PaperlayerView exact revision on arXiv

Source-linked AI summary

PhaseSync-Exo: Human Clock Anchored Reference Adaptation for Dynamic Gait Tracking

Kaijie Qi, Yuehan Wang, Kaiming Xu, Chong Li, Jiakuo Yu

arXiv:2609.38824v1cs.RO

TL;DR

Human-aware exoskeleton walking must preserve gait timing while tracking diverse motions under dynamic constraints. PhaseSync-Exo combines wearable-sensing reconstruction, factorized retargeting with recurrent control, and a human-clock-anchored adapter. Its simulation comparisons report lower human-clock phase error than robot-relative correction and fixed-clock continuation, with lower transient hip RMSE than immediate reset and no falls across timing rollouts.

  • Problem

    Exoskeletons must preserve human gait timing while allowing motion to remain dynamically feasible, since mistimed assistance can oppose the wearer and direct trajectory replay can violate physical constraints.

    Source: §I · p. 1

  • Method

    PhaseSync-Exo combines causal two-IMU gait reconstruction, factorized amplitude–cadence retargeting with recurrent control, and a human-clock-anchored adapter that adjusts reference rate.

    Source: §I · p. 1

  • Results

    HCA reduces human-clock phase MAE from 72.02° to 13.90° versus fixed-clock continuation, while all 420 timing rollouts complete without falls.

    Source: C. PhaseSync Exo Method · p. 6, §IX · p. 7

  • Takeaways & Limitations

    The simulation results support continuous phase acquisition and sustained alignment to an independent human clock without retraining the dynamics policy.

    Source: §IX · p. 7

Abstract

from arXiv · show

Human-aware exoskeleton walking requires reconstructing gait, tracking diverse motions under dynamic constraints, and preserving human timing. We present PhaseSync-Exo, which combines two-IMU CNN-Transformer reconstruction, factorized amplitude-cadence retargeting with curriculum-trained recurrent control, and a human-clock-anchored adapter (HCA). HCA combines human-clock attraction with robot-relative feedback to adjust reference rate while preserving forward progression and continuity. The reconstruction module achieves a mean absolute error (MAE) of 3.24 degrees on a held-out recording, and the frozen tracking policy completes 356 of 357 amplitude-frequency trials. Two complementary dynamic comparisons isolate HCA's timing benefit without retraining. Against robot-relative correction, HCA reduces human-clock phase MAE from 95.55 degrees to 10.99 degrees, limiting reference drift. When human and delivered phases initially differ, it reduces phase MAE from 72.02 degrees to 13.90 degrees versus fixed-clock continuation. Compared with immediate phase reset, HCA reduces transient reference-tracking hip RMSE by 22.7% without reference jumps. All 420 timing rollouts complete without falls. These simulation results support continuous phase acquisition and sustained alignment to an independent human clock.

I. INTRODUCTION

PhaseSync-Exo addresses the need to preserve human gait timing while keeping exoskeleton motion dynamically feasible. It separates motion diversity, temporal alignment, and dynamic control through gait reconstruction, factorized retargeting, recurrent control, and a human-clock-anchored adapter.

Motivation

  • The central challenge is preserving a human gait’s timing and hip–knee pattern while allowing dynamically critical robot motion to change.Upright stability alone can oppose the wearer if assistance arrives at the wrong phase, while direct replay can violate contact, actuator, and balance constraints.Source: §I · p. 1

Approach

  • PhaseSync-Exo separates motion diversity, temporal alignment, and dynamic control.The hierarchy uses a motion descriptor to select amplitude–frequency conditions, a phase synchronizer to select current and preview frames, and a GRU–PPO controller to realize them under physical constraints.Source: §I · p. 1
  • A causal two-IMU CNN–Transformer reconstructs continuous hip, knee, and ankle kinematics as human motion input.Evaluation on a held-out recording separates reconstruction performance from downstream controller tracking accuracy.Source: §I · p. 1
  • The factorized multi-trajectory framework separates joint excursion from cadence and couples exact-FK retargeting with recurrent asymmetric actor–critic control.Its complementary curricula train joint amplitude–cadence variation or wide-frequency coverage, with the latter evaluated for transfer to unseen amplitudes without policy changes.Source: §I · p. 1
  • The human-clock-anchored adapter combines circular human and robot phase errors in a positive bounded reference-rate update.Frozen-policy comparisons assess drift suppression against robot-relative correction, continuous phase acquisition against fixed-clock continuation, and transient tracking error against immediate phase reset.Source: §I · p. 1

II. RELATED WORK

The related work frames PhaseSync-Exo as combining physics-based motion imitation with explicit motion-family parameterization and circular timing correction. Its learning controller realizes dynamics, while synchronization remains an interpretable mechanism outside the learned policy.

A. Physics-Based Motion Imitation

  • PhaseSync-Exo parameterizes a retargeted motion family and closes the timing loop around it, keeping reference diversity and synchronization explicit rather than implicit in a trajectory-specific policy.The cited physics-based imitation approaches use motion data to supervise dynamics.Source: A. Physics-Based Motion Imitation · pp. 1–2

B. Reinforcement Learning for Legged and Exoskeleton Control

  • Reinforcement learning realizes dynamics rather than defining the desired human frame; synchronized references enter the actor, while privileged simulator state is confined to the training critic.This separation preserves an interpretable timing mechanism outside the learned controller.Source: B. Reinforcement Learning for Legged… · p. 2

C. Gait Phase and Human–Robot Coordination

  • PhaseSync-Exo couples circular phase comparison to a parameterized retargeted motion bank and constrains correction at the cursor-rate level.A scalar phase estimate alone does not specify which member of a motion family to track, and nearest-frame correction can violate temporal continuity.Source: C. Gait Phase and Human–Robot… · p. 2

III. DESIGN OF THE PROPOSED EXOSKELETON

The system combines a six-joint exoskeleton, wearable-sensing gait reconstruction, and a factorized retargeted motion bank controlled by a recurrent policy. Exact robot kinematics and curriculum training support tracking across motion amplitudes and cadences.

III. DESIGN OF THE PROPOSED EXOSKELETON

  • The exoskeleton has six actuated sagittal joints, and its retargeted library stores joint, torso, and tracked-body references at 60 Hz.The gait descriptor specifies cadence and hip, knee, and ankle amplitude scales; bounded policy actions operate under contact, actuator, and balance constraints.Source: §III · p. 2

IV. HUMAN GAIT KINEMATICS RECONSTRUCTION

  • Two thigh- and shank-mounted IMUs provide 18 channels to a CNN–Transformer, which uses a trailing 32-sample window to estimate sagittal hip, knee, and ankle angles.The module reconstructs joint kinematics from recordings without activity or phase labels for downstream reference generation.Source: §IV · p. 2

V. FACTORIZED MULTI TRAJECTORY TRACKING

  • Exact robot forward kinematics recomputes tracked-body poses from scaled joint angles, while a bounded root-height correction preserves the source foot-height profile.Joint excursion and cadence are factorized; the amplitude–cadence curriculum varies both, whereas wide-frequency training uses nominal amplitude and expands cadence coverage.Source: §V · p. 2
  • A recurrent actor tracks current and preview references using a 70-dimensional observation, with PD control executing bounded residual joint-position commands under motor limits.A separate privileged-state critic supports PPO training, while deployment requires only the actor and joint controller.Source: §V · pp. 2, 4

VI. PHASESYNC EXO METHOD

The synchronization interface distinguishes human, delivered-reference, and estimated robot phase in cycles. It uses the human estimate as an independent timing signal and updates the delivered cursor through a bounded adapter.

A. Reference Phase and Synchronization Interface

B. Proprioceptive Robot Phase Estimate

The robot phase estimator matches encoder-derived joint positions and velocities to nearby motion-cycle candidates. The bounded phase correction combines robot-relative feedback with human-clock attraction to adjust reference rate while preserving forward progression.

B. Proprioceptive Robot Phase Estimate

  • The estimator selects among 29 candidates within ±0.22 cycle by matching joint positions and velocities, then filters the normalized winning offset with EMA coefficient 0.2.Hip, knee, and ankle weights are 1.0/1.25/0.5, with 0.30 rad/2.50 rad/s normalization; locally unwrapped queries handle cycle boundaries.Source: B. Proprioceptive Robot Phase Estimate · p. 4

C. Human Clock Anchored Bounded Phase Correction

  • RRC applies gain 1.5 to filtered robot-minus-delivered phase error, but persistent correction can accumulate human-clock drift despite positive reference progression.RRC has no attraction to human timing; HCA instead uses a separate circular human-minus-delivered error.Source: C. Human Clock Anchored Bounded Phase… · p. 4
  • HCA combines human-clock attraction and robot-relative feedback, with gains of 6/1.5 and a correction bound of b = 0.10.Robot lag slows the reference, while human-clock attraction pulls it back; the policy and robot estimator remain unchanged.Source: C. Human Clock Anchored Bounded Phase… · p. 4
  • With positive human cadence and b<1, the cursor progresses forward; locally, human-reference error has a restoring term, while filtered robot error acts as a disturbance and constant bias leaves finite offset.Capability-change experiments measure tracking, human-clock error, and recovery with the same policy and adapter gains.Source: C. Human Clock Anchored Bounded Phase… · p. 4

D. Synchronized Frame Query

The synchronized cursor queries the selected motion variant at its corrected rate. Interpolated current and preview commands remain temporally consistent, while the recurrent policy realizes them under physical constraints.

D. Synchronized Frame Query

  • Joint positions are linearly interpolated between adjacent frames, while body quaternions use shortest-path normalized interpolation.Reference velocities are scaled by the effective playback rate consistently with cursor advancement.Source: D. Synchronized Frame Query · p. 4
  • The reference query uses the synchronized cursor to select adjacent frames from the chosen motion variant.The reference velocities share the corrected cursor rate.Source: D. Synchronized Frame Query · p. 4
  • Preview frames at +2 and +5 policy steps use the same corrected rate, giving the actor temporally consistent current and future commands.The recurrent policy must realize these commands under contact, balance, and motor constraints.Source: D. Synchronized Frame Query · p. 4

E. Multi Reference Recurrent Asymmetric Actor Critic

The recurrent actor receives reference, proprioceptive, and preview information, while a privileged critic is used only during training. Actions are mapped to residual position targets and trained with motor-limit and stability objectives.

E. Multi Reference Recurrent Asymmetric Actor Critic

  • The actor uses a 70-D observation and a 256-unit GRU followed by 512–256–128 ELU layers and a Gaussian six-action head; the 139-D privileged critic is not deployed.The 39-D base interface includes reference and encoder state, torso information, and previous action; previews and gait-conditioning variables add 31 dimensions.Source: E. Multi Reference Recurrent… · p. 4

F. Action Mapping Motor Limits and Stability Objectives

  • Clipped actor outputs map to residual position targets around the default pose.Targets update at 60 Hz and are tracked by joint PD control at 240 Hz.Source: F. Action Mapping Motor Limits and… · p. 4
  • The reward combines body-reference and direct-joint tracking with action-rate, saturation, joint-limit, contact, torque, and power terms; hip and knee weigh more in phase matching, while all six joints contribute to balance.Training enforces the modules' torque–speed envelope.Source: F. Action Mapping Motor Limits and… · p. 4

VII. TRAINING AND EVALUATION DESIGN

The estimator is evaluated on synchronized human gait recordings, while simulation training uses separate motion and dynamics curricula with broad environment randomization. The reported reconstruction result is measured over all test windows.

A. Simulation and Curriculum Training

  • Simulation training uses 1024 environments, 240 Hz physics, 60 Hz control, and 10 s episodes; PPO collects 24 steps per environment and performs five epochs with four minibatches.The motion and dynamics variations follow separate curricula, and evaluations freeze the WFC policy at iteration 39,999.Source: A. Simulation and Curriculum Training · pp. 4–5

B. Randomization and Deployment Interface

  • Per-environment randomization covers friction, restitution, torso center-of-mass and motor zero offsets, external pushes, sensor bias and noise, and 0–2-step sensing delay.Each foot uses six independent contact pads to reduce simulator dependence of large-area contact.Source: B. Randomization and Deployment… · p. 5

A. Human Gait Kinematics Reconstruction

  • Five synchronized walking recordings contain 21,772 raw samples at approximately 19.75 Hz, with two IMUs providing 18 input channels paired to hip, knee, and ankle angle targets.Source: A. Human Gait Kinematics Reconstruction · p. 5
  • 3.242° MAE, 5.607° RMSE, and R² = 0.819 are reported across 4,267 test windows, versus 9.754°, 13.280°, and −0.016 for the training-mean baseline.The reported reductions are 66.8% in MAE and 57.8% in RMSE.Source: A. Human Gait Kinematics Reconstruction · p. 5

B. Factorized Multi Trajectory Tracking

The factorized tracking evaluation spans amplitude variants, cadence settings, and starting phases, with WFC tested for transfer beyond its nominal-amplitude training.

B. Factorized Multi Trajectory Tracking

C. PhaseSync Exo Method

With the tracking policy frozen, HCA is tested for maintaining alignment to the human clock and reacquiring initially offset phases; the reported advantage is timing-specific.

C. PhaseSync Exo Method

  • 22.7%: HCA reduced 0–2 s delivered-reference hip RMSE versus immediate phase reset, from 4.23° to 3.27°, without reference jumps.The immediate reset produced reference jumps in all 80 trials, including 40 backward assignments.Source: C. PhaseSync Exo Method · p. 6
  • All 420 timing-study rollouts completed without falls or tracking termination.Source: C. PhaseSync Exo Method · p. 6
  • Actuator derating confirmed continued walking, but not a uniform accuracy or recovery gain; the established advantage concerns reference-clock consistency and continuous phase acquisition.Source: C. PhaseSync Exo Method · p. 6
  • 88.5%: HCA lowered human-clock phase MAE versus RRC from 95.55° to 10.99° when clocks coincided at onset.Delivered-reference hip RMSE stayed close at 3.36°/3.42°, while mean final reference lag fell from 0.905 to 0.018 cycle.Source: C. PhaseSync Exo Method · p. 6, Table IV · p. 7
  • 80.7%: HCA reduced source-phase MAE versus FC continuation from 72.02° to 13.90° in initially misaligned conditions.Steady MAE fell 84.8% to 10.98°; all 80 paired conditions improved in both phase MAE and human-trajectory hip RMSE.Source: C. PhaseSync Exo Method · p. 6, Table V · p. 7

IX. CONCLUSION

PhaseSync-Exo combines gait reconstruction and multi-trajectory tracking with a human-clock-anchored interface. Its simulation results support drift suppression and continuous phase acquisition without retraining the dynamics policy.

IX. CONCLUSION

  • The estimator achieved 3.24° MAE on a held-out recording, and the frozen WFC policy completed 356/357 amplitude-frequency trials.Source: §IX · p. 7
  • All 420 timing rollouts completed without falls; the findings provide a simulation basis for subsequent physical human–exoskeleton validation.Source: §IX · p. 7
  • HCA reduced human-clock phase MAE by 88.5% versus RRC and by 80.7% versus FC continuation, and lowered transient hip RMSE by 22.7% versus immediate reset without reference jumps.Source: Figure 9 · p. 7, §IX · p. 7
Loading 2609.38824v1…