Learn a movement prior
A low-level policy learns to imitate dog motion on flat ground. It is then frozen, preserving a reusable set of movement skills.
Motion Priors Reimagined: Adapting Flat-Terrain Skills for Complex Quadruped Mobility
Flat-ground dog motion becomes a reusable skill layer, with learned joint corrections for rough terrain and local navigation.
Open the original interactive walkthrough and transcript ↗
This video is embedded from Takahiro Miki’s paper explainer. The original page includes interactive animations, a transcript, and a detailed walkthrough with the paper’s figures.
Reading the results
The quoted 75–95 successes in 100 trials evaluate the student policy with noisy observations in simulation. Hardware demonstrations show deployment on ANYmal-D. The paper also reports that the high-level policy can favor one gait rather than reliably switch between all learned gaits.
The key ideas
Adapt through residuals
A goal-conditioned high-level policy supplies latent motion commands and joint-level residual corrections. Penalizing large residuals balances terrain adaptation with fidelity to the learned movement patterns.
Transfer to the robot
A student learns from the teacher using noisy simulated observations, before deployment on the real robot. Hardware experiments demonstrate stair traversal and local navigation among obstacles.