# Legs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion

**Authors:** Xuxin Cheng, Ashish Kumar, Deepak Pathak  
**arXiv:** [2303.11330](https://arxiv.org/abs/2303.11330)  
**Date:** 2023-03-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.11330v2/

## Key points

- The paper explores the limitations of robotic quadrupeds in comparison to their biological counterparts
- Dogs and other animals display a variety of agile skills beyond locomotion that involve using their legs to interact with objects and climb walls
- The authors propose training quadruped robots not only to walk but also to use their front legs for manipulation tasks in the real world
- Skill learning is decoupled into two categories - locomotion and manipulation
- The authors train these skills in simulation using curriculum learning and transfer them to the real world using their proposed sim2real variant
- They combine these skills into a robust long-term plan by learning a behavior tree from an expert demonstration which encodes a high-level task hierarchy
- The method is evaluated in both simulation and real-world scenarios showing successful executions of short as well as long-range tasks
- The approach could have significant implications for various fields where such robots are used including search and rescue operations or hazardous material handling tasks.

## Summary

The paper titled "Legs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion" by Xuxin Cheng, Ashish Kumar, and Deepak Pathak explores the limitations of robotic quadrupeds in comparison to their biological counterparts. While locomotion has seen significant progress for walking or running across challenging terrains, dogs and other animals display a variety of agile skills beyond locomotion that involve using their legs to interact with objects and climb walls. To bridge this gap, the authors propose training quadruped robots not only to walk but also to use their front legs for manipulation tasks in the real world. To handle this challenging optimization problem, they decouple skill learning into two categories - locomotion and manipulation. Locomotion includes any movement such as walking or climbing a wall while manipulation involves using one leg to interact while balancing on the other three legs. The authors train these skills in simulation using curriculum learning and transfer them to the real world using their proposed sim2real variant that builds upon recent success in locomotion research. Finally, they combine these skills into a robust long-term plan by learning a behavior tree from an expert demonstration which encodes a high-level task hierarchy. The method is evaluated in both simulation and real-world scenarios showing successful executions of short as well as long-range tasks. The authors demonstrate how robustness helps confront external perturbations. Videos showcasing the results are available at https://robot-skills.github.io/. Overall, this paper presents an innovative approach towards pushing quadrupedal agility beyond just locomotion by incorporating manipulation skills into robotic quadrupeds' capabilities which could have significant implications for various fields where such robots are used including search and rescue operations or hazardous material handling tasks.