Robotics communication infrastructure. One binary, predictable latency.

I was getting into robotics programming for the first time, I tried to learn ROS 2, I found it very boring and complex, with a lot of abstractions, so I thought I can make a much simpler version that is performant, reliable, and lightweight, so I'm building glu.

• it must be just a robotics communication middleware, not an OS.
• it must be one lightweight binary that can be used in compute constrained systems, with no abstraction overhead and massive layers to setup.
• it must be very easy and boring to use, if you have to think about communication while using it instead on working on the robot logic and brain then glu had failed.
• it must scale well for real-world massive projects.

Multimodal fusion layers with linear-scaling attention for vision-language models.

Cross-modal attention gets expensive quadratically as you fuse longer sequences. GRF is an attempt at fusion layers that scale linearly without gutting downstream quality — explored through a pre-print on multimodal fusion in transformer architectures.

• a not so bad trial to make multimodal models more compute efficient.