This thesis aims to investigate systems and tools for teleoperating a humanoid robot. Robot teleoperation is crucial to send and control robots in environments that are dangerous or inaccessible for humans (e.g., disaster response scenarios, contaminated environments, or extraterrestrial sites). The term teleoperation most commonly refers to direct and continuous remote control of a robot. In this case, the human operator guides the motion of the robot with her/his own physical motion or through some physical input device. One of the main challenges is to control the robot in a way that guarantees its dynamical balance while trying to follow the human references. In addition, the human operator needs some feedback about the state of the robot and its work site through remote sensors in order to comprehend the situation or feel physically present at the site, producing effective robot behaviors. Complications arise when the communication network is non-ideal. In this case the commands from human to robot together with the feedback from robot to human can be delayed. These delays can be very disturbing for the human operator, who cannot teleoperate their robot avatar in an effective way.
Another crucial point to consider when setting up a teleoperation system is the large number of parameters that have to be tuned to effectively control the teleoperated robots. Machine learning approaches and stochastic optimizers can be used to automate the learning of some of the parameters.
In this thesis, we proposed a teleoperation system that has been tested on the humanoid robot iCub. We used an inertial-technology-based motion capture suit as input device to control the humanoid and a virtual reality headset connected to the robot cameras to get some visual feedback. We first translated the human movements into equivalent robot ones by developping a motion retargeting approach that achieves human-likeness while trying to ensure the feasibility of the transferred motion. We then implemented a whole-body controller to enable the robot to track the retargeted human motion. The controller has been later optimized in simulation to achieve a good tracking of the whole-body reference movements, by recurring to a multi-objective stochastic optimizer, which allowed us to find robust solutions working on the real robot in few trials.
To teleoperate walking motions, we implemented a higher-level teleoperation mode in which the user can use a joystick to send reference commands to the robot. We integrated this setting in the teleoperation system, which allows the user to switch between the two different modes.
A major problem preventing the deployment of such systems in real applications is the presence of communication delays between the human input and the feedback from the robot: even a few hundred milliseconds of delay can irremediably disturb the operator, let alone a few seconds. To overcome these delays, we introduced a system in which a humanoid robot executes commands before it actually receives them, so that the visual feedback appears to be synchronized to the operator, whereas the robot executed the commands in the past. To do so, the robot continuously predicts future commands by querying a machine learning model that is trained on past trajectories and conditioned on the last received commands.
