Multi-expert learning of adaptive legged locomotion

Abstract

Achieving versatile robot locomotion requires motor skills that can adapt to previously unseen situations. We propose a multi-expert learning architecture (MELA) that learns to generate adaptive skills from a group of representative expert skills. During training, MELA is first initialized by a distinct set of pretrained experts, each in a separate deep neural network (DNN). Then, by learning the combination of these DNNs using a gating neural network (GNN), MELA can acquire more specialized experts and transitional skills across various locomotion modes. During runtime, MELA constantly blends multiple DNNs and dynamically synthesizes a new DNN to produce adaptive behaviors in response to changing situations. This approach leverages the advantages of trained expert skills and the fast online synthesis of adaptive policies to generate responsive motor skills during the changing tasks. Using one unified MELA framework, we demonstrated successful multiskill locomotion on a real quadruped robot that performed coherent trotting, steering, and fall recovery autonomously and showed the merit of multi-expert learning generating behaviors that can adapt to unseen scenarios.

Locomotion Control With Frequency and Motor Pattern Adaptations

“…Furthermore, it could be a requirement to learn new motor patterns for new tasks. Motor pattern adaptation has been applied both to CPG-based locomotion controllers (Nakanishi et al, 2004 ; Oliveira et al, 2011 ) and deep neural network locomotion controllers (Clune et al, 2011 ; Hwangbo et al, 2019 ; Lee et al, 2020 ; Schilling et al, 2020a , b ; Yang et al, 2020 ).…”

Meta-Learning for Fast Adaptive Locomotion with Uncertainties in Environments and Robot Dynamics

“…actuator properties, a model-free RL scheme can train different locomotion policies separately and deploy on a real quadrupedal robot [7]. By using a multi-expert learning, an hierarchical RL architecture can learn to fuse multiple motor skills and generate multimodal locomotion coherently on a real quadruped [8]. However, in general, model-free RL algorithms have limited sample efficiency, resulting in long training time to produce viable policies.…”

Design and Implementation of Symmetric Legged Robot for Highly Dynamic Jumping and Impact Mitigation

“…Legged robots have good potential for traversing difficult obstacles, and the execution of highly dynamic maneuvers—such as jumping and running—has attracted a lot of attention. Many state-of-the-art legged robots, such as Atlas [ 1 ], BHR [ 2 ], Bigdog [ 3 ], ANYmal [ 4 ], MIT Cheetah [ 5 ], Jueying [ 6 ], HyQ [ 7 ], SCalf [ 8 ], Minitaur [ 9 ], Stanford doggo [ 10 ], GOAT [ 11 ], and Salto [ 12 ], have achieved significant advancements. Many of the abovementioned legged robots can perform versatile and stable gait.…”