State Dependent Phase Dynamics for Robotic Synchronization

44 concepts29 papers1 proposal

In classical robotics, motion is almost always indexed by wall-clock time ($t$). A trajectory generator outputs a sequence of target positions or torques corresponding to specific, monotonically increasing timestamps. If a robot is programmed to hand an object to a human at $t = 2.0$ seconds, it will execute that trajectory regardless of whether the human is ready, distracted, or moving at a completely different pace.

This rigid reliance on absolute time causes catastrophic failures in real-world, interactive environments. Humans do not operate on fixed internal clocks. As shown by Joelle Pineau et al., fixed-speed navigation routinely fails in assistive contexts because human walking speeds vary widely and unpredictably. Furthermore, Christopher K. Fourie argues in [Adaptive Planning with Evidence Based Prediction for Improved Fluency in Routine Human-Robot Collaborative Tasks] that treating human actions as isolated, reactive events prevents robots from achieving the fluid temporal coordination characteristic of high-performing teams.

To build robots that can seamlessly dance, walk, hug, or assist, we must decouple the spatial definition of a task from its temporal execution. The core idea behind dynamic temporal synchronization is to replace absolute time $t$ with a state-dependent phase variable $\phi$.


From Time to Phase: The Core Intuition

Imagine a pre-recorded trajectory of a hand reaching out to grasp a cup. Instead of indexing this trajectory as $x(t)$, we index it as $x(\phi)$, where the phase variable $\phi \in [0, 1]$ represents the progress of the task.

In their foundational work on motor primitives, Auke Ijspeert et al. demonstrated that replacing absolute time with a state-dependent phase variable makes movement execution invariant to temporal scaling and robust to physical perturbations.

$$\frac{d\phi}{dt} = \dot{\phi}$$

Here, $\dot{\phi}$ is the phase velocity—the speed at which the robot progresses through its task. If the robot is unimpeded and its partner is moving at the expected speed, $\dot{\phi}$ remains nominal.

  • If the robot is physically blocked or its partner pauses: The system sets $\dot{\phi} = 0$. The phase stops advancing, and the robot freezes in place.
  • When the obstruction is cleared or the partner resumes: $\dot{\phi}$ becomes positive again, and the movement picks up exactly where it left off, rather than jumping ahead to a future time step.

Warping Time via Phase Velocity Inference

To synchronize with a human partner, a robot must infer the human's current phase and phase velocity in real-time from sensory data.

[Human Motion] ──> [Recursive Filter / ProMP] ──> Estimates Phase (ϕ) & Velocity (ϕ̇) ──> [Robot Trajectory x(ϕ)]

In [Learning Interactive Behaviors for Musculoskeletal Robots Using Bayesian Interaction Primitives], Joseph Campbell et al. introduced a recursive filtering framework that tracks both phase and phase velocity during physical human-robot collaboration. By treating the human's movement speed as a latent variable, the robot can dynamically stretch or compress its own trajectory timing.

The authors extended this concept to highly intimate, whole-body interactions like hugging in [Learning Whole-Body Human-Robot Haptic Interaction in Social Contexts], demonstrating that tracking latent phase variables allows a robot to handle non-linear timing variations, such as unexpected mid-hug pauses, without requiring explicit pre-programmed triggers. Similarly, Michael Drolet et al. utilized concurrent phase and velocity estimation to blend different interaction styles smoothly on the fly.

Instead of recursive filtering, internal motion models can also be aligned using explicit scaling factors. Alexandros Paraschos et al. showed that model-based estimation of time modulation factors allows Probabilistic Movement Primitives (ProMPs) to rapidly synchronize with a human's pace during the early stages of an interaction, ensuring precise spatial-temporal alignment.


Entrainment and Phase Oscillators

When a task is highly repetitive or periodic (such as walking or rhythmic wiping), synchronization can be framed as entrainment, where two dynamical systems couple their rhythms.

Guilherme Maeda et al. implemented this by utilizing a phase oscillator as a temporal engine. By treating an external moving target's state as the primary clock, the robot's phase oscillator automatically accelerates or decelerates its trajectory to maintain perfect synchronization with unpredictable external stimuli.

This coupling can also be purely physical, driven by bidirectional force feedback rather than vision:

  1. Virtual Reflexes: In wearable robotics, Ali Reza Manzoori et al. bypassed explicit intent-recognition algorithms entirely. They generated assistive torque timing through the continuous interaction between a user's real-time kinematics and a virtual reflex model, allowing an exoskeleton to adapt naturally to changes in gait speed.
  2. Velocity-Proportional Injection: Auke Ijspeert et al. demonstrated that injecting torque proportional to the user's joint velocity creates an inherent synchronization loop. When the user slows down, the system naturally reduces power, preventing the device from fighting the wearer's intent across varying terrains.
  3. Bio-mimetic Stimulation: For clinical rehabilitation, Ruxin He et al. used real-time movement velocity measurements to dynamically adjust the duration and onset of functional electrical stimulation (FES), ensuring that muscle stimulation remains perfectly synchronized with the patient's current gait pace.

Predictive Rhythms and Mutual Adaptation

True temporal fluency mimics human-to-human interaction, where synchronization is predictive rather than purely reactive.

Alessandra Sciutti et al. demonstrated that by calculating a running average of detected motion intervals, a robot can predict the timing of future human actions and adjust its own motor execution parameters proactively. When scaling this to groups, Tariq Iqbal et al. discovered that robots achieve significantly higher perceived fluency by identifying and following the single most synchronous individual in a team, rather than trying to synchronize to a noisy group average.

This mirrors how humans naturally interact. In a study of human dyads by Tamara Lorenz et al., researchers found that when two people perform incongruent movements, they maintain phase synchrony through mutual adaptation—specifically, the unhampered partner increases their "dwell time" at targets to wait for the partner navigating an obstacle.

To embed this predictive capability into robotic control, modern architectures explicitly condition trajectory parameters on external human states:

  • João Silvério et al. conditioned task duration on external sensory inputs, allowing a cobot to stretch or compress its trajectory timing to match human approach speeds.
  • Yasemin Göksu et al. conditioned bimanual robot motions directly on the real-time trajectory of a human receiver to maintain a natural interaction rhythm during handovers.
  • Jason Qin et al. showed that modeling this real-time spatial-temporal alignment is crucial for delicate physical assistance tasks, such as feeding.
  • Haoyu Zhao et al. explicitly modeled periodic phase and rhythm to manage smooth transitions between distinct whole-body humanoid behaviors, such as shifting from standing to walking.
  • Albert Mukovskiy et al. synchronized the blending weights of non-periodic actions (like reaching) with a leading periodic oscillator (like walking) to synthesize natural, predictive full-body movements.
  • Elmar Rueckert et al. demonstrated that the motor system rapidly couples auxiliary postural tasks with primary volitional goals within a very small number of repetitions, highlighting the efficiency of phase-based coupling.
  • Even at the hardware level, Kevin M. Godby et al. showed that modeling the statistical distribution of temporal latencies between commands and sensory feedback allows a robot to robustly distinguish its own actions from external environmental movements.

Robustness Beyond Interaction: Loco-Manipulation

While phase dynamics are intuitive for human-robot interaction, they are equally powerful for autonomous robots interacting with a passive environment.

In [Guided Reinforcement Learning for Robust Multi-Contact Loco-Manipulation], Jean-Pierre Sleiman et al. applied state-dependent phase dynamics to motion imitation in reinforcement learning. When a quadrupedal robot attempts to open a heavy, spring-loaded door, a standard time-indexed policy will fail if the robot slips; the policy continues counting time and tries to walk through the door before it has successfully been opened.

By replacing the fixed temporal clock with adaptive phase dynamics, the policy automatically slows or pauses the phase progression whenever the robot deviates from the reference path due to a slip or a heavy load. The controller is mathematically prevented from attempting future stages of the task until the current physical errors are corrected.

By moving from $x(t)$ to $x(\phi)$, robotics shifts from rigid scheduling to organic synchronization, allowing machines to share the temporal rhythms of the physical world.

Key Foundational Work

Towards robotic assistants in nursing homes: Challenges and results (2003) — contributed 1 concept elements to this direction.
Learning Attractor Landscapes for Learning Motor Primitives (2002) — 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_dynamic_temporal_synchronization_8b9f9061,
  title = {dynamic temporal synchronization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-temporal-synchronization/rdr_8b9f906197b342f99ab721476a62a99c},
  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_dynamic_temporal_synchronization_8b9f9061_v1,
  title = {dynamic temporal synchronization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-temporal-synchronization/rdr_8b9f906197b342f99ab721476a62a99c/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