Robix

ByteDance Seed's high-level robot 'brain' model that unifies conversation, reasoning, and task planning.

Robix was released by the ByteDance Seed team in September 2025, in 7B and 32B parameter sizes. It serves as the high-level cognitive layer in a hierarchical robot system: given the camera feed and what the user says, a single vision-language model handles reasoning, long-horizon task planning, and natural-language interaction all at once, producing both atomic instructions for a lower-level controller and things to say back to the user; the lower level can be ByteDance's own GR-3 VLA model. Robix can proactively ask clarifying questions about ambiguous instructions, revise its plan in real time when interrupted mid-execution, and apply commonsense reasoning, such as filtering food choices by dietary restrictions. Training has three stages: continued pretraining to strengthen embodied-reasoning abilities such as spatial understanding, supervised fine-tuning that unifies interaction and planning into reasoning-action sequences, and reinforcement learning to improve consistency between reasoning and action. The authors report it outperforms commercial large-model baselines on real tasks such as clearing a dinner table and grocery shopping.

ExampleA user asks the robot to clear the table, then mid-task adds, 'leave the cup for now' — Robix immediately revises its plan, confirms verbally, and sends the lower-level VLA a new atomic instruction.

Also called

Robix-7B, Robix-32B, Robix: A Unified Model for Robot Interaction, Reasoning and Planning

Related

Seed GR-3 · Hierarchical Architecture · Dual-System Architecture (System 1 / System 2) · Embodied Reasoning · LLM-based Task Planning · ByteDance Seed

Sources

Robix (arXiv 2509.01106)
Robix project page

As of

2025-09