A four-legged robot can quickly figure out how to navigate new obstacle courses by climbing up and down large boxes, crouching to scoot beneath tables and even jumping across gaps.
The study shows how artificial intelligence training can help robots quickly adapt to new indoor and outdoor environments with a versatility that surpasses traditional robotics software based on pre-programmed behaviours.
鈥淲e replace the standard software of most robots with neural networks,鈥� says at the US computing firm NVIDIA and ETH Zurich in Switzerland. 鈥淭his allows the robot to achieve behaviours that were not possible otherwise.鈥�
Advertisement
Rudin and his colleagues trained three separate neural networks: a perception module to understand the robot鈥檚 surroundings, a navigation module to plan out its path and navigate obstacles and a locomotion module to control the motors that let the robot perform specific movements. The training occurred during millions of trial-and-error runs in a computer simulation, which included diverse scenarios with randomised environmental obstacles.
Next, the robot used its training to navigate three real-world obstacle courses. In each scenario, the robot鈥檚 perception module first used an onboard laser sensor to create a 3D map of its surroundings. That map allowed the robot鈥檚 navigation module to plot a course of action, which its locomotion module then used to execute the necessary robotic movements.
In demonstrations, the trained robot nimbly crossed challenging terrains at speeds of up to 2 metres per second and leapt across gaps of up to 1 metre. 鈥淲e do not perform any pre-mapping or pre-planning,鈥� says at NVIDIA and ETH Zurich, a coauthor on the study. 鈥淓verything is happening online and in real-time.鈥�
But the AI training has its limits. The robot鈥檚 capabilities still depend upon having a diversity of training simulations, which require a lot of time and effort for human researchers to develop, says Hoeller. The robot could also potentially perform more precise 3D mapping and plan better foot placements if it had more onboard computing power, he says.
Science Robotics
