Particularly interesting how they utilize lidar odometry to estimate global position, maintaining consistency with the teleoperator. Their mixture of experts training system for humanoid RL is novel and represents a creative application I haven't encountered before in this domain.

Guided diffusion is a great way to extrapolate a lot of extra utility out of diffusion models. Their application of it to whole body reference motion tracking is, just as a whole, very creative. Diffusion works are also always compelling because they connect to flow matching, consistency models, DDIM, and diffusion steering improvements, allowing transfer of proven concepts from image/video generation to robotics.

Great reminder to question basic assumptions - more reasoning tokens aren't always helpful if they're wrong. Interesting how tiny normalization choices can completely change what you're actually optimizing for.

Clever approach using entropy during inference to guide training efficiency, allowing them to ignore 80% of tokens during gradient updates. Another entropy-centric paper explores using model intrinsics, but for encouraging exploration during learning.