anonymous-report-421/GPT-as-Policy

What it solves

This project evaluates the effectiveness of using a large language model (GPT-6 Astra) as a control policy for robots, specifically testing whether the model can directly control a robot or act as a high-level reviewer to correct the actions of another policy (like $\pi_{0.5}$).

How it works

The project implements two primary evaluation methods:

1. GPT-6 Astra Direct: The model is used as the primary controller to execute tasks.
2. Hybrid Policy: A base policy ($\pi_{0.5}$) proposes actions, and GPT-6 Astra reviews these actions, optionally correcting them before they are executed.

These policies are tested across ten RoboDojo tasks. The system integrates with simulators (Isaac Sim 5.1) and policy servers to execute rollouts and record success rates and scores.

Who it’s for

Researchers and developers in embodied AI and robotics who are studying how frontier LLMs can be used for robotic control and action correction.

Highlights

• Hybrid Correction Approach: Demonstrates that a hybrid policy (base policy + GPT correction) outperforms direct LLM control.
• Performance Metrics: Achieves a 48% success rate in hybrid mode compared to 26% in direct mode on RoboDojo tasks.
• Integration: Provides a framework for connecting LLMs to the RoboDojo and RoboLab environments.
• Reasoning Capabilities: Utilizes the xhigh reasoning setting of GPT-6 Astra for policy execution.