Unified Robot Models

Unified robot models are Xu Huazhe’s preferred direction for the action and behavior parts of household robots in 166: 许华哲再次具身创业：不想错过最大的西瓜. The claim is that a robot cannot become general by solving isolated tasks one by one; it needs a model whose training lets skill, perception, action, and generalization compound.

Xu connects the route to reinforcement learning, post-training, data filtering, and the use of failure or suboptimal data. The goal is not to throw every trace into one dataset, but to let the robot learn from exploration and mistakes without narrowing itself into a fixed task policy.

Key Claims

• A collection of small models can work for fixed tasks, but may not produce general household intelligence.
• Failure data and mediocre data can be useful if the training system can judge quality and learn from them.
• Post-training should improve task success without destroying generalization.
• A unified model route creates a milestone problem for startups because early task results may look weaker than specialized systems.

Connections

• AI Native Robotics and Physical AGI — route and goal.
• Vision Language Action Models, World Models, and World Model VLA Fusion — adjacent model-route vocabulary.
• Real Robot Data Strategy and Household Robot Data Flywheel — data requirements.
• AI Commercialization Pressure — startup pressure created when long-term model training lacks immediate task-specific wins.
• Poke Robotics — company context.