If we are to fulfil the dream of robots becoming fixtures in our homes and providing assistance with everyday tasks, we need to focus on making them autonomous, a leading robotics researcher has said.
Dr Petar Kormushev, of the Faculty of Engineering at Imperial College London’s Dyson School of Design Engineering and former member of the WALK-MAN robot team in this year’s DARPA Robotics Challenge, made the comments at Imperial’s Robotics Forum Showcase on Friday.
“If we really want robots to go into our homes and do tasks for us, they need to be autonomous and learn from their actions,” he said.
“For me the obvious answer to this problem is to give robots the opportunity to learn for themselves.”
Images courtesy of DARPA
However, while this may be the best way to achieve the robot butlers we’ve long dreamed of, making robots behave autonomously remains a phenomenal challenge.
The DARPA Robotics Challenge, which required robots to perform a serious of tasks to simulate a disaster situation, provided a clear demonstration of this, as Kormushev, who was involved in WALK-MAN’s locomotion and automation, explains:
“Although these tasks were predefined, the biggest challenge was doing them partially autonomously,” he said.
However, he was critical of the level of autonomy asked of the robots in development, suggesting that DARPA should have required more of the teams.
“Most of the tasks were based on teleoperation,” he said, referring to the practice of controlling the robots from a distance rather than allowing them to complete the predefined task on their own,”and that was a bit of a disappointment.”
However, while Kormushev is a proponent of further automation and self-learning in robots, he accepts that there are limiting factors on how well a robot can learn.
Perhaps the most popular method of self-learning in robots is imitation learning, where the robot is programmed to observe and learn by copying the actions it sees.
“But the limiting factor is that the robot can only be as good as the person they are being taught by,” he said, adding that this approach relied too heavily on objects being placed in specific locations, which made it unsuitable for advanced interaction in an ever-changing household environment.
Instead he suggests that reinforcement learning, where the focus is on the appearance of the object and positive feedback teaches the robot it is taking the correct course of action, similar to how children learn.
However, ultimately if we are to have robot butlers tending to our every need, programmers will need to prioritise making robots able to react and respond to their changing environment.
“It’s obvious that precision and speed are no longer the most important features of the robot, at least in our home,” he said.
