Human–robot collaboration (HRC) is a production model where humans and robots collaborate side by side in shared workspaces, particularly within complex environments. It plays a crucial role in the context of Industry 5.0, emphasizing sustainability with a focus on human well-being. The environment in HRC is dynamic and the safety of human is a critical concern of HRC. Currently, there are standards such as ISO/TS 15066 that put strict constraints on the robot speed and guarantee the safety of human but also restrict the efficiency of HRC. Moreover, robot learning is emerging and promising for improving the intelligence of robot in HRC and improving the safety of human. Therefore, in this chapter, safe HRC using multimodal perception and robot learning is proposed. First, a multimodal perception architecture for HRC is introduced, based on which, human–robot collision detection is realized with a bounding box and a modified and dynamic safe distance threshold is used to assess the safety risk of HRC without sacrificing the efficiency. Second, compared with the existing safety control strategy, a dynamic risk index which changes dynamically according to the distance between human and robot and the velocity of human and robot is constructed. A gradient risk index minimization method based on a dynamic risk index is used to modify the trajectory of robot, so as to avoid collision with human without reducing the efficiency of HRC significantly. Third, in terms of safe HRC control based on deep reinforcement learning, a Markov decision process for safe HRC is constructed, a reward function design method combining extrinsic reward function with intrinsic rewards function is constructed, an IRDDPG algorithm that combines the intrinsic reward function with the Deep Deterministic Policy Gradient is adopted to solve the multi-objective optimization problem of safety and task efficiency. The experiments verify the effectiveness of the proposed methods.