Embodied AI has become a pivotal domain in modern robotics, driven by the growing need for robots to interact seamlessly with their environments and autonomously perform complex tasks. While significant progress has been made in single-agent systems, the next frontier lies in multi-agent collaboration and the ability to handle diverse manipulation tasks with greater precision and adaptability.
This challenge, hosted within the MEIS workshop at CVPR 2025, aims to push the boundaries of Embodied AI through robotic dual-arm manipulation tasks. Leveraging the RoboTwin simulation platform alongside the AgileX Cobot-Magic and Tack-Pika robotic platforms, the challenge is designed to tackle the complexities of robotic manipulation in both simulated and real-world environments. Participants will engage in Rigid Object Manipulation and Tactile Manipulation, each addressing unique aspects of robotic dexterity and control.
Looking ahead, the insights gained from this challenge will pave the way for more sophisticated multi-agent systems capable of handling real-world tasks with high levels of autonomy. The integration of advanced AI techniques—such as generative models and imitation learning will be key to overcoming current limitations and unlocking the full potential of robotic manipulation.
We invite researchers, engineers, and innovators worldwide to participate in this challenge and contribute to the advancement of Embodied AI in robotic manipulation. Together, we can push the boundaries of what’s possible and shape a future where intelligent, capable robots seamlessly support both industrial applications and everyday human activities.
🏆 Awards (Simulation & Real-world Track, Bonus only for Real-world Track): 🥇 First Prize ($1500) 🥈 Second Prize ($1000) 🥉 Third Prize ($500)
RoboTwin (ECCV 2024 Workshop Best Paper) :Webpage | Repo 2nd MEIS Workshop @CVPR2025 :Webpage Social Media :量子位
Poster
Competition Content
Task Overview
This track focuses on the precise manipulation of rigid objects, requiring participants to develop algorithms that enable robots to grasp, move, and interact with rigid objects in a controlled manner. These tasks will evaluate the robot's ability to plan and execute complex actions, with solutions validated through both simulation and real-world trials.
Evaluation
Expert data collected from real robots (Cobot-Magic), along with corresponding scenario data generated using the RoboTwin simulation platform, will be provided as combined expert datasets. Participants are required to train their algorithms solely using the provided data. The better the algorithm performs in real-world scenarios, the higher the score.
If the number of participants in the pre-registration phase exceeds expectations, we will consider conducting the first round of the competition on the RoboTwin simulation platform. The top-performing teams will advance to the final round, which will be held in real-world environments.
Important Update: Due to an overwhelming number of pre-registered teams and logistical constraints, we regret to announce that Track 2: Tactile Manipulation has been canceled. The competition will now focus solely on rigid object manipulation, but with a bonus on one tactile manipulation task.Detailed competition rules and procedures will be announced during the online briefing session.