AgiBot Embodied Intelligence Research Center has released the second generation of its integrated embodied “large-and-small brain” system, named GenieReasoner, representing a major upgrade to its embodied AI architecture.

To address the long-standing modality alignment challenge between semantic reasoning and action control in Vision-Language-Action (VLA) models, AgiBot proposed a new model architecture that supports unified discretized pre-training across perception, cognition and control. This enables tighter integration between high-level reasoning and low-level motor execution within a single end-to-end framework.

In addition, AgiBot introduced a flow-matching mechanism to mitigate the action precision bottleneck of traditional discrete tokenizers, improving the continuity and accuracy of action generation while preserving the efficiency advantages of discrete representations.

The company said GenieReasoner is designed to serve as a core cognitive engine for embodied agents, allowing robots to better understand complex environments, reason about tasks, and execute actions in a more coherent and reliable manner.