🚨 New Paper Alert – ICRA 2025 🚨 Language-Conditioned Offline RL for Multi-Robot Navigation Ever wish you could just tell a team of robots what to do? Like: 🗣 “Robot 1, head to the southwest corner.” 🗣 “Robot 3, the north edge is your target.” Now you can. In our latest work presented at ICRA this week by Ajay Shankar, we show how natural language can be used to guide a team of 5 robots—but with a twist: 🔁 Instead of using large language models (LLMs) to directly control robots (a growing trend), we use them to translate human language into a compact, consistent representation—a distilled “latent vector” that captures the essence of the command. 🤖 Then, our robots—trained offline using reinforcement learning—know how to act based on that vector. Say it 10 different ways? Same action.
✨ Why this matters: It generalizes across language variety It separates understanding from acting It works across a multi-robot system with flexible, high-level language commands (e.g. “go to the north edge”, “head to the top left”)
#ICRA2025 #Robotics #OfflineRL #NaturalLanguage #MultiRobot #LanguageToAction #HumanRobotInteraction