Legs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion

Abstract: Locomotion has seen dramatic progress for walking or running across challenging terrains. However, robotic quadrupeds are still far behind their biological counterparts, such as dogs, which display a variety of agile skills and can use the legs beyond locomotion to perform several basic manipulation tasks like interacting with objects and climbing. In this paper, we take a step towards bridging this gap by training quadruped robots not only to walk but also to use the front legs to climb walls, press buttons, and perform object interaction in the real world. To handle this challenging optimization, we decouple the skill learning broadly into locomotion, which involves anything that involves movement whether via walking or climbing a wall, and manipulation, which involves using one leg to interact while balancing on the other three legs. These skills are trained in simulation using curriculum and transferred to the real world using our proposed sim2real variant that builds upon recent locomotion success. Finally, we combine these skills into a robust long-term plan by learning a behavior tree that encodes a high-level task hierarchy from one clean expert demonstration. We evaluate our method in both simulation and real-world showing successful executions of both short as well as long-range tasks and how robustness helps confront external perturbations. Videos at https://robot-skills.github.io

• The paper presents a novel decoupling of locomotion and manipulation, significantly boosting quadrupedal agility through targeted skill training.
• It introduces a reinforcement learning curriculum and sim2real adaptation that enable effective climbing, button pressing, and object interaction.
• The use of a behavior tree framework with robust online adaptation yields reliable long-horizon task planning in dynamic real-world scenarios.

Leveraging Quadrupedal Agility for Enhanced Functional Capabilities in Robotic Systems

The paper presented addresses a notable gap in robotic quadrupeds, emphasizing the distinction between effective locomotion and the diverse manipulation skills exhibited by biological counterparts. The research centers on significantly advancing the capabilities of quadrupeds, not only in traversing complex terrains but also in executing basic manipulation tasks. These tasks include climbing, button pressing, and object interaction, thereby redefining the quadrupedal robots' operational repertoire.

The paper employs a novel approach by decoupling skill learning into two distinct categories: locomotion and manipulation. Locomotion encompasses movements such as walking and climbing, while manipulation focuses on using one leg for interaction while maintaining balance on the other three. This separation is critical, as it simplifies the challenging optimization problem into more manageable components. The skills are initially trained in simulated environments using reinforcement learning, leveraging a curriculum-based approach to enhance learning efficiency. Subsequently, these skills are transferred to real-world scenarios through their proposed sim2real variant, building on recent advances in sim2real adaptation techniques.

Remarkably, the research extends beyond isolated skill execution to encompass long-term task planning. The authors accomplish this by utilizing a behavior tree framework to encode a high-level task hierarchy. This framework learns from a single expert demonstration, enabling the robot to execute robust, long-horizon tasks. The ability to handle perturbations and interruptions during task execution highlights the robustness of their approach.

Quantitative evaluations in both simulation and real-world settings underscore the effectiveness of the proposed methods. The use of a Unified State Estimator (USE) and Regularized Online Adaptation (ROA) proves instrumental in achieving superior performance compared to existing baselines. Importantly, the terrain curriculum is shown to enhance policy learning, enabling the quadrupeds to handle more challenging environments, such as vertical walls.

Furthermore, the researchers demonstrate the practical implications of their work through extensive real-world experiments. These experiments include tasks such as climbing and pressing buttons at varying heights, showcasing the potential applicability of their methods in human environments. This expanded range of capabilities for quadrupeds could significantly enhance their effectiveness in contexts requiring both mobility and interaction with objects, such as search and rescue missions or assistance in daily human tasks.

The implications of this work are manifold. Theoretically, it bridges a crucial gap between locomotion and manipulation, offering a framework for future research in integrated skill learning for robotic systems. Practically, the enhanced capabilities and robustness achieved could be instrumental in accelerating the deployment of quadrupedal robots in dynamic and complex environments.

Looking toward the future, further integration of high-level decision-making and low-level command tracking in an end-to-end framework could be a promising direction. Such advancements could lead to even more autonomous robotic systems capable of adapting to and interacting with their environments with minimal human intervention.

In conclusion, this research demonstrates a significant step forward in the development of versatile and capable quadrupedal robots. By equipping these robots with enhanced manipulation skills alongside traditional locomotion, the authors pave the way for more dynamic and useful robotic applications, both in research and real-world scenarios.