SLoMo

This repo contains some instructions associated with the paper "SLoMo: A General System for Legged Robot Motion Imitation from Casual Videos". link, arXiv

reconstruction

For reconstruction, we use PPR, a physics-informed reconstruction method. Please refer to the instructions for more details. For general monocular reconstruction, please checkout Lab4d.

We provided some of the trajectories generated from the reconstruction process in the data folder.

offline policy optimization

To optimization an policy/reference trajectory offline, we use examples provided from ContactImplicitMPC.jl to solve a contact-implicit trajectory optimization problem or optimize an imitation-learning policy using the motion_imitation repo.

online control

To control the robot online, we use CI-MPC to track offline trajectories. One can also inference the network trained from imitation learning to generate controls on quadrupeds.

citation

If you find our work useful, please consider citing our paper:

@inproceedings{zhang_slomo_2023,
        title = {SLoMo: A General System for Legged Robot Motion Imitation from Casual Videos},
        url = {https://ieeexplore.ieee.org/abstract/document/10246373},
        author = {Zhang, John Z. and Yang, Shuo and Yang, Gengshan and Bishop, Arun L. and Gurumurthy, Swaminathan and Ramanan, Deva and Manchester, Zachary},
        journal = {IEEE Robotics and Automation Letters},
        year = {2023}
}