Imitation Learning enables natural human-robot interaction

Imitation Learning enables natural human-robot interaction

Telepresence robots are used to bring friends, relatives, or healthcare professionals closer to patients in care homes, hospitals, or at home. This significantly improves quality of life for people who are not easily able to travel or see their loved ones in person regularly. However, these robots quickly reach their limits when the remote pilot wants to interact naturally with several people, moving around among them and dipping in and out of conversations – for example at family gatherings.

These situations are difficult for conventional telepresence robots because the operator needs to manually control the robot and it is difficult to pick up on subtle behavioural cues in a group of people seen through a webcam. Alternatively, an autonomous robot can be used, but the behaviour of these is often clumsy. That’s because it’s extremely difficult to explain to a robot how to behave like a human.

For example, imagine needing to get from the hallway, through the living room, to the kitchen during a family event with small groups of people standing, or sitting, in the room. How does a robot navigate this situation without being a nuisance and getting in the way? This is extremely hard to define a priori for all possible situations, so a robot needs some kind of social intelligence of its own.

The TERESA project, led by Morpheus (now Latent Logic’s) founder and Chief Scientist, Shimon Whiteson, created a telepresence robot that displays social intelligence, allowing it to integrate smoothly and naturally into complex social situations. This social intelligence is learned by Imitation Learning, meaning that the Robot doesn’t plan its actions to get from A to B as quickly or efficiently as possible, but to do it in such a way that most resembles the way it has seen humans do it. TERESA lets the operator provide high-level instructions such as which part of a room or building to move to, or whom to approach, but the robot learns how to implement these instructions like a human would. Just a few hours of video data were needed to train the robot to do this.

At Morpheus (now Latent Logic), we’ve recognised that teaching machines to learn by simply observing human behaviour has tremendous potential in a range of applications and are now working towards building behaviour models of humans in traffic. These models enable realistic simulation-based testing of autonomous vehicles.

You can read more about the TERESA project at