Embodied AI

A Unified KV Cache for Multi-Task Inference on Robots

2026/09/01

OxyGen treats the KV cache as one shared resource across tasks and control frames, letting Mixture-of-Transformers VLAs keep full action frequency while generating language concurrently, up to 3.7× faster on an RTX 4090 and Jetson AGX Thor.

Deploying and Accelerating Cosmos 3 Robot Policy on RTX 4090s

2026/08/21

How quantization, sampler tuning, kernel work, and rollout topology accelerate NVIDIA's 16B Cosmos 3 robot policies on 24 GB GPUs — from single-request serving to full 1,200-episode evaluations.

Vision-Language-Action (VLA) Models: A Review of Recent Progress

2025/09/16

Recent VLAs are moving from discrete to continuous control and from single-system to dual-system designs.