Eliminating Temporal Discontinuities in Human Motion Synthesis

48 concepts41 papers1 proposal

When humans move, our actions flow seamlessly from one state to the next. A runner slowing down to a walk naturally absorbs their momentum, adjusting their posture, joint velocities, and foot placements over several intermediate strides. However, when artificial intelligence models attempt to synthesize long-term human motion, they often stumble into the "uncanny valley" of temporal discontinuity. Instead of fluid transitions, characters appear to "teleport," snap instantaneously between poses, or experience jarring joint jitter.

This breakdown in realism occurs because motion generation systems frequently treat long sequences as a series of isolated, independent fragments. Whether stitching together pre-recorded motion capture clips or generating sequences from sequential text prompts, managing the boundary conditions between distinct actions remains a fundamental challenge in computer animation and robotics.

The Boundary Problem in Motion Generation

To understand why temporal discontinuities happen, we must look at how motion data is processed. Many state-of-the-art synthesis frameworks rely on window-based or chunk-based generation. As noted by Yang et al. and Chu et al., generating motion in independent temporal windows frequently introduces visual artifacts right at the boundaries where these windows meet.

This issue is equally prevalent in predictive modeling. When a neural network transitions from processing an observed historical sequence to generating a predicted future sequence, a noticeable "jump" or jerk often occurs at the initial prediction frame. Martinez et al. and Gui et al. highlighted that standard sequence models struggle to maintain physical smoothness across this boundary, violating the basic laws of continuous physical motion.

Isolated Generation:
[ Window A: Sprinting ] ──(Instant Snap / Teleport)──> [ Window B: Lying Down ]

Continuous Generation:
[ Window A ] ───[ Learned Physical Transition Phase ]───> [ Window B ]

The core of the problem is a lack of contextual awareness. As Jiang et al. explained, when a system initiates a new motion segment, it often lacks a mechanism to account for the exact terminal state—such as the final joint angles and velocities—of the preceding segment. Without this information, the model cannot calculate a physically plausible path forward.

Why Naive Stitching and Blending Fail

Historically, the simplest way to combine discrete motion fragments was through concatenative synthesis or linear blending. However, these mathematical shortcuts fail to replicate true human kinetics.

Xu et al. and Duan et al. demonstrated that human perception is highly sensitive to the worst transition in an animation sequence rather than the average quality of all frames combined. A single discontinuous jump completely breaks the illusion of realism. Standard linear blending between two highly disparate poses often results in severe physical artifacts, such as foot skating or awkward joint clipping.

Furthermore, naive concatenation introduces hidden physical assumptions that break fluid movement. For instance, in humanoid robotics, decomposing complex trajectories into simpler building blocks (like individual steps) and stitching them together can introduce artificial zero-velocity assumptions at the transition points, as shown by Lembono et al.. These momentary "stops" prevent the system from maintaining momentum, resulting in robotic, jerky locomotion. Even in mechanical control systems, as observed by Lee et al., relying on piecewise linear models introduces control signal discontinuities at the boundaries between different states.

The Physics of Motion Cycles and Semantic Stages

Human movement is inherently structured around physical cycles (such as walking strides) and semantic stages (such as navigating toward a chair before sitting). Interrupting these cycles mid-stream is a primary source of transition artifacts.

Early work by Ormoneit et al. emphasized that representing cyclic movements requires enforcing periodic constraints so that the conclusion of a cycle aligns perfectly with its start. When stitching together disparate segments to match external cues (like a musical beat), models often cut actions short. Au et al. demonstrated that explicitly enforcing "action completeness"—penalizing transitions that occur mid-action—preserves the integrity of complex physical cycles like spins or waves, significantly reducing visual artifacts.

Mid-Cycle Interruption (High Artifacts):
[ Stride Phase 1 ] ──> [ Stride Phase 2 ] ──(Cut Short)──> [ Sit Down ]

Action Completeness (Smooth Flow):
[ Stride Phase 1 ] ──> [ Stride Phase 2 ] ──> [ Stride Phase 3 (Neutral) ] ──> [ Sit Down ]

Similarly, transitioning between highly contrasting semantic behaviors presents a steep challenge. Jin et al. noted that generative models struggle to create seamless transitions when blending highly contrasting local actions, such as moving from a high-velocity sprint to a stationary state. Without accounting for the varying speeds and underlying dynamics required to bridge these distinct behavioral states, the resulting animations suffer from uncoordinated, abrupt posture changes (Zhang et al.; Tang et al.).

Modern Solutions: Explicit Transition Modeling

To overcome these limitations, modern motion synthesis has shifted away from static concatenation toward dynamic, context-aware interpolation and transition generation.

1. Contextual Refinement and In-Betweening

Instead of assuming a straight line between two poses, researchers treat the gap between actions as an unsupervised generation problem. For example, Tang et al. framed sign language transition generation as a supervised "mask-and-recover" task, forcing the model to synthesize the complex, non-linear coarticulation that naturally occurs between isolated signs. To ensure these boundaries remain clean, the authors of another study introduced a secondary refinement stage that specifically analyzes the terminal frames of a preceding action and the initial frames of the subsequent action to eliminate kinematic discontinuities (Qian et al.).

2. Compositional Diffusion and Blended Encodings

In text-to-motion synthesis, diffusion models have been adapted to generate long-term sequences by natively incorporating transition mechanics. Su et al. developed a framework utilizing "Compositional Transition Sampling," which synthesizes smooth transitions between different text-prompted actions without requiring manual post-processing alignment. To maintain structural continuity over extended durations, Barquero et al. introduced a method that strategically blends absolute and relative positional encodings, allowing periodic motions to extrapolate smoothly without breaking down at chunk boundaries.

3. Dedicated Transition Memory

Rather than forcing a single network architecture to handle both the unique characteristics of an action and the mechanics of switching between them, recent architectures decouple these responsibilities. The authors of a recent study introduced a framework featuring dual memory banks: one dedicated to specific action characteristics and another explicitly for action transitions (Tang et al.). By dynamically adjusting attention based on the required transition speed and dynamics, the model successfully prioritizes short-term boundary smoothness without degrading the long-term distinctiveness of the generated actions.

Key Foundational Work

Implicit Probabilistic Models of Human Motion for Synthesis and Tracking (2002) — contributed 1 concept elements to this direction.
Learning and Tracking Cyclic Human Motion (2000) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_motion_transition_discontinuity_80bf3999,
  title = {motion transition discontinuity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/motion-transition-discontinuity/rdr_80bf39999a704aed97289d9c4c074a47},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_motion_transition_discontinuity_80bf3999_v1,
  title = {motion transition discontinuity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/motion-transition-discontinuity/rdr_80bf39999a704aed97289d9c4c074a47/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake