Efficient VLAs, MoE Locomotion, and Parametric Humanoids

MoERLocomotion local-global mixture-of-experts for legged robot RL

This introduces a method for applying reinforcement learning to legged robots more efficiently by combining local and global Mixture-of-Experts (MoE). For builders, it implies you can deploy more robust locomotion policies without the massive compute overhead usually required.

Simple-VLA efficient vision-language-action models for robot agents

Introduces Simple-VLA as a framework for Vision-Language-Action models that prioritizes simplicity and efficiency. The significant reduction in training complexity makes building and fine-tuning multi-modal robot agents much more accessible for smaller teams.

Reactive navigation for dynamic environments via switching dynamical systems

This paper tackles persistent re-planning in dynamic environments by using a switching dynamical system with AUTOL. It's a critical step toward robust real-world navigation where obstacles aren't static.

Learning smooth and safe locomotion trajectory generation via RL

Focuses on achieving smooth and safe motion in Reinforcement Learning-based locomotion. The ability to generate smooth trajectories directly avoids the need for heavy post-processing in your control pipeline.

Proportional humanoids optimizing body shape for task performance

Presents a new method for adjusting humanoid body proportions to optimize task performance. This is essentially 'parametric design for tomorrow's androids', crucial for anyone iterating on custom hardware.