Statler
State-maintaining language models for embodied reasoning.
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 theworld_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
andworld_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.