The previous locomotion approach worked well in terms of sim-to-real transfer: the policy could track velocity and handle pushes robustly. However, the gait was still slightly limping and conservative, and further improvement required difficult reward tuning. To address this, we explore locomotion with motion priors, where the policy learns more explicit gait patterns from motion data, reducing reward hand-tuning and enabling more controlled motion. While some imitation-based methods require motion data to be included in the observation space or fed as specific reference trajectories, this can bias the policy even when imitation is not desired. AMP is a better fit because it uses a motion discriminator to evaluate the overall motion distribution rather than tracking a specific frame, requires no imitation data in the observation space, and keeps imitation rewards separate from task rewards so their influence can be balanced.
- How to generate motion prior data (kimodo, AMASS, OMOMO etc)
Motion prior data contains desired motion examples, such as walking or turning, without including full robot physics, dynamics, or actuator constraints. This data can come from human motion datasets such as AMASS, object-interaction datasets such as OMOMO, or generated motion from models like Kimodo. To generate a wide range of motions covering different walking directions, we chose Kimodo.
- How to retarget to Asimov (GMR)
After generating the motion data in smplx format, we retarget it to Asimov using GMR (general motion retargeting). GMR maps the source human motion onto Asimov’s robot morphology by matching key body poses, such as pelvis, feet, and upper-body motion, while respecting Asimov’s joint structure and physical constraints.
Figure5. General motion retargeting on Asimov_1
- Adversarial motion prior
AMP is the core approach used to guide the policy toward more natural gaits. The total reward is composed of task rewards from the reward function and AMP style rewards from the motion prior. To balance task performance and gait style, we use a 0.7 task reward and 0.3 style reward ratio. AMP trains a discriminator to distinguish between motion prior data and policy-generated motions. If the policy motion looks similar to the motion prior distribution, it receives a higher style reward.