Rapid Perception-Based Terrain-Adaptive Agile Locomotion of a Humanoid Robot

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

This award supports research that enables robots to perform hazardous and difficult tasks typically performed by humans such as search and rescue in natural disasters, thereby promoting the progress of science, advancing prosperity and advancing human welfare. These operations often involve navigating wreckage, disconnected terrains, and other dangerous areas that put humans at risk. The two-legged form of humans makes them naturally adept at navigating these difficult scenarios. While significant progress has been made in developing human-like bipedal (two-footed) robots, they still lack the rapid navigation and locomotion capabilities needed for highly irregular terrains. This project will solve this challenge and hence advance robotics by enabling bipedal robots to dynamically navigate chaotic and unstructured environments. These robots will enhance societal safety by performing hazardous tasks, thus safeguarding lives during natural disasters and industrial accidents. Beyond technological advancements, all project outcomes will be made available as open-source, aiming to reduce the entry barriers to humanoid robot research. A new undergraduate course will also be developed using the newly developed simulation software that integrates perception, control, and drivetrain mechanisms. This research aims to make fundamental contributions to humanoid robots that are capable of rapid navigation in complex environments, representing a paradigm shift, and the extension of their applications from controlled spaces to open, dynamic, and unstructured settings. It will achieve this goal by the integration of three critical areas: rapid perception to assess terrain shape and estimate body pose relative to steppable areas during swift movements, locomotion planning to formulate optimal movement strategies based on perception data, and robot hardware design to enable precise body control and forceful ground push-off. This project will accomplish the integrated system-level agility of the humanoid robot rather than solely focus on improving the performance of individual components. The technical approach includes: (1) an event camera-based pose tracking framework for accurate, real-time pose estimation during rapid movements, involving novel tracking algorithms and effective steppable region representation; (2) a novel locomotion controller combining a neural network policy for high-level decision-making with a low-level motion controller, utilizing a simplified model-based trajectory optimization for efficient and scalable learning; and (3) a unique humanoid robot leg design featuring innovative co-actuation and toe mechanisms for explosive ground reaction forces and secure ground contact. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.