Statler

State-maintaining language models for embodied reasoning.

[project page] [paper]

With other members of RIPL at TTIC, we developed Statler based on the existing work of Code as Policies. The gist of Statler is a state-maintaining language model, where an explicit representation of the world state is maintained over time and intended for long-horizon operations. The paper is accepted to ICRA 2024.

  • Upon receiving the user’s query, Statler feeds it to the world_model_reader, which extends it with the current world state and uses it as the prompt to elicit code generation from a large language model (LLM). Specifically, GPT 3.5 is used.
  • The generated code not only enables the robot to perform the action in response, but also calls the world_model_writer with the necessary information to update the world state.
  • world_model_reader and world_model_writer are both instances of general-purpose GPT-driven LLMs separatedly tuned via few-shot prompting. We found that splitting the reasoning burden into different LLMs results in improved accuracy and stability for long-horizon operations.