UniVLA presents a novel approach to generalist policy planning using an embodiment-agnostic action space, achieving state-of-the-art results across various benchmarks with efficient training. It includes a comprehensive methodology for extracting latent actions from cross-embodiment videos and guidance on pre-training and fine-tuning models for real-world robot tasks.

Google DeepMind is advancing robotics by enabling robots to learn and improve autonomously through competitive play, using table tennis as a testbed. By having robots play against each other and incorporating vision language models for coaching, they aim to overcome the limitations of traditional programming and machine learning approaches that require extensive human input. This research seeks to create machines capable of continuous self-improvement and skill acquisition in dynamic environments.

Researchers have developed the Video Joint Embedding Predictive Architecture (V-JEPA), an AI model that learns about its environment through videos and exhibits a sense of "surprise" when presented with contradictory information. Unlike traditional pixel-space models, V-JEPA uses higher-level abstractions to focus on essential details, enabling it to understand concepts like object permanence with high accuracy. The model has potential applications in robotics and is being further refined to enhance its capabilities.

Vision-language-action (VLA) models enhance robotic manipulation by integrating action generation with vision-language capabilities. This paper reviews post-training strategies for VLA models, drawing parallels with human motor learning to improve interaction with environments. It introduces a taxonomy focusing on environmental perception, embodiment awareness, task comprehension, and multi-component integration, while identifying key challenges and trends for future research.