TechPrufer
Get started
Physical Intelligence
Physical Intelligence logoPhysical Intelligence · Feature trail

Hierarchical Interactive Robot (Hi Robot)

1 post · Feb 2025

What this trail covers

The Hi Robot system introduces a hierarchical inference process for robots, combining a reactive 'System 1' VLA model (π0) with a deliberative 'System 2' VLM. This allows robots to break down complex tasks, process natural language prompts and feedback, and adapt their behavior more effectively, leading to improved instruction-following accuracy and task completion. This iteration enhances the system by training the high-level policy with synthetic data and demonstrating improved performance over GPT-4o and flat VLA policies in tasks like table bussing, sandwich making, and grocery shopping.

All Physical Intelligence trails
  1. Teaching Robots to Listen and Think HarderThis post introduces the 'Hi Robot' system, which employs a hierarchical inference process for robots. It combines a reactive 'System 1' VLA model (π0) with a deliberative 'System 2' VLM. The System 2 VLM acts as an 'inner voice,' breaking down complex tasks into intermediate steps for the System 1 VLA model. This approach allows for better processing of complex prompts and real-time contextual feedback, leading to improved instruction-following accuracy. The training of the high-level policy is enhanced by using synthetically labeled datasets that simulate realistic multi-step interactions. Evaluation shows Hi Robot outperforms GPT-4o and a flat VLA policy in various real-world tasks.