Four Legged Robot Masters Forests Stairs And Obstacles With Animal Like Agility

Washington DC – A four-legged robot has demonstrated the ability to quickly switch between different motor skills to navigate uneven terrain, move through forests, climb stairs, and jump over obstacles such as fallen branches and stepping stones. Researchers developed a new reinforcement learning framework that allows the robot to select from a set of pre-trained movement skills and autonomously adapt to different environments. The system, called action pretrained transformer-based reinforcement learning, or APT-RL, helps the robot perform agile and high-speed movements using only onboard perception and computation. The framework includes three learning phases. The robot first learns basic gait patterns, including trotting and bounding, on flat terrain using a large two-dimensional motion dataset. It then learns to choose suitable movements for different obstacles and environments through reinforcement learning. In the final stage, the robot improves its actions using real-world sensory information. Tests showed that the robot equipped with APT-RL could travel across paved pathways and forested areas, climb stairs, and jump over obstacles by switching between different motor skills. Researchers said future versions of the framework could help enable autonomous search and rescue operations or exploration missions.