The paper introduces a novel approach to inverse reinforcement learning (IRL) called *hybrid IRL*, which aims to reduce the computational burden and improve sample efficiency. Traditional IRL methods require solving a reinforcement learning (RL) problem repeatedly, leading to high interaction complexity and inefficiency. The proposed hybrid IRL approach combines online and expert data to train policies, focusing on states that are similar to the expert's behavior. This reduces the need for extensive exploration and computational overhead. The key contributions of the paper are: 1. **Reduction to Expert-Competitive RL**: The authors derive a reduction from inverse RL to expert-competitive RL, showing that as long as the policy search procedure guarantees average performance relative to the expert, it can achieve strong policy performance. 2. **Model-Free and Model-Based Algorithms**: Two hybrid IRL algorithms, HyPE (Hybrid Policy Emulation) and HyPER (Hybrid Policy Emulation with Resets), are proposed. HyPE uses the HyQ algorithm for hybrid RL, while HyPER uses the LAMPS algorithm. Both algorithms provide performance guarantees and are shown to be more sample-efficient than standard IRL methods on continuous control tasks. 3. **Empirical Validation**: Extensive experiments on the MuJoCo locomotion benchmark and the D4RL antmaze-large environment demonstrate that HyPE and HyPER achieve higher rewards with fewer interactions compared to other IRL methods. HyPER, in particular, shows superior performance without requiring resets to expert states, making it suitable for real-world applications. The paper highlights the benefits of hybrid IRL in reducing exploration and improving efficiency, making it a promising approach for imitation learning tasks.The paper introduces a novel approach to inverse reinforcement learning (IRL) called *hybrid IRL*, which aims to reduce the computational burden and improve sample efficiency. Traditional IRL methods require solving a reinforcement learning (RL) problem repeatedly, leading to high interaction complexity and inefficiency. The proposed hybrid IRL approach combines online and expert data to train policies, focusing on states that are similar to the expert's behavior. This reduces the need for extensive exploration and computational overhead. The key contributions of the paper are: 1. **Reduction to Expert-Competitive RL**: The authors derive a reduction from inverse RL to expert-competitive RL, showing that as long as the policy search procedure guarantees average performance relative to the expert, it can achieve strong policy performance. 2. **Model-Free and Model-Based Algorithms**: Two hybrid IRL algorithms, HyPE (Hybrid Policy Emulation) and HyPER (Hybrid Policy Emulation with Resets), are proposed. HyPE uses the HyQ algorithm for hybrid RL, while HyPER uses the LAMPS algorithm. Both algorithms provide performance guarantees and are shown to be more sample-efficient than standard IRL methods on continuous control tasks. 3. **Empirical Validation**: Extensive experiments on the MuJoCo locomotion benchmark and the D4RL antmaze-large environment demonstrate that HyPE and HyPER achieve higher rewards with fewer interactions compared to other IRL methods. HyPER, in particular, shows superior performance without requiring resets to expert states, making it suitable for real-world applications. The paper highlights the benefits of hybrid IRL in reducing exploration and improving efficiency, making it a promising approach for imitation learning tasks.