Efficient Locomotion Control in Legged Robots

The field of legged robotics is moving towards more efficient and robust locomotion control methods. Researchers are exploring innovative approaches to improve sample efficiency, reduce energy expenditure, and enhance overall performance. One notable direction is the integration of model-based reinforcement learning with physics-grounded energy models, which has shown promising results in sim-to-real transfer and energetic efficiency. Another area of focus is the development of hierarchical control frameworks that incorporate reduced-order models and nonlinear model predictive control to achieve versatile step planning and robust locomotion. Noteworthy papers in this area include:

  • A paper proposing a hyperparameter-free gradient optimization method to minimize energy expenditure without conflicting with task performance, demonstrating a 64% reduction in energy usage while maintaining comparable task performance.
  • A paper presenting a computationally efficient hierarchical control framework for humanoid robot locomotion, enabling robust locomotion across diverse environments and improving push recovery success rate by 36%.
  • A paper introducing a framework that integrates sim-to-real reinforcement learning with a physics-grounded energy model, achieving a 32 percent reduction in the full Cost of Transport of ANYmal.

Sources

First Order Model-Based RL through Decoupled Backpropagation

A Layered Control Perspective on Legged Locomotion: Embedding Reduced Order Models via Hybrid Zero Dynamics

Non-conflicting Energy Minimization in Reinforcement Learning based Robot Control

Tangential Action Spaces: Geometry, Memory and Cost in Holonomic and Nonholonomic Agents

Hierarchical Reduced-Order Model Predictive Control for Robust Locomotion on Humanoid Robots

Learning to Walk with Less: a Dyna-Style Approach to Quadrupedal Locomotion

Towards bridging the gap: Systematic sim-to-real transfer for diverse legged robots

RT-HCP: Dealing with Inference Delays and Sample Efficiency to Learn Directly on Robotic Platforms

Built with on top of