Robix is an integrated model that integrates robotic reasoning, task planning, and natural language interaction into a single vision-language architecture. Acting as a high-level cognitive layer in a hierarchical robotic system, Robix dynamically generates atomic commands for low-level controllers and verbal responses for human interaction, enabling robots to follow complex instructions, plan long-term tasks, and interact naturally with humans within an end-to-end framework. Robix introduces new capabilities such as proactive conversation during task execution, real-time interruption handling, and context-aware common-sense reasoning. At its core, Robix leverages thought-chain reasoning and employs a three-stage training strategy: (1) continuous pretraining to enhance basic implementation reasoning abilities, including 3D spatial understanding, visual-based, and task-oriented reasoning; (2) supervised fine-tuning to model human-robot interaction and task planning as integrated reasoning-action sequences; and (3) reinforcement learning to improve reasoning-action consistency and long-term task consistency. Extensive experiments show that Robix outperforms open-source and commercial benchmarks (e.g., GPT-4o and Gemini 2.5 Pro) in executing interactive tasks, demonstrating strong generalization across a variety of instruction types (e.g., open, multi-step, constrained, null, and interrupted) and across a variety of user-related tasks, such as table cleaning, grocery shopping, and diet filtering.
