快猫短视频

An AI has been used to train a small robot dog to perform cleaning tasks. The hardware cost a total of $6300, which is less than a tenth of the price tag of the well-known robot dogs built by US tech firm Boston Dynamics.

This type of self-taught robotic body coordination relies on an AI training regimen that could pave the way for affordable robot dogs and possibly even humanoid robots that could be used as helpers in homes and workplaces. 鈥淎 single robot should do multiple tasks, and those robots should be safe and low cost,鈥� says at Carnegie Mellon University in Pennsylvania.

Pathak and his colleagues trained their AI to smoothly coordinate the movements of a four-legged robot with an arm attached to its back while a human operator directed the robot鈥檚 general activity through a joystick or other means. They used reinforcement learning to train the AI through trial-and-error repetition in both computer simulations and in a physical robot.

The team first trained the AI to control the robot鈥檚 legs independently from the arm, before training it on the whole-body control challenge of coordinating leg and arm movements, says , also at Carnegie Mellon University. For example, the AI learned on its own to bend and stretch certain legs to help maximise the reach of the robot鈥檚 arm as it tossed cups into a garbage bin or wiped a whiteboard with an eraser.

Another training trick involved having one version of the AI teach another what it was learning about how to most smoothly move the robot鈥檚 limbs. The researchers had a 鈥渢eacher鈥� AI agent that had trained in a simulation become the role model for a 鈥渟tudent鈥� AI agent that tried to mimic the teacher鈥檚 body movements while navigating an environment using only information available through the robot鈥檚 cameras or other sensors.

鈥淚nstead of just blindly training by using reinforcement learning in a simulation, we take into consideration whether this kind of [AI] agent trained in the simulation can actually adapt to the robot by using only the limited amount of observations available onboard,鈥� says at Stanford University in California.

Traditionally, teaching robots to walk and manipulate an arm at the same time has been a 鈥渟uper hard problem鈥�, requiring human engineers to hand-code the robot鈥檚 software with knowledge of how its body works and manually tune the robotic movements, says at the University of Texas at Austin, who was not involved in the study.

This time-consuming human engineering has enabled companies such as Boston Dynamics to create dancing robot dogs capable of opening doors 鈥� but an AI-powered approach such as this one could make whole-body control much more accessible and affordable for many different robot bodies. 鈥淜udos to the team because they鈥檙e doing it with a low-cost robot,鈥� says Sentis.

Reference: