Report on Current Developments in Legged Robotics

General Direction of the Field

The field of legged robotics is witnessing a significant shift towards more sophisticated and integrated control strategies, aimed at enhancing the performance, adaptability, and robustness of legged systems in complex environments. Recent advancements are characterized by a blend of theoretical innovations and practical implementations, pushing the boundaries of what legged robots can achieve.

One of the key trends is the expansion of control methodologies to incorporate both linear and angular dynamics, enabling more precise and versatile motion planning. This is particularly evident in the development of frameworks that integrate angular components into traditional linear control models, such as the Divergent Component of Motion (DCM). These advancements allow for more nuanced control over the rotational dynamics of legged robots, which is crucial for tasks requiring agile maneuvers and precise posture regulation.

Another notable direction is the adoption of hierarchical control architectures, which facilitate the separation of complex tasks into manageable layers. This approach not only simplifies the control problem but also enhances the robustness and adaptability of the system. For instance, hierarchical Model Predictive Control (MPC) schemes are being employed to manage posture regulation and push-recovery in humanoid robots, demonstrating significant improvements in stability and recovery performance under disturbances.

The field is also seeing a surge in the use of optimization-based control designs, particularly in the context of trajectory optimization and impedance control. These methods leverage advanced mathematical techniques to generate optimal trajectories and control inputs, enabling legged robots to perform dynamic maneuvers with high precision. The integration of pseudospectral collocation and Lie Group optimization is a testament to this trend, offering versatile frameworks for continuous-time multi-phase problems.

Moreover, there is a growing emphasis on sim-to-real transfer, with researchers developing autotuning methods that bridge the gap between simulation and real-world performance. These methods, often based on differential programming and neural networks, aim to optimize control parameters efficiently, ensuring that simulation-learned models can be effectively transferred to physical hardware.

Noteworthy Papers

• Angular Divergent Component of Motion: Introduces spatial DCM with angular objectives, merging linear and angular dynamics into a unified framework, validated through simulations and hardware experiments.
• Galileo: A Pseudospectral Collocation Framework: Presents a novel transcription scheme for optimizing trajectories on Lie Groups, demonstrating feasibility on various legged robots.
• Adapting Gait Frequency for Posture-regulating Humanoid Push-recovery: Proposes a hierarchical-MPC-based scheme for posture regulation and push-recovery, showing significant improvements in recovery performance.
• Autotuning Bipedal Locomotion MPC with GRFM-Net: Develops an autotuning method for bipedal locomotion control, effectively reducing the sim-to-real gap and improving parameter transferability.