Complete framework for agile quadruped locomotion: integrating real-time control, planning, and perception in multi-contact environments

Abstract

Real-time synthesis of legged locomotion manoeuvres in challenging environments, such as industrial staircases with distinct contact surfaces, remains an unresolved problem. These environments, known as multi-contact environments, require the simultaneous determination of footstep locations several steps ahead while generating whole-body motions close to the robot's operational limits. The practical constraint of state estimation and perception errors necessitates rapid re-planning of motions. With traditional model-based methods, this constraint prevents using a single large optimisation to solve the entire locomotion problem, leading to problem decomposition into smaller, more manageable sub-problems to meet computing time requirements.

However, such decomposition introduces issues such as loose coupling between sub-problems due to varying models, constraints, and horizons, highlighting the need for an efficient architecture. Decomposition, particularly at the contact level to manage complexity, exacerbates the challenge of addressing combinatorial problems in multi-contact environments.

We propose a fully decomposed control architecture to address the locomotion problem, leveraging mixed-integer optimisation with conservative assumptions as a planner to select only the contact surfaces while continuously updating footstep positions and leg trajectories within these surfaces. The approach was initially validated on the quadruped robot Solo and later extended to include real-time perception with the larger industrial robot ANYmal-B. Additionally, we explore integrating learning-based methods along with trajectory optimisation frameworks to leverage the strengths of both approaches and enhance the planner's performance.