The field of legged robot locomotion is rapidly advancing, with a focus on developing more agile, efficient, and adaptable systems. Recent research has explored the use of reinforcement learning and model predictive control to improve the stability and robustness of legged robots in various environments. Notable developments include the integration of whole-body control and manipulation capabilities, as well as the use of risk-aware reinforcement learning to adapt to unknown conditions. Additionally, researchers have investigated the importance of architectural design in achieving robust humanoid locomotion, highlighting the benefits of layered control architectures. Overall, these advancements are paving the way for the development of more sophisticated and reliable legged robots for a range of applications. Noteworthy papers include: Towards Dynamic Quadrupedal Gaits, which presents a unified reinforcement learning framework for generating versatile quadrupedal gaits, and Residual MPC, which blends reinforcement learning with model predictive control to achieve highly robust behaviors. Architecture Is All You Need also demonstrates the importance of architectural design in achieving robust humanoid locomotion.
Advancements in Legged Robot Locomotion
Sources
ATRos: Learning Energy-Efficient Agile Locomotion for Wheeled-legged Robots
Towards Safe Maneuvering of Double-Ackermann-Steering Robots with a Soft Actor-Critic Framework
Towards Dynamic Quadrupedal Gaits: A Symmetry-Guided RL Hierarchy Enables Free Gait Transitions at Varying Speeds
Data-driven simulator of multi-animal behavior with unknown dynamics via offline and online reinforcement learning
Gain Tuning Is Not What You Need: Reward Gain Adaptation for Constrained Locomotion Learning
Preference-Conditioned Multi-Objective RL for Integrated Command Tracking and Force Compliance in Humanoid Locomotion
NaviGait: Navigating Dynamically Feasible Gait Libraries using Deep Reinforcement Learning
Autonomous Legged Mobile Manipulation for Lunar Surface Operations via Constrained Reinforcement Learning
Residual MPC: Blending Reinforcement Learning with GPU-Parallelized Model Predictive Control
Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations
Risk-Aware Reinforcement Learning with Bandit-Based Adaptation for Quadrupedal Locomotion
Generative Models From and For Sampling-Based MPC: A Bootstrapped Approach For Adaptive Contact-Rich Manipulation
Architecture Is All You Need: Diversity-Enabled Sweet Spots for Robust Humanoid Locomotion