CooHOI: Learning Cooperative Human-Object Interaction with Manipulated Object Dynamics This paper introduces a novel framework, CooHOI, for learning cooperative human-object interaction tasks. The framework employs a two-phase learning approach: first, individual skill acquisition using the Adversarial Motion Priors (AMP) framework, followed by transfer to multi-agent collaboration using Multi-Agent Proximal Policy Optimization (MAPPO). The key idea is to use the dynamics of the manipulated object as feedback for implicit communication and coordination between agents. Unlike previous methods that rely on tracking-based approaches, CooHOI is inherently efficient and does not require motion capture data of multi-character interactions. It can be seamlessly extended to include more participants and a wide range of object types. The framework enables physically simulated characters to execute multi-agent human-object interaction (HOI) tasks with naturalness and precision. The method involves training a single-agent carrying policy using the AMP framework, followed by parallel training to develop cooperative strategies. The dynamics of the object are used as feedback information, allowing agents to adjust their actions based on changes in the object's dynamics. This approach facilitates implicit communication and coordination between agents, leading to effective cooperation in carrying tasks. Experiments demonstrate that the framework enables characters to exhibit natural-looking behaviors while successfully completing cooperative tasks using only motion capture data from a single agent. The results show that the framework is effective in handling various object categories and can be extended to different types of objects and numbers of agents. The framework also provides insights into the capabilities of policy models and the effectiveness of using object dynamics for communication in learning cooperative object transporting tasks.CooHOI: Learning Cooperative Human-Object Interaction with Manipulated Object Dynamics This paper introduces a novel framework, CooHOI, for learning cooperative human-object interaction tasks. The framework employs a two-phase learning approach: first, individual skill acquisition using the Adversarial Motion Priors (AMP) framework, followed by transfer to multi-agent collaboration using Multi-Agent Proximal Policy Optimization (MAPPO). The key idea is to use the dynamics of the manipulated object as feedback for implicit communication and coordination between agents. Unlike previous methods that rely on tracking-based approaches, CooHOI is inherently efficient and does not require motion capture data of multi-character interactions. It can be seamlessly extended to include more participants and a wide range of object types. The framework enables physically simulated characters to execute multi-agent human-object interaction (HOI) tasks with naturalness and precision. The method involves training a single-agent carrying policy using the AMP framework, followed by parallel training to develop cooperative strategies. The dynamics of the object are used as feedback information, allowing agents to adjust their actions based on changes in the object's dynamics. This approach facilitates implicit communication and coordination between agents, leading to effective cooperation in carrying tasks. Experiments demonstrate that the framework enables characters to exhibit natural-looking behaviors while successfully completing cooperative tasks using only motion capture data from a single agent. The results show that the framework is effective in handling various object categories and can be extended to different types of objects and numbers of agents. The framework also provides insights into the capabilities of policy models and the effectiveness of using object dynamics for communication in learning cooperative object transporting tasks.