Training for Human Motion Synthesis and Control, and label-driven retargeting penalties in PHUMA: Physically-Grounded Humanoid Locomotion Dataset. The second is reward fine-tuning with foot-ground guidance and an anti-freezing reward in Skeleton2Stage: Reward-Guided Fine-Tuning for Physically Plausible Dance Generation. The third is physics-simulator projection during sampling in PhysDiff: Physics-Guided Human Motion Diffusion Model, or after estimation in MultiPhys: Multi-Person Physics-aware 3D Motion Estimation.

All three share two features. They treat contact as a binary state and act on the stance phase by zeroing velocity. None of them measures or shapes the approach to contact.

Two of these papers also report problems our diagnosis explains, and so motivate this work. Skeleton2Stage found that physics rewards push the model toward standing still. MultiPhys found that simulator refinement cut skating but raised acceleration error relative to its kinematic input. Our proposal differs in acting on the geometry of the touchdown event itself: its angle and its tangential arrest time. It also treats the height-threshold label as a noisy quantity whose unreliability we can predict.

Research Program

The work proceeds in three stages. Stage three is committed only if stage one confirms the diagnosis.

Stage one: event-level diagnosis. We extract every foot touchdown from ground-truth and generated motion on HumanML3D (Generating Diverse and Natural 3D Human Motions from Text), AIST++, and the PHUMA locomotion corpus. The generators are a discrete-token model, Generating Human Motion from Textual Descriptions with Discrete Representations, a text-to-motion diffusion model, and the EDGE dance model used by Skeleton2Stage.

For each event we record:

  • the touchdown angle $\theta$;
  • the tangential arrest time, meaning frames until horizontal foot speed falls below a fixed tolerance;
  • the skid distance during stance;
  • label stability, meaning how often a height-threshold contact label flips when the trajectory receives small Gaussian noise.

We fit skid distance against $\theta$ separately for real and generated data. We also compare how $\theta$ is distributed in real versus generated touchdowns.

Stage two: simulator jitter localization. We pass generated motions through a simulator tracking controller like those in PhysDiff and MultiPhys. We then attribute excess acceleration error to time windows around touchdowns, binned by input $\theta$, and count contact make/break cycles per touchdown as a direct chatter measure.

Stage three: the method. It has two pieces.

The first is a transversality objective applied at touchdown events. It penalizes $\theta$ below a floor estimated from the real-data distribution, and it penalizes tangential arrest time above the real-data median. The objective acts only at events, so a character that never steps is neither rewarded nor punished by it. It also never asks a moving foot to stop mid-air.

The second is an event-aligned foot parameterization for the diffusion generator. The model outputs a continuous touchdown time per foot, not a per-frame label. During the predicted stance, the foot's world position is copied from its position at that time, a hard reset rather than a soft penalty. The rest of the leg is solved by a differentiable inverse-kinematics layer inside training, so the body learns to be consistent with planted feet.

Both pieces are applied by fine-tuning existing pretrained generators with matched budgets. This keeps the comparison about the idea, not about architecture.

Baselines

The naive reference point is the unmodified pretrained generator, meaning EDGE on AIST++ and the text-to-motion models on HumanML3D, evaluated with our event-level metrics and the standard ones.

The strongest existing method on dance is Skeleton2Stage. It fine-tunes EDGE with a simulator imitation reward, a foot-ground deviation reward with test-time guidance, and an anti-freezing reward. It reports reducing foot-ground deviation on EDGE from 11.85 to 2.51, a 70% human win rate over EDGE, and 3s inference versus 30s for simulator projection. It is the strongest competitor because it already avoids online simulation and explicitly fights freezing, which is exactly our claimed advantage. We will reuse its released setup if available, or reimplement it with the same EDGE checkpoint, the same fine-tuning steps and reward-model compute, and the same hyperparameter search budget as our method. Beating it means lower foot-ground deviation than 2.51 in our reproduction, at equal or better FID and motion velocity, with a lower skid distance at low-$\theta$ touchdowns specifically.

On text-to-motion, the strongest existing method is PhysDiff, which reports an 86% reduction in its physical error metric on HumanML3D and more than 20% FID improvement. We compare against it where its projection can be reproduced. We also test our fine-tuned model as input to PhysDiff-style projection, since the two are complementary.

A strong heuristic baseline, and the most direct threat, is foot-locking with inverse-kinematics cleanup. It uses the same contact detector our method uses at evaluation, and the same frame budget.

A zero-velocity loss baseline fine-tunes with $\sum_t c_t \|\dot p^{xy}_{foot,t}\|^2$ under identical steps.

Two ablations isolate the key ingredient. One uses the event-aligned parameterization without the transversality term. The other uses the transversality term with thresholded per-frame labels in place of continuous touchdown times. If the transversality term adds nothing over the parameterization alone, the grazing lens was diagnostic but not prescriptive, and we will report it that way.

Why not just…?

Why not just lock the feet with inverse kinematics after generation? PhysDiff reports that when generated motion is too far from physical, "a simple post-hoc fix often results in jerky, unnatural movements." Our diagnosis sharpens why. Post-hoc locking relies on a height-threshold label, and near grazing that label is unstable, so locks switch on and off erratically at exactly the touchdowns that skid. This is a real threat on clean locomotion, where labels may be reliable. Stage one's label-stability measurement and the matched foot-lock baseline will show whether the gap is confined to low-$\theta$ events. If foot-locking matches us overall, the contribution shrinks to the diagnosis.

Why not just add a zero-velocity loss during training? It acts on stance, not on the approach, so it leaves the grazing approach intact and asks the network to stop a foot that is still skimming. Skeleton2Stage's need for a separate anti-freezing reward suggests such penalties are easiest to satisfy by moving less. We will measure mean body velocity alongside sliding for this baseline. Our prediction is that it reduces skid only by shrinking $v_h$, while our method reduces it at matched velocity.

Why not just project through a physics simulator? Simulator projection works; PhysDiff's numbers are strong. But it is roughly ten times slower per Skeleton2Stage, and MultiPhys reports higher acceleration error than its kinematic input. Stage two tests whether that jitter is concentrated at grazing touchdowns. If it is, fixing touchdowns upstream should cut simulator jitter too, and the fair comparison is simulator-alone versus our model plus simulator.

Why not just train on more data? PHUMA reports that unfiltered internet video at scale, the Humanoid-X data, gave only 50.6% imitation success versus 92.7% for physically curated PHUMA, with skating around 50% before curation. Scale without contact-aware curation does not remove the artifact. We will also check whether PHUMA-curated data shifts the generated $\theta$ distribution, which would indicate that curation works partly by restoring transversality.

Decisive Checks

The first check is the kill switch. If generated touchdowns do not have lower $\theta$ than real ones, or if skid distance is not concentrated at low $\theta$, the grazing thesis is wrong and stage three is not pursued. The second check is whether simulator acceleration error localizes to low-$\theta$ touchdowns, supporting the chatter explanation for MultiPhys-style jitter. The third check is whether our fine-tuned EDGE beats the reproduced Skeleton2Stage on foot-ground deviation and on low-$\theta$ skid at matched FID and matched mean body velocity, the last being the test of freezing. On PHUMA, we will report the non-skating metric against its reported improvement from about 50% to about 90%. No reported result was found on Lacuna for PopDanceSet breakdowns, so there we will report only our own reproductions.