Robot systems, methods, control modules, and computer program products that leverage large language models
Robot control systems, methods, control modules and computer program products that leverage one or more large language model(s) (LLMs) in order to achieve at least some degree of autonomy are described. Robot control parameters and/or instructions may advantageously be specified in natural language (NL) and communicated with the LLM via an NL prompt or query. The LLM module provides a task plan in NL, which can be evaluated for at least one fault or error. If at least one fault or error is identified, the LLM module can be queried to provide a resolution.
1. A method of operation of a robot system, the method comprising:
accessing a task plan by at least one processor of the robot system, the task plan expressed in a natural language (NL) and including a plurality of steps;
determining, by the at least one processor, that the task plan includes at least one fault;
generating, by the at least one processor, an NL description of the at least one fault;
providing a NL query to a large language model (LLM) module, the NL query including at least the NL description of the at least one fault and an NL request for a resolution plan;
receiving the resolution plan from the LLM module, the resolution plan expressed in NL; and
executing an updated task plan by the robot system, the updated task plan based at least in part on the resolution plan.
2. The method of claim 1 , wherein:
the method further includes generating an NL description of any steps of the task plan which do not include faults; and
the NL query further includes the NL description of any steps of the task plan which do not include faults.
3. The method of claim 1 , wherein the resolution plan includes an NL description of at least one resolution selected from a group of resolutions consisting of:
adjustment to order of steps in the task plan;
adjustment to success criteria for any step in the task plan;
adjustment to any actions expressed in any step of the task plan;
adjustment to any targets of interaction for any step of the task plan;
replacement of any step of the task plan with at least one new step;
removal of at least one step of the task plan; and
the updated task plan to replace all steps of the task plan.
4. The method of claim 1 , wherein determining that the task plan includes at least one fault comprises:
executing at least one step of the task plan; and
identifying that execution of the at least one step produces at least one fault scenario.
5. The method of claim 4 , wherein each respective fault scenario includes at least one scenario selected from a group of scenarios consisting of:
execution of the step produces results which are not in alignment with the work objective;
execution of the step inhibits execution of at least one other step in the task plan; and
execution of the step produces at least one unacceptable effect in the environment.
6. The method of claim 1 , wherein determining that the task plan includes at least one fault comprises: determining that the task plan violates any rule in a set of rules specified in at least part of a reasoning engine.
7. The method of claim 1 , wherein determining that the task plan includes at least one fault comprises:
generating a simulation in which a simulated instance of the robot system executes a simulated instance of the task plan; and
identifying at least one fault scenario in the simulation.
8. The method of claim 7 , wherein the at least one fault scenario includes at least one scenario selected from a group of scenarios consisting of:
simulated execution of the task plan produces simulated results which are not in alignment with the work objective;
simulated execution of at least one step in the task plan inhibits execution of at least one other step in the task plan; and
simulated execution of the task plan produces at least one unacceptable effect in the environment model.
9. The method of claim 1 , further comprising generating the task plan by the LLM module.
10. A robot control module comprising at least one non-transitory processor-readable storage medium storing processor-executable instructions or data that, when executed by at least one processor of a processor-based system, cause the processor-based system to:
access a task plan, the task plan expressed in a natural language (NL) and including a plurality of steps;
determine, by the at least one processor, that the task plan includes at least one fault;
generate, by the at least one processor, an NL description of the at least one fault;
provide a NL query to a large language model (LLM) module, the NL query including at least the NL description of the at least one fault and an NL request for a resolution plan;
receive the resolution plan from the LLM module, the resolution plan expressed in NL; and
execute an updated task plan by the processor-based system, the updated task plan based at least in part on the resolution plan.
11. The robot control module of claim 10 , wherein:
the processor-executable instructions or data further cause the at least one processor to generate an NL description of any steps of the task plan which do not include faults; and
the NL query further includes the NL description of any steps of the task plan which do not include faults.
12. The robot control module of claim 10 , wherein the resolution plan includes an NL description of at least one resolution selected from a group of resolutions consisting of:
adjustment to order of steps in the task plan;
adjustment to success criteria for any steps in the task plan;
adjustment to any actions expressed in any steps of the task plan;
adjustment to any targets of interaction for any steps of the task plan;
replacement of any steps of the task plan with at least one new step;
removal of at least one step of the task plan; and
the updated task plan to replace all steps of the task plan.
13. The robot control module of claim 10 , wherein the processor-executable instructions or data which cause the at least one processor to determine that the task plan includes at least one fault cause the processor-based system to:
execute at least one step of the task plan; and
identify, by the at least one processor that execution of the at least one step produces at least one fault scenario.
14. The robot control module of claim 13 , wherein each respective fault scenario includes at least one scenario selected from a group of scenarios consisting of:
execution of the step produces results which are not in alignment with the work objective;
execution of the step inhibits execution of at least one other step in the task plan; and
execution of the step produces at least one unacceptable effect in the environment.
15. The robot control module of claim 10 , wherein the processor-executable instructions or data which cause the at least one processor to determine that the task plan includes at least one fault cause the at least one processor to: determine that the task plan violates any rule in a set of rules specified in at least part of a reasoning engine.
16. The robot control module of claim 10 , wherein the processor-executable instructions or data which cause the at least one processor to determine that the task plan includes at least one fault cause the at least one processor to:
generate a simulation in which a simulated instance of the processor-based system executes a simulated instance of the task plan; and
identify at least one fault scenario in the simulation.
17. The robot control module of claim 16 , wherein the at least one fault scenario includes at least one scenario selected from a group of scenarios consisting of:
simulated execution of the task plan produces simulated results which are not in alignment with the work objective;
simulated execution of at least one step in the task plan inhibits execution of at least one other step in the task plan; and
simulated execution of the task plan produces at least one unacceptable effect in the environment model.
18. The robot control module of claim 10 wherein the processor-executable instructions or data, when executed by at least one processor of the processor-based system, further cause the processor-based system to engage the LLM module to generate the task plan.
19. A robot system comprising:
a robot body;
at least one sensor;
a robot controller including at least one processor and at least one non-transitory processor-readable storage medium, the at least one non-transitory processor-readable storage medium storing processor-executable instructions which when executed by the at least one processor cause the robot system to:
access a task plan, the task plan expressed in a natural language (NL) and including a plurality of steps;
determine, by the at least one processor, that the task plan includes at least one fault;
generate, by the at least one processor, an NL description of the at least one fault;
provide a NL query to a large language model (LLM) module, the NL query including at least the NL description of the at least one fault and an NL request for a resolution plan;
receive the resolution plan from the LLM module, the resolution plan expressed in NL; and
execute an updated task plan by the robot system, the updated task plan based at least in part on the resolution plan.
20. The robot system of claim 19 wherein the at least one non-transitory processor-readable storage medium of the robot system stores the LLM module.