# US 2021/0171024 A1 · Deep learning based motion control of a group of autonomous vehicles > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2021/0171024 A1 - Document ID: US20210171024A1 - Status: Published application - Kind: A1 - Application number: 17/106,436 - Filed: 2020-11-30 - Published: 2021-06-10 - Earliest claimed: 2019-12-06 - Assignees: Elektrobit Automotive GmbH - Applicants: Elektrobit Automotive GmbH - Inventors: Sorin Mihai GRIGORESCU - Web page: https://patentgrep.com/patent/US20210171024A1 ## CPC - G05D 1/0221: {involving a learning process} - B60W 30/095: Predicting travel path or likelihood of collision - B60W 60/0011: {involving control alternatives for a single driving scenario, e.g. planning several paths to avoid obstacles} - B60W 60/0027: {using trajectory prediction for other traffic participants} - G05D 1/0289: {with means for avoiding collisions between vehicles (vehicle fittings for automatically controlling speed including means for detecting potential obstacles B60K31/0008; avoiding obstacles by action on the steering system B62D; radar, sonar, lidar systems designed for anti-collision purposes G01S13/93, G01S15/93, G01S17/93)} - G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting - G06N 3/006: based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO] - G06N 3/044: Recurrent networks, e.g. Hopfield networks - G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/049: Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs - G06N 3/08: Learning methods - G06N 3/084: Backpropagation, e.g. using gradient descent - G06N 3/092: Reinforcement learning - G08G 1/164: {Centralised systems, e.g. external to vehicles} - G05D: SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES - B60W: CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G08G: TRAFFIC CONTROL SYSTEMS (guiding railway traffic, ensuring the safety of railway traffic B61L; arrangement of road signs or traffic signals E01F9/00; radar or analogous systems, sonar systems, lidar systems specially adapted for traffic control G01S13/91, G01S15/88, G01S17/88; {radar or analogous systems, sonar systems, lidar systems specially adapted for anti-collision purposes G01S13/93, G01S15/93, G01S17/93}) ## Abstract A controller for an agent of a group of agents, in particular for a group of autonomous or semi-autonomous vehicles, and to a computer program implementing such a controller. A temporal deep network for such a controller and to a method, a computer program and an apparatus for training the temporal deep network. The controller includes a temporal deep network designed to calculate a desired trajectory for the agent, a nonlinear model predictive controller designed to calculate commands for the agent based on the desired trajectory and desired trajectories of the other agents of the group of agents, and an augmented memory designed to integrate historic system states of the group of agents for the temporal deep network. ## Claims ### Claim 1 A controller for an agent (Ai) of a group of agents (An), the controller comprising: a temporal deep network configured to calculate a desired trajectory for the agent (Ai); a nonlinear model predictive controller configured to calculate commands for the agent (Ai) based at least in part on the desired trajectory and desired trajectories of other agents (An) of the group of agents (An); and an augmented memory configured to integrate historic system states of the group of agents (An) for the temporal deep network. ### Claim 2 (depends on 1) The controller according to claim 1, wherein the historic system states of the group of agents (An) comprise historic states and observations of the group of agents (An). ### Claim 3 (depends on 1) The controller according to claim 1, wherein the temporal deep network comprises a long short-term memory recurrent neural network. ### Claim 4 (depends on 1) The controller according to claim 1, wherein the controller is configured to consider a collision avoidance constraint for each agent (An). ### Claim 5 (depends on 1) The controller according to claim 1, wherein the controller is configured to share the desired trajectory of the agent (Ai) and observations of the agent (Ai) with the other agents (An) of the group of agents (An). ### Claim 6 A computer program code comprising instructions, which, when executed by at least one processor, cause the at least one processor to implement a controller for an agent (Ai) of a group of agents (An), the controller comprising: a temporal deep network configured to calculate a desired trajectory for the agent (Ai); a nonlinear model predictive controller configured to calculate commands for the agent (Ai) based at least in part on the desired trajectory and desired trajectories of other agents (An) of the group of agents (An); and an augmented memory configured to integrate historic system states of the group of agents (An) for the temporal deep network. ### Claim 7 (depends on 1) A temporal deep network for a controller according to claim 1 for an agent (Ai) of a group of agents (An), wherein the temporal deep network is configured to: calculate a desired trajectory for the agent (Ai) based at least in part on historic observations of the agent (Ai); calculate a reference trajectory for the agent (Ai); and calculate historic states of all agents (An) of the group of agents. ### Claim 8 (depends on 7) The temporal deep network according to claim 7, wherein the temporal deep network comprises a long short-term memory recurrent neural network. ### Claim 9 (depends on 7) The temporal deep network according to claim 7, wherein the temporal deep network is trained based on inverse reinforcement learning. ### Claim 10 A computer program code comprising instructions, which, when executed by at least one processor, cause the at least one processor to implement a temporal deep network to: calculate a desired trajectory for an agent (Ai) based at least in part on historic observations of the agent (Ai); calculate a reference trajectory for the agent (Ai); and calculate historic states of all agents (An). ### Claim 11 A method for training a temporal deep network, the method comprising: acquiring trajectories from manually driving agents (An) in a test environment; and training the temporal deep network using inverse reinforcement learning based at least in part on trajectories acquired from the manually driving agents (An) in the test environment. ### Claim 12 (depends on 11) The method according to claim 11, wherein parameters of the temporal deep network are learned by minimizing a loss function in a maximum likelihood estimation setup. ### Claim 13 A computer program code comprising instructions, which, when executed by at least one processor, cause the at least one processor to acquire trajectories from manually driving agents (An) in a test environment; and train a temporal deep network using inverse reinforcement learning based at least in part on the trajectories acquired from the manually driving agents (An) in the test environment. ### Claim 14 An apparatus for training a temporal deep network comprising: a processor configured to train the temporal deep network using inverse reinforcement learning based on trajectories acquired from manually driving agents (An) in a test environment. ### Claim 15 An autonomous or semi-autonomous vehicle, characterized in that the autonomous or semi-autonomous vehicle comprises: a controller comprising: a temporal deep network configured to calculate a desired trajectory for an agent (Ai); a nonlinear model predictive controller configured to calculate commands for the agent (Ai) based at least in part on the desired trajectory and desired trajectories of other agents (An) of a group of agents (An); and an augmented memory configured to integrate historic system states of the group of agents (An) for the temporal deep network.