Generating Natural and Expressive Robot Gestures through Iterative Reinforcement Learning with Human Feedback using LLMs

A new system uses large language models and iterative reinforcement learning with human feedback to generate more natural and expressive gestures for the humanoid robot Pepper, according to research submitted in June 2026 [1]. The study integrates ChatGPT into Pepper to produce co-speech gestures aligned with conversational output [1]. While this baseline approach enables flexible gesture generation, the resulting motions are often perceived as stiff and unnatural [1]. To address this limitation, the researchers introduced an iterative reinforcement learning with human feedback (RLHF) system that finetunes gesture generation based on user evaluations [1]. The RLHF system improved the LLM's co-speech generative capabilities, producing movements described as more expressive, relevant, and fluid [1]. The findings were validated through an iterative user study comparing Pepper's generated gestures [1]. Expressive gestures are essential for natural and effective communication, complementing speech when verbal cues alone are insufficient, such as for pointing [2]. For social robots like Pepper, producing natural and expressive movements is critical for improving human-robot interaction and long-term acceptance [2]. Generating gestures has historically been challenging due to reliance on expert-authored animations, resulting in rigid behaviors impractical for dynamic environments [2]. Machine learning approaches have also struggled to capture perceived naturalness, a difficulty that increases with more degrees of freedom [2]. Recent advances in large language models have enabled dynamic code generation, offering new opportunities for runtime gesture synthesis from natural language [2]. Large language models, or LLMs, are a class of AI models trained on vast text corpora to understand and generate human-like text [3]. The field has seen rapid development, with companies such as DeepSeek demonstrating that models can be trained for significantly lower costs than previously assumed — its V3 model was reportedly trained for US$6 million, compared to the US$100 million cost for OpenAI's GPT-4 in 2023 [4]. The paper was submitted to arXiv on June 17, 2026, and is hosted under the Computer Science > Robotics category [1]. arXiv is operated by Cornell University and serves as a primary distribution platform for preprints in physics, mathematics, computer science, and related fields. The research was developed through arXivLabs, a framework that allows collaborators to develop and share new features directly on the arXiv website [1].