A team of scientists from Carnegie Mellon University and the University of California have recently announced a breakthrough in the training of a four-legged robotic dog. This robot has been trained to use its front legs to interact with its environment, such as activating switches or climbing walls.
In order to achieve this, the scientists separated the training into separate locomotion and manipulation strategies, as reinforced learning was not effective in teaching the robot to move while using its front legs to manipulate objects due to the contradictory movement sequences.
The scientists initially carried out both training sessions in a simulation, however, this was met with extra hassle and effort in comparison to increased learning. The skills trained in this way had to then be combined.
From human commands, the robot creates a behavior tree that includes the tasks in a series of interconnected subtasks for locomotion and manipulation. The robot can then process and execute these subtasks one after the other. This also means that if a subtask fails, the robot can go back in the behavior tree until it has successfully completed a task, making the system more robust.
The robot used for this experiment was a Unitree Go 1 robotic dog, equipped with Intel’s RealSense technology for perception. The robot was able to stand up against a wall and press a button situated about one meter high for wheelchair users. The door then opened, allowing the robot to go through it.
The scientists see this as a significant step in enabling non-humanoid robots to move into environments designed for humans. Additionally, the scientists had the robotic dog kick a soccer ball with its front legs to demonstrate other object manipulations that are possible.
This scientific paper, titled “Legs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion, is scheduled to be presented at the ICRA 2023 robotics fair, set to take place in London from May 29 to June 2.
