A new algorithm called BrainBody-LLM, developed by researchers at NYU Tandon School of Engineering, enables robots to plan and execute tasks in a way that closely mimics human-like reasoning and movement. The system uses two large language models (LLMs) working together—the Brain LLM and the Body LLM—to break down complex tasks into actionable steps and refine movements in real time. The Brain LLM acts as the planner, interpreting high-level user instructions such as "Eat chips on the sofa" and decomposing them into logical, sequential steps using real-world knowledge. The Body LLM then translates each step into executable robot commands, controlling physical movements with precision. If a required action cannot be performed due to environmental constraints, the Body LLM signals this by outputting a special token, allowing the system to adapt dynamically. A key innovation of the BrainBody-LLM framework is its closed-loop feedback mechanism. This allows the system to continuously monitor the robot’s actions and environmental responses, incorporating error signals and context cues to correct and refine its plans on the fly. This feedback loop enhances robustness, especially when dealing with unpredictable or complex real-world scenarios. The researchers tested the algorithm in both simulated and real-world environments. In simulations using the VirtualHome platform, the model demonstrated improved task completion rates. In real-world trials, a Franka Research 3 robotic arm successfully carried out a variety of household tasks, achieving an average success rate of 84%—a 17% improvement over existing state-of-the-art models. The team believes the algorithm’s strength lies in its ability to combine high-level reasoning with low-level control, enabling robots to handle tasks that require both understanding and physical execution. Unlike many current LLM-based systems that operate in isolation, BrainBody-LLM integrates feedback from the physical world, making it more reliable and adaptive. Looking ahead, the researchers aim to expand the system by incorporating additional sensory inputs such as 3D vision, depth sensing, and joint-level control. These enhancements could further improve the fluidity and realism of robotic movements, bringing them closer to human capabilities. The work represents a significant step forward in using LLMs not just for language understanding, but as core components of intelligent robotic systems. The approach may inspire other researchers to develop similar frameworks, accelerating progress in AI-driven robotics.