System and methods for pixel based model predictive control
Techniques are disclosed that enable model predictive control of a robot based on a latent dynamics model and a reward function. In many implementations, the latent space can be divided into a deterministic portion and stochastic portion, allowing the model to be utilized in generating more likely robot trajectories. Additional or alternative implementations include many reward functions, where each reward function corresponds to a different robot task.
1 . A robot comprising:
one or more actuators;
memory storing a trained latent robot dynamics model and a trained reward function for a robot task;
one or more processors configured to:
predict multiple candidate sequences of actions using a latent state observation of the robot and the latent robot dynamics model;
select, from the multiple candidate sequences of actions, using the trained reward function, a sequence of actions for performing the robot task; and
control one or more actuators based on the selected sequence of actions to cause performance of the robot task.
2 . The robot of claim 1 , wherein the latent robot dynamics model is a deterministic belief state model (DBSM).
3 . The robot of claim 2 , wherein the DBSM includes an encoder network and a decoder network.
4 . The robot of claim 3 , wherein the encoder network is used in deterministically updating a deterministic activation vector based on processing the latent state observation of the robot, and one or more historical latent state observations of the robot.
5 . The robot of claim 4 , wherein the DBSM further includes a transition function and wherein the transition function is used to generate one or more future states of the robot based on the latent state observation and the sequence of actions for performing the robot task.
6 . The robot of claim 5 , wherein the decoder network is used to generate the sequence of actions for performing the robot task based on the latent state observation of the robot, the one or more historical latent state observations of the robot, and the one or more future states of the robot.
7 . A robot comprising:
one or more actuators;
memory storing a trained latent robot dynamics model and a trained reward function for a robot task;
one or more processors configured to:
use a latent state observation of the robot, the latent robot dynamics model, and the trained reward function, to determine a sequence of actions for performing the robot task; and
control one or more actuators based on the determined sequence of actions to cause performance of the robot task;
wherein, subsequent to controlling the one or more actuators based on the determined sequence of actions to cause performance of the robot task, one or more of the processors are further configured to:
identify an updated latent state observation of the robot;
use the updated latent state observation of the robot, the latent robot dynamics model, and the trained reward function, to determine an updated sequence of actions for performing the robot task; and
control the one or more actuators based on the determined updated sequence of actions to cause performance of the robot task.
8 . A robot comprising:
one or more actuators;
memory storing a trained latent robot dynamics model and a trained reward function for a robot task;
one or more processors configured to:
use a latent state observation of the robot, the latent robot dynamics model, and the trained reward function, to determine a sequence of actions for performing the robot task; and
control one or more actuators based on the determined sequence of actions to cause performance of the robot task;
wherein, subsequent to controlling the one or more actuators based on the determined sequence of actions to cause performance of the robot task, one or more of the processors are further configured to:
use an additional latent state observation of the robot, the latent robot dynamics model, and the trained reward function, to determine an additional sequence of actions for performing an additional robot task; and
control the one or more actuators based on the additional sequence of actions to cause performance of the additional robot task.
9 . A method implemented by one or more processors, the method comprising:
predicting multiple candidate sequences of actions using a latent state observation of a robot and a latent robot dynamics model;
selecting, from the multiple candidate sequences of actions, using a trained reward function a sequence of actions for performing a robot task, wherein the trained reward function is for the robot task; and
controlling one or more actuators based on the selected sequence of actions to cause performance of the robot task.
10 . The method of claim 9 , wherein the latent robot dynamics model is a deterministic belief state model (DBSM).
11 . The method of claim 10 , wherein the DBSM includes an encoder network and a decoder network.
12 . The method of claim 11 , wherein the encoder network is used in deterministically updating a deterministic activation vector based on processing the latent state observation of the robot, and one or more historical latent state observations of the robot.
13 . The method of claim 12 , wherein the DBSM further includes a transition function and wherein the transition function is used to generate one or more future states of the robot based on the latent state observation and the sequence of actions for performing the robot task.
14 . The method of claim 13 , wherein the decoder network is used to generate the sequence of actions for performing the robot task based on the latent state observation of the robot, the one or more historical latent state observations of the robot, and the one or more future states of the robot.
15 . A method implemented by one or more processors, the method comprising:
using a latent state observation of a robot, a latent robot dynamics model, and a trained reward function, to determine a sequence of actions for performing a robot task, wherein the trained reward function is for the robot task; and
controlling one or more actuators based on the determined sequence of actions to cause performance of the robot task;
wherein, subsequent to controlling the one or more actuators based on the determined sequence of actions to cause performance of the robot task, the method further comprises:
identifying an updated latent state observation of the robot;
using the updated latent state observation of the robot, the latent robot dynamics model, and the trained reward function, to determine an updated sequence of actions for performing the robot task; and
controlling the one or more actuators based on the determined updated sequence of actions to cause performance of the robot task.
16 . A method implemented by one or more processors, the method comprising:
using a latent state observation of a robot, a latent robot dynamics model, and a trained reward function, to determine a sequence of actions for performing a robot task, wherein the trained reward function is for the robot task; and
controlling one or more actuators based on the determined sequence of actions to cause performance of the robot task;
wherein, subsequent to controlling the one or more actuators based on the determined sequence of actions to cause performance of the robot task, the method further comprises:
using an additional latent state observation of the robot, the latent robot dynamics model, and the trained reward function, to determine an additional sequence of actions for performing an additional robot task; and
controlling the one or more actuators based on the additional sequence of actions to cause performance of the additional robot task.