Ruofeng Wang

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-121

May 17, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-121.pdf

This work introduces DiffuseLoco, a framework for training multi-skill diffusion policies for dynamic legged locomotion from offline datasets, which then enables real-time control of robots in the real world. Learning multiple locomotion skills within a single policy presents a significant challenge in legged control. To address this, offline learning from multi-modal datasets with diffusion models can yield a policy with a rich distribution of locomotion skills. However, due to the larger-scale model and iterative denoising process, diffusion models have their own limitations in achieving real-time control onboard the robot. DiffuseLoco is devel- oped to tackle these issues, utilizing several improvements such as goal-conditioning, receding horizon control, delayed inputs, and an accelerated computing pipeline. We highlight Dif- fuseLoco with its multi-modality in locomotion skills, zero-shot transfer to real quadrupedal robots, and capability in real-time control on edge computing devices. Through over 200 benchmarking in real-world experiments, DiffuseLoco demonstrates better stability and ve- locity tracking performance compared to state-of-the-art baselines. The five skills we bench- marked include walking at three different speeds, as well as turning left and right. In the experiments, our DiffuseLoco policy is capable of switching skills smoothly and exhibits ro- bustness against various environments. In addition, we conduct extensive ablation studies to support the design choices in DiffuseLoco. This work opens a new possibility of leveraging imitation learning to create multi-skill controllers for legged locomotion from offline datasets.

Advisors: Borivoje Nikolic

---

BibTeX citation:

@mastersthesis{Wang:EECS-2024-121,
    Author= {Wang, Ruofeng},
    Title= {Real-Time Legged Locomotion Control with Diffusion from Offline Datasets},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-121.html},
    Number= {UCB/EECS-2024-121},
    Abstract= {This work introduces DiffuseLoco, a framework for training multi-skill diffusion policies for dynamic legged locomotion from offline datasets, which then enables real-time control of robots in the real world. Learning multiple locomotion skills within a single policy presents a significant challenge in legged control. To address this, offline learning from multi-modal datasets with diffusion models can yield a policy with a rich distribution of locomotion skills. However, due to the larger-scale model and iterative denoising process, diffusion models have their own limitations in achieving real-time control onboard the robot. DiffuseLoco is devel- oped to tackle these issues, utilizing several improvements such as goal-conditioning, receding horizon control, delayed inputs, and an accelerated computing pipeline. We highlight Dif- fuseLoco with its multi-modality in locomotion skills, zero-shot transfer to real quadrupedal robots, and capability in real-time control on edge computing devices. Through over 200 benchmarking in real-world experiments, DiffuseLoco demonstrates better stability and ve- locity tracking performance compared to state-of-the-art baselines. The five skills we bench- marked include walking at three different speeds, as well as turning left and right. In the experiments, our DiffuseLoco policy is capable of switching skills smoothly and exhibits ro- bustness against various environments. In addition, we conduct extensive ablation studies to support the design choices in DiffuseLoco. This work opens a new possibility of leveraging imitation learning to create multi-skill controllers for legged locomotion from offline datasets.},
}

EndNote citation:

%0 Thesis
%A Wang, Ruofeng 
%T Real-Time Legged Locomotion Control with Diffusion from Offline Datasets
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 17
%@ UCB/EECS-2024-121
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-121.html
%F Wang:EECS-2024-121