Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning — ThinkLLM