MirroS has developed Code-as-World, a system that transforms real-world videos into executable MuJoCo physics simulations. This system employs an agentic loop to create and verify these simulated environments, which are then used to train physical reasoning capabilities. The approach achieved a score of 55.4 on the QuantiPhy validation, surpassing Gemini-3.1 Flash, and its checkpoints are publicly available. AI
IMPACT Enables more realistic training of AI agents in simulated physical environments, potentially accelerating progress in robotics and embodied AI.
RANK_REASON The item describes a new system for converting videos into physics simulations and its performance on a validation benchmark. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Mastodon — mastodon.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →