Simulation-based control of autonomous devices
In an embodiment, a system calculates a distribution of possible parameters for a simulation that cause the simulation to match a measured behavior in the real world. In an embodiment, the system selects a plurality of simulation parameters based on a statistical distribution that represents an initial estimate of possible parameter values. In an embodiment, using the results produced by the simulation, an updated distribution of possible parameters is constructed based on a density of the results modeled using Fourier features. In an embodiment, the updated distribution of possible parameters can be used to select a particular set of parameters for the simulation, which cause the simulator to approximate the measured behavior.
1 . One or more processors, comprising circuitry to:
obtain one or more Fourier-transformed representations computed for one or more measured results of a robotic device performing a task in the physical world;
simulate a virtual robotic device performing the task using one or more simulation parameters selected based on the one or more Fourier-transformed representations;
generate one or more multimodal probability distributions based, at least in part, on one or more results of the simulation; and
cause, using a controller of the robotic device and based at least on a random sampling of the one or more multimodal probability distributions, the robotic device to repeat at least one performance of the task in the physical world.
2 . The one or more processors of claim 1 , wherein the one or more Fourier-transformed representations model a density function based, at least in part, on results of one or more additional simulations.
3 . The one or more processors of claim 2 , wherein the density function is parameterized as a set of Fourier features.
4 . The one or more processors of claim 1 , wherein the one or more multimodal probability distributions are to be generated further based, at least in part, on one or more additional simulations performed with a set of parameters chosen in accordance with a predicted prior distribution of parameters.
5 . The one or more processors of claim 1 , wherein the one or more multimodal probability distributions, as a result of being applied to a simulator, are to cause the simulator to approximate the one or more measured results with greater accuracy.
6 . The one or more processors of claim 5 , wherein the simulator is to simulate the virtual robotic device performing the task using the one or more selected simulation parameters.
7 . The one or more processors of claim 5 , wherein applying the one or more multimodal probability distributions to the simulator includes randomizing, by the simulator, over the one or more simulation parameters based, at least in part, on the one or more multimodal probability distributions.
8 . The one or more processors of claim 1 , wherein the circuitry is to cause the one or more multimodal probability distributions to be updated to predict a result of the robotic device performing the task.
9 . The one or more processors of claim 1 , wherein the one or more multimodal probability distributions comprise one or more mixtures of Gaussian distributions.
10 . A system, comprising one or more processors to:
obtain one or more Fourier-transformed representations computed for one or more measured results of a robotic device performing a task in the physical world;
simulate a virtual robotic device performing the task using one or more simulation parameters selected based on the one or more Fourier-transformed representations;
generate one or more multimodal probability distributions based, at least in part, on one or more results of the simulation; and
cause, using a controller of the robotic device and based at least on a random sampling of the one or more multimodal probability distributions, the robotic device to repeat at least one performance of the task in the physical world.
11 . The system of claim 10 , wherein the one or more Fourier-transformed representations model a density function based, at least in part, on results of one or more additional simulations.
12 . The system of claim 11 , wherein:
the density function is modeled as a set of Fourier features; and
the set of Fourier features is selected using Halton sequences.
13 . The system of claim 11 , wherein the density function is modeled as a set of randomly selected Fourier features.
14 . The system of claim 10 , wherein the one or more multimodal probability distributions are to be used to estimate a second distribution of parameter values, and wherein the one or more multimodal probability distributions are to be generated further based, at least in part, on one or more additional simulations performed by a simulator using sets of parameters chosen in accordance with a first distribution of parameter values, the first distribution of parameter values generated prior to the second distribution of parameter values.
15 . The system of claim 14 , wherein the simulator is to produce results for individual parameter sets in the sets of parameters.
16 . The system of claim 10 , wherein the one or more multimodal probability distributions are to model a non-Gaussian distribution that indicates a plurality of parameter solutions.
17 . The system of claim 10 , wherein the one or more multimodal probability distributions are to be applied to a simulator to improve an accuracy thereof.
18 . A non-transitory computer-readable storage medium having stored thereon a set of instructions that, as a result of being performed by one or more processors, cause the one or more processors to at least:
obtain one or more Fourier-transformed representations computed for one or more measured results of a robotic device performing a task in the physical world;
simulate a virtual robotic device performing the task using one or more simulation parameters selected based on the one or more Fourier-transformed representations;
generate one or more multimodal probability distributions based, at least in part, on one or more results of the simulation; and
cause, using a controller of the robotic device and based at least on a random sampling of the one or more multimodal probability distributions, the robotic device to repeat at least one performance of the task in the physical world.
19 . The non-transitory computer-readable storage medium of claim 18 , wherein the one or more Fourier-transformed representations model a density based, at least in part, on parameter-result pairs produced by one or more additional simulations.
20 . The non-transitory computer-readable storage medium of claim 19 , wherein the density is modeled as a set of Fourier features.
21 . The non-transitory computer-readable storage medium of claim 20 , wherein the set of Fourier features is determined in accordance with a quasi Monte Carlo strategy.
22 . The non-transitory computer-readable storage medium of claim 18 , wherein the instructions, as a result of being performed by the one or more processors, further cause the one or more processors to use additional simulations selected in accordance with the one or more multimodal probability distributions to produce a refined distribution of the one or more simulation parameters.
23 . The non-transitory computer-readable storage medium of claim 18 , wherein the one or more multimodal probability distributions are to be generated further based, at least in part, on one or more additional simulations performed with a set of parameters chosen in accordance with a bounded uniform prior.
24 . The non-transitory computer-readable storage medium of claim 18 , wherein the one or more multimodal probability distributions are to be generated further based, at least in part, on one or more additional simulations performed with a set of parameters chosen in accordance with a Gaussian prior.
25 . The non-transitory computer-readable storage medium of claim 18 , wherein the one or more multimodal probability distributions are to improve an accuracy of a simulator as a result of being applied to the simulator.