AgentGarten
A research organization focused on reinforcement learning and visual agent research. It proposes a shared neural renderer method to improve the learning efficiency of agents.
Timeline (last 90 days)
This method requires that a world's dynamics be expressed as structured conditions and exported to the renderer, so its scope of application is limited to code-defined worlds.
The same pipeline was reused in four other worlds, with improvements in all of them.
The Adversarial Forcing distillation step may not generalize to unseen physics or textures.
The "millions versus 4 rounds" comparison is built on traditional RL baselines; how much the efficiency gap would shrink if a stronger pixel-level model were used as the control currently has no answer.
AgentGarten proposes using a single shared neural renderer to provide real-time visual observations for multiple code-defined worlds, and the agent converges within 4 rounds under the condition of obtaining information only through rendered observations.
The paper reports that in hide-and-seek experiments the agent learned to build panel shelters by round 4 and learned to build ramps to climb over walls by round 10.
The shared neural renderer turns rendering from a graphics pipeline hand-written separately for each world into a unified neural rendering service, thereby reducing the visual learning burden for each new world.
Each round's experience is distilled into an inheritable "playbook", on which subsequent agents continue to improve.
Matrix-Game 4.0 requires only a single consumer-grade GPU, whereas Code World Model requires 1× H200 141GB, StreamRender-H3 requires 8× GB200, and AgentGarten requires 1× H100 datacenter card.
Matrix-Game 4.0 is more open-sourced than Code World Model, StreamRender-H3 (already open) and AgentGarten (partially open).
1 more timeline events
All Facts (11)
This method requires that a world's dynamics be expressed as structured conditions and exported to the renderer, so its scope of application is limited to code-defined worlds.
50%UnverifiedThe same pipeline was reused in four other worlds, with improvements in all of them.
50%UnverifiedThe Adversarial Forcing distillation step may not generalize to unseen physics or textures.
50%UnverifiedThe "millions versus 4 rounds" comparison is built on traditional RL baselines; how much the efficiency gap would shrink if a stronger pixel-level model were used as the control currently has no answer.
50%UnverifiedAgentGarten proposes using a single shared neural renderer to provide real-time visual observations for multiple code-defined worlds, and the agent converges within 4 rounds under the condition of obtaining information only through rendered observations.
50%UnverifiedThe paper reports that in hide-and-seek experiments the agent learned to build panel shelters by round 4 and learned to build ramps to climb over walls by round 10.
50%UnverifiedThe shared neural renderer turns rendering from a graphics pipeline hand-written separately for each world into a unified neural rendering service, thereby reducing the visual learning burden for each new world.
50%UnverifiedEach round's experience is distilled into an inheritable "playbook", on which subsequent agents continue to improve.
50%UnverifiedMatrix-Game 4.0 requires only a single consumer-grade GPU, whereas Code World Model requires 1× H200 141GB, StreamRender-H3 requires 8× GB200, and AgentGarten requires 1× H100 datacenter card.
50%UnverifiedMatrix-Game 4.0 is more open-sourced than Code World Model, StreamRender-H3 (already open) and AgentGarten (partially open).
50%UnverifiedMatrix-Game 4.0 has an end-to-end latency of 112 milliseconds, whereas Code World Model is about 10.2 seconds, StreamRender-H3 about 545 milliseconds, and AgentGarten about 878 milliseconds.
50%