Quadruped Agile Locomotion: Sparsely Gated Mixture of Experts with Visual Input
Quadruped Agile Locomotion: Sparsely Gated Mixture of Experts with Visual Input
This research investigates applying sparsely gated mixture-of-experts (MoE) architectures to vision-based quadruped locomotion. By comparing MoE policies against standard MLPs under matched active parameter counts, the study demonstrates an 8% increase in training reward on a Unitree Go2 robot. Results suggest that borrowing scalability techniques from large language models can enhance network capacity for dynamic tasks like climbing and jumping.
The study compares sparsely gated mixture-of-experts (MoE) architectures against standard MLPs for vision-based quadruped locomotion, reporting an 8% reward increase during training on a Unitree Go2 robot when active parameters are matched.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2604.19344v2 Announce Type: replace Abstract: Agile locomotion is important for quadruped robots in the real world. Recent approaches have demonstrated impressive capabilities including dynamic climbing and jumping but typically rely on sequential multilayer perceptron (MLP) architectures with densely activated layers. In contrast, sparsely gated mixture-of-experts (MoE) architectures have become popular in the
Open source S2
sults show an 8\% increase in reward during training comparing the MoE to a standard MLP and even show benefits comparing to MLPs with equivalent total parameters. Experimental results on a real Unitree Go2 quadruped robot demonstrate performance in the real world. These results highlight that sparsely gated MoE architectures may pose advantages in tasks where network capacity is important and that borrowing techniqu
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org