Human-AI Shared Control Via Adaptive Neuron Editing in Reinforcement Learning Policy

While incorporating humans into the inference loop has been shown to significantly improve interpretability and learning efficiency, existing human-AI shared control algorithms suffer from either additional labor costs and retraining overhead or failure to achieve control objectives due to inaccurate explanations of neural representations. To achieve low-cost, highperformance human-AI shared control, this study introduces the Adaptive Neuron Editing (ANE) algorithm. ANE performs a fine-grained causal analysis of neural representations, adaptively adjusting editing values based on the current hidden states. We validate the effectiveness of ANE's neural representation analysis and evaluate its performance across various reinforcement learning tasks, including quadruped and bipedal robot locomotion control. The results demonstrate ANE's superiority over previous work on nearly all performance metrics. Furthermore, this work successfully enabled a wide range of complex motor skills in bipedal robots, which were then utilized to complete a highdifficulty parkour challenge.