Learning Safety-Assured Collaborative Quadrupedal Manipulation

Collaboration is very common in animals, ranging from flocks of wolves to packs of bees. For instance, a team of ants can cooperatively carry heavy loads to their nest. Drawing inspiration from collaborative teamwork in nature, robots can also work together to achieve tasks that are beyond the ability of individual robots. However, this is challenging in several aspects. First, a high-level task and motion planning policy is required to plan and coordinate the motion of each robot, while taking feedback to estimate each agent's and the environment's state. Moreover, it is imperative to design a unified dynamic loco-manipulation control strategy to achieve manipulation and locomotion objectives simultaneously in the context of collaboration between legged robots. Finally, safety is of great importance in collaborative tasks, since a fault in one robot can lead to a cascade of problems to other robots. This project aims to address these challenges by developing a perceptive, collaborative and safe loco-manipulation framework. We propose a learning safety-assured framework for quadrupedal robots that collaboratively manipulate large objects. 


  • Bike Zhang, UC Berkeley
  • Zhongyu Li, UC Berkeley
  • Akshara Rai, Facebook AI Research
  • Koushil Sreenath, UC Berkeley