Dyna announced Dyna-2, a world-action model pre-trained on one million hours of human video, which revealed new scaling laws in human data. This work has led to the development of the physical robot agent, Dyna-2.1, capable of executing long-horizon tasks with autonomy. The discussion focuses on reliable, long-horizon autonomy as the main challenge for real-world robotics. Dyna's approach involves world-action models, predicting future video to guide robot actions, and leveraging synthetic data for training.
Perseverance surpassed Opportunity's distance record on Mars in just five years. This achievement occurred when Perseverance's drive reached 299.6 feet. The record was set on October 11, 2026. This event was captured by one of Perseverance's navigation cameras. The milestone involved exceeding Opportunity's previous record of 28.06 miles.
Gemini 4 Argon demonstrates a significant leap in physical understanding, code generation, and visual-semantic reasoning, scoring 87% on the Embodied Reasoning (ERQA) benchmark. Argon can be applied across the robotics lifecycle, enabling tasks like robot control, real-to-simulation reconstruction, and physical understanding. It facilitates complex tasks such as mobile manipulation, zero-shot control, and steering video world models using iterative visual feedback. Argon unlocks real-to-sim by building calibrated MuJoCo twins from real robot feeds and using visual reasoning to verify physical contact forces. Furthermore, Argon can generate new simulation data to improve real-world robot performance by synthesizing thousands of domain-randomized episodes.
Cooperative autonomous driving uses natural language communication to enable vehicles to share observational data. The UNCAP framework allows vehicles to determine who to communicate with and how much to trust the received information. It achieves this by having vehicles broadcast minimal location data, selectively exchange natural language descriptions with calibrated confidence, and reason about mutual information to assess relevance. Experiments showed UNCAP improved efficiency, safety, and reduced decision uncertainty compared to other methods. Ultimately, this approach suggests that structured language communication can be a powerful foundation for future cooperative AI systems.