System and method for robot learning from human demonstrations with formal logic
Systems and methods are provided for learning control policies that are safe, robust, and interpretable in developing robotic systems. In one example, a signal temporal logic is used to evaluate and rank the quality of demonstrations. The temporal logic based specifications allow generation of non-Markovian rewards, and also define causal dependencies between tasks such as sequential task specifications.
1. A system for reinforcement learning, the system comprising:
one or more processors;
a computer-readable medium storing executable instructions that, when executed, cause the system to perform operations comprising:
receive a set of demonstrations, the set of demonstrations obtained via interaction of an agent with an environment determined according to sensor input from one or more sensors of the agent;
receive a set of specifications, the set of specifications providing descriptions of one or more tasks and/or one or more objectives;
convert the set of specifications into a signal temporal logic formal language in real-time;
evaluate the set of demonstrations based on the set of specifications in the signal temporal logic formal language;
generate a robustness value for each demonstration in the set of demonstrations based on the evaluation;
infer rewards for each demonstration based on the robustness value;
learn a control policy based on the inferred rewards; and
provide one or more control signals to one or more actuators of the agent based on the control policy.
2. The system of claim 1 , wherein the computer-readable medium stores further instructions that when executed cause the system to:
verify the learned control policy based on the set of specifications to determine a final policy.
3. The system of claim 1 , wherein the computer-readable medium stores further instructions that when executed cause the system to:
for each demonstration comprising a set of state and action pairs, generate a state reward corresponding to each state; and
generate a candidate reward for each demonstration based on the state reward for each state in the set of states for each demonstration.
4. The system of claim 3 , wherein inferring rewards for each demonstration comprises ranking each demonstration based on the robustness value, and determining a learner reward for the agent based on the ranks and corresponding candidate rewards for each demonstration.
5. The system of claim 1 , wherein the control policy is determined based on a reinforcement learning algorithm.
6. The system of claim 1 , wherein the set of specifications are provided in natural language.
7. The system of claim 1 , wherein the agent is selected from the group consisting of a cyber-physical system, a robotic system, an autonomous vehicle, an insulin delivery system, and a drone.
8. The system of claim 1 , wherein each specification in the set of specifications is represented as a directed acyclic graph (DAG).
9. A system, comprising:
one or more sensors configured to acquire environmental data of an environment interacting with the system;
one or more processors;
a computer-readable medium storing executable instructions that, when executed, cause the system to perform operations comprising:
evaluate a current state of the system according to the environmental data from the one or more sensors; and
determine an action to be performed by the system based on the current state according to a control policy;
wherein the control policy is learned based on inferred rewards from a plurality of demonstrations; and
wherein the plurality of demonstrations are evaluated and ranked based on a robustness value of each demonstration using a set of specifications in a formal language; and
provide control signals to one or more actuators based on the control policy.
10. The system of claim 9 , wherein the formal language is selected from the group consisting of is selected from the group consisting of a temporal logic, a Signal Temporal Logic (STL), a Linear Temporal Logic (LTL), and a Computation Tree Logic (CTL).
11. The system of claim 9 , wherein the inferred rewards are based on uncertainties in the environment; and wherein the inferred reward increases with increase in uncertainty.
12. A method for performing reinforcement learning, the method comprising:
receiving a set of demonstrations, the set of demonstrations obtained via interaction of an agent with an environment;
receiving a set of specifications, the set of specifications providing descriptions of one or more tasks and/or one or more objectives;
converting the set of specifications into a temporal logic formal language;
evaluating the set of demonstrations based on the set of specifications in the temporal logic formal language;
generating a robustness value for each demonstration in the set of demonstrations based on the evaluation;
inferring rewards for each demonstration based on the robustness value;
learning a control policy based on the inferred rewards; and
storing the control policy; and
provide one or more control signals to one or more actuators of the agent based on the control policy.
13. The method of claim 12 , wherein the temporal logic is selected from the group consisting of Signal Temporal Logic (STL), Linear Temporal Logic (LTL), and Computation Tree Logic (CTL).
14. The method of claim 12 , further comprising, for each demonstration comprising a set of state and action pairs, generating a state reward corresponding to each state.
15. The method of claim 14 , further comprising, generating a candidate reward for each demonstration based on the state reward for each state in the set of states.
16. The method of claim 15 , wherein inferring rewards for each demonstration comprises ranking each demonstration based on the robustness value, and determining a learner reward for the agent based on the ranks and corresponding candidate rewards for each demonstration.
17. The method of claim 15 , wherein the agent is selected from the group consisting of a cyber-physical system, a robotic system, an autonomous vehicle, an insulin delivery system, and a drone.