Research Analyzer
← Back ICRA 2026

DynaFlow: Dynamics-Embedded Flow Matching for Physically Consistent Motion Generation from State-Only Demonstrations

Sowoo Lee, Dongyun Kang, Jaehyun Park, Hae-Won Park

PDF

AI summary

Key figure (auto-extracted from paper)
DynaFlow generates physically consistent, deployable robot motions directly from state-only demonstrations by embedding a differentiable simulator into a flow matching model.
Flow matching differentiable simulation physically consistent motion state-only learning quadruped robotics generative control

Problem

Generative motion models often produce physically inconsistent trajectories and rely on scarce action data or complex auxiliary controllers, hindering real-world deployment.

Approach

The method integrates a differentiable physics simulator directly into a flow matching framework, using analytical gradients to learn actions from state-only data while guaranteeing dynamic feasibility by construction.

Key results

  • Generates strictly dynamically feasible trajectories from inconsistent kinematic data
  • Learns deployable action sequences end-to-end without ground-truth action labels
  • Successfully reproduces diverse gaits and executes long-horizon open-loop motions on a physical quadruped robot
  • Translates infeasible retargeted motion capture data into executable behaviors on hardware

Why it matters

Enables scalable, physically reliable motion generation for robotics without costly action data or complex control pipelines, bridging the gap between kinematic data and real-world deployment.

Abstract

This paper introduces DynaFlow, a novel frame- work that embeds a differentiable simulator directly into a flow matching model. By generating trajectories in the action space and mapping them to dynamically feasible state trajectories via the simulator, DynaFlow ensures all outputs are physically consistent by construction. This end-to-end differentiable archi- tecture enables training on state-only demonstrations, allowing the model to simultaneously generate physically consistent state trajectories while inferring the underlying action sequences required to produce them. We demonstrate the effectiveness of our approach through quantitative evaluations and showcase its real-world applicability by deploying the generated actions onto a physical Go1 quadruped robot. The robot successfully reproduces diverse gaits present in the dataset, executes long- horizon motions in open-loop control and translates infeasible kinematic demonstrations into dynamically executable stylistic behaviors. These hardware experiments validate that DynaFlow produces deployable, highly effective motions on real-world hardware from state-only demonstrations, effectively bridging the gap between kinematic data and real-world execution.

Index terms

Learning from Demonstration Imitation Learning

Related papers