Today for AI
HOT RADAR
agentHEAT 10.5°

AI CLUSTERED EVENT · 10/9/2026

MirroS Releases AgentGarten: Combining Code-Defined Physics with Diffusion Rendering for Embodied Agent Evolution

1 reports archived1 independent sourcesupdated 10/9/2026, 10:03:48
Synthesis & Latest Updates
1 Sources Cross-Validated

The MirroS team has released and open-sourced AgentGarten, an embodied AI training framework that connects executable code environments with real-time neural renderers. By defining precise physical causal rules via code and generating realistic visual feedback at over 30 fps using diffusion models, it resolves the conflict between the crude visuals of traditional engines and the uncontrollable physics of video-based world models. In hide-and-seek experiments, agents utilized an 'experiment manual' mechanism to record trial-and-error experiences, successfully emerging complex strategies like moving barriers and building ramps, validating the feasibility of Physical Recursive Self-Improvement (RSI).

LATEST/The MirroS team has released and open-sourced AgentGarten, an embodied AI training framework that connects executable code environments with real-time neural renderers. By defining precise physical causal rules via code and generating realistic visual feedback at over 30 fps using diffusion models, it resolves the conflict between the crude visuals of traditional engines and the uncontrollable physics of video-based world models. In hide-and-seek experiments, agents utilized an 'experiment manual' mechanism to record trial-and-error experiences, successfully emerging complex strategies like moving barriers and building ramps, validating the feasibility of Physical Recursive Self-Improvement (RSI).

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. 量子位官网T2·78 pts

    MirroS Releases AgentGarten: Combining Code-Defined Physics with Diffusion Rendering for Embodied Agent Evolution

    Original: 代码造世界，扩散绘现实：AgentGarten让智能体在实时试炼场中边玩边进化

    • AgentGarten employs a dual-track architecture: the code layer ensures deterministic and queryable physical rules, while the diffusion rendering layer provides high-fidelity, real-time visual feedback.
    • It introduces an 'experiment manual' mechanism where agents autonomously review and record findings after each interaction round, forming a continuous experience accumulation loop.
    • In hide-and-seek tasks, agents spontaneously evolved advanced game strategies such as constructing cover and adjusting tool positions without pre-set scripts.