Generative Learning of World (GLOW)

KnowinDream

Gen­er­a­tive 3D World Data Engine

Visual Input

Robot Cam­eras (Images / Video)

Language

User Text Instruc­tion

Robot State & Execution Feedback

Pro­pri­o­cep­tion / Roll­out Traces

Gen­er­ates higher-quality embod­ied datadis­tri­bu­tions equipped with rich 3D struc­ture.

TRAININGDATA

KnowinGLOW

One Uni­fied Autore­gres­sive Model, Mul­ti­ple Capa­bil­i­ties.

BrainSkill

Spa­tial Under­stand­ing · Seman­tic Ground­ing

ActSkill

CoT-Guided Closed-LoopAction Gen­er­a­tor

Autore­gres­sive CoT-guided action gen­er­a­tion with closed-loop visual feed­back.

ACTIONCHUNK

KnowinWorld

Inter­ac­tive World Dynam­ics Sand­box

Action roll­out

Robot Execution

Action chunks are con­verted into smooth arm motion locally.

Embodied Alignment

World Dynamics World-grounded rollout
Active Exploration Skill composition
Interactive Reasoning Brain-Act consistency

KnowinAgent

Long-Horizon Run­time & Capa­bil­ity Evo­lu­tion

Stateful Memory Recovery & Replanning Playful Exploration Skill Consolidation
RLHF