Bimanual Long-Horizon Lifecare Robotics with Temporal Context LLM Planner and Transformer Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Lifecare tasks often require complex bimanual manipulation.
• Existing robotic systems struggle with long-horizon tasks without extensive human data.
• Manual reward function design is time-consuming and limits scalability.

Purpose of the Study:

• To present a novel robot intelligence, TCP-TRL, for learning complex bimanual lifecare tasks.
• To eliminate the need for long-horizon human demonstrations and manually designed reward functions.
• To enhance robot adaptability and performance in healthcare settings.

Main Methods:

• TCP-TRL integrates a Large Language Model (LLM) with Retrieval-Augmented Generation (RAG) for high-level planning (TCP).
• TCP translates text into temporal action contexts, guiding a Transformer-Based Reinforcement Learning (TRL) framework.
• A compact dataset of primitive actions and task-specific info guides LLM planning, generating structured plans and sparse rewards for TRL training.

Main Results:

• TCP-TRL achieved an 81.86% average success rate on four lifecare tasks.
• Performance matched state-of-the-art TRL models trained with human demonstrations and designed rewards.
• The system demonstrated robustness and preserved temporal dependencies during training.

Conclusions:

• TCP-TRL offers a scalable, human intervention-free solution for bimanual long-horizon manipulation.
• It overcomes LLM hallucinations and reduces dependence on large datasets.
• This approach is relevant for dual-arm robots in hospital and homecare environments.