Animal demonstrations → real-world locomotion

OmniMimicDynamics-completed Motion Augmentation for
Multi-style Omnidirectional Quadruped Locomotion

One policy. Four gait styles. Forward, backward, lateral, and turning motion.

Unitree Go2Single multi-gait policyDirect deployment · no fine-tuning

Real-world deployment

Four gaits. One policy.

Trot · Pace · Canter · Pronk

Simulation-trained policy, deployed directly on the real robot without fine-tuning.
Real-time playback (1×). Shorter clips hold their final frame.

Download 4K video

Overview

Natural gait styles,
beyond the demonstrated directions.

Animal demonstrations provide natural, distinctive gait styles, but their narrow directional coverage leaves little style-consistent supervision for backward, lateral, and turning commands.

OmniMimic combines temporal reversal, constrained dynamics completion, and sagittal reflection to construct robot-specific motion and physical supervision beyond the observed directions. A demonstration-guided curriculum progressively expands the command distribution, while a shared actor and softly mixed residual experts learn a single policy for four gaits.

The deployed actor uses robot state, commands, and gait code—without reference motions, physical labels, or guidance torques.

63.1%

Lower command-tracking RMSE

Uniform Cartesian command grid

12.9%

Lower foot-position RMSE

Forward and backward reference velocities

Simulation comparison

Four-gait mean relative to the matched APEX baseline.

These numbers describe simulation evaluations, not the real-world videos.

The framework

From limited demonstrations
to omnidirectional control.

OmniMimic framework: dynamics-completed augmentation supplies physical labels; demonstration-guided commands expand during training; a shared-residual policy learns four gaits with training-only physics supervision and actor-only deployment. — OmniMimic framework. Physical supervision guides training; only the actor is retained at deployment. View full resolution ↗

01 DMA

Dynamics-completed
Motion Augmentation

Time reversal and sagittal reflection extend the motion directions. Constrained dynamics completion supplies contact-force and torque labels for training-only physical supervision.

02 DOCE

Demonstration-guided
Omnidirectional Command Expansion

Training begins near the demonstrated velocities, then progressively expands toward the target command ranges for forward, lateral, and yaw motion.

03 SMP

Shared-residual
Multi-gait Policy

A shared actor produces the complete action. A gait-conditioned soft gate mixes bounded residual corrections, preserving distinct styles within one policy.

More real-world demos