Computer-implemented method and device for controlling a mobile robot based on semantic environment maps

A control unit of a mobile robot configured to determine a motion trajectory for the mobile robot based on occupancy priors indicating probabilities of presence of dynamic objects and/or individuals in a map of an environment, wherein the occupancy priors are determined by a reward function defined by reward function parameters, and wherein, for determining of the reward function parameters, the control unit is configured to:

• provide a number of semantic maps;
• provide a number of training trajectories for each of the number of semantic maps;
• compute a gradient as a difference between an expected mean feature count and an empirical mean feature count depending on each of the number of semantic maps and on each of the number of training trajectories, wherein the empirical mean feature count is a first sum of averages, each of the averages of the first sum being an average number of features accumulated over the provided training trajectories of a different semantic map of a plurality of the semantic maps, and wherein the expected mean feature count is a second sum of averages, each of the averages of the second sum being an average number of features accumulated by trajectories generated, for a different semantic map of the plurality of semantic maps, depending on current reward function parameters;
• and
• update the reward function parameters depending on the gradient.

A robotic device, comprising:

• a control unit control unit configured to determine a motion trajectory for the robotic device based on occupancy priors indicating probabilities of presence of dynamic objects and/or individuals in a map of an environment, wherein the occupancy priors are determined by a reward function defined by reward function parameters, and wherein, for determining of the reward function parameters, the control unit is configured to:
• provide a number of semantic maps,
• provide a number of training trajectories for each of the number of semantic maps,
• compute a gradient as a difference between an expected mean feature count and an empirical mean feature count depending on each of the number of semantic maps and on each of the number of training trajectories, wherein the empirical mean feature count is a first sum of averages, each of the averages of the first sum being an average number of features accumulated over the provided training trajectories of a different semantic map of a plurality of the semantic maps, and wherein the expected mean feature count is a second sum of averages, each of the averages of the second sum being an average number of features accumulated by trajectories generated, for a different semantic map of the plurality of semantic maps, depending on current reward function parameters, and
• update the reward function parameters depending on the gradient;
• and an actuation unit for interaction with the environment;
• wherein the actuation unit is controlled by the control unit according to the determined motion trajectory.

A non-transitory machine readable medium on which are stored instructions, the instructions, when executed by a computer, causing the computer to determine a motion trajectory for a mobile robot based on occupancy priors indicating probabilities of presence of dynamic objects and/or individuals in a map of an environment, wherein the occupancy priors are determined by a reward function defined by reward function parameters, wherein for the determining of the reward function parameters, the instructions cause the computer to perform the following steps:

• providing a number of semantic maps;
• providing a number of training trajectories for each of the number of semantic maps;
• computing a gradient as a difference between an expected mean feature count and an empirical mean feature count depending on each of the number of semantic maps and on each of the number of training trajectories, wherein the empirical mean feature count is a first sum of averages, each of the averages of the first sum being an average number of features accumulated over the provided training trajectories of a different semantic map of a plurality of the semantic maps, and wherein the expected mean feature count is a second sum of averages, each of the averages of the second sum being an average number of features accumulated by trajectories generated, for a different semantic map of the plurality of semantic maps, depending on current reward function parameters;
• and
• updating the reward function parameters depending on the gradient.