Youzhi Liu, Li Gao, Yifei Qian, Liu Liu, Yang Cai, Ziqiao Li
This paper presents a method to automatically generate diverse and expressive motions for quadruped robots using generative AI, without relying on animal data, and successfully deploys them on a real robot.
Quadruped robot motion data is limited and difficult to collect because it depends on observing, reconstructing, and retargeting from animals. This restricts the robot's behavioral repertoire to a few basic gaits, making it hard to achieve rich and expressive motions.
The authors reframe data scarcity as a generation problem, removing animals from the loop. An LLM proposes motion prompts, and a video diffusion model synthesizes the corresponding robot behaviors. To reliably extract 3D reference trajectories from the generated videos, they introduce an Identity Consistency Loss that enforces appearance coherence across frames. The extracted trajectories are used to train tracking policies for deployment on a real robot.
The study releases Quad-Imaginarium, an open-source dataset of 7,488 language-annotated quadruped motions (18.5 hours). Validation on a real Unitree Go2 robot with 392 randomly sampled motions achieved a 96.7% deployment success rate, complemented by a 97.6% success rate across the full dataset in simulation. This demonstrates a new paradigm for generating and deploying rich robot motions without animal data.