Merge of 1d8f72e

PiperOrigin-RevId: 601753731
Change-Id: Ia0d30df816c8f83285a1b2a619f57500c3a78816

Author: copybara-github

README.md

@@ -16,9 +16,9 @@ real-time predictive control with [MuJoCo](https://mujoco.org/), developed by
 Google DeepMind.
 
 MJPC allows the user to easily author and solve complex robotics tasks, and
-currently supports three shooting-based planners: derivative-based iLQG and
-Gradient Descent, and a simple yet very competitive derivative-free method
-called Predictive Sampling.
+currently supports multiple shooting-based planners. Derivative-based methods include iLQG and
+Gradient Descent, while derivative-free methods include a simple yet very competitive planner
+called Predictive Sampling and the Cross Entropy Method (with diagonal covariance).
 
 - [Overview](#overview)
 - [Graphical User Interface](#graphical-user-interface)

docs/OVERVIEW.md

@@ -132,7 +132,7 @@ Values in brackets should be replaced by the designer.
 
 `Agent` settings can be specified by prepending `agent_` for corresponding class members.
 
-`Planner` settings can similarly be specified by prepending the corresponding optimizer name, (e.g., `sampling_`, `gradient_`, `ilqg_`).
+`Planner` settings can similarly be specified by prepending the corresponding optimizer name, (e.g., `sampling_`, `cross_entropy_`, `gradient_`, `ilqg_`).
 
 It is also possible to create GUI elements for  parameters that are passed to the residual function.  These are specified by the prefix `residual_`, when the suffix will be the display name of the slider:
 
@@ -280,12 +280,15 @@ Additionally, custom labeled buttons can be added to the GUI by specifying a str
 
 The purpose of `Planner` is to find improved policies using numerical optimization.
 
-This library includes three planners that use different techniques to perform this search:
+This library includes multiple planners that use different techniques to perform this search:
 
 - **Predictive Sampling**
   - random search
   - derivative free
   - spline representation for controls
+- **Cross Entropy Method**
+  - all properties of Predictive Sampling
+  - refits a nominal policy to mean of elite samples instead of using the best
 - **Gradient Descent**
   - requires gradients
   - spline representation for controls

mjpc/CMakeLists.txt

@@ -40,6 +40,8 @@ add_library(
   tasks/acrobot/acrobot.h
   tasks/cartpole/cartpole.cc
   tasks/cartpole/cartpole.h
+  tasks/cube/solve.cc
+  tasks/cube/solve.h
   tasks/fingers/fingers.cc
   tasks/fingers/fingers.h
   tasks/hand/hand.cc
@@ -54,6 +56,8 @@ add_library(
   tasks/manipulation/common.h
   tasks/manipulation/manipulation.cc
   tasks/manipulation/manipulation.h
+  tasks/op3/stand.cc
+  tasks/op3/stand.h
   tasks/panda/panda.cc
   tasks/panda/panda.h
   tasks/particle/particle.cc
@@ -75,6 +79,8 @@ add_library(
   planners/cost_derivatives.h
   planners/model_derivatives.cc
   planners/model_derivatives.h
+  planners/cross_entropy/planner.cc
+  planners/cross_entropy/planner.h
   planners/robust/robust_planner.cc
   planners/robust/robust_planner.h
   planners/sampling/planner.cc

mjpc/grpc/CMakeLists.txt

@@ -37,7 +37,7 @@ findorfetch(
 
 find_package(ZLIB REQUIRED)
 set(gRPC_ZLIB_PROVIDER "package" CACHE INTERNAL "")
-
+set(ZLIB_BUILD_EXAMPLES OFF)
 set(_PROTOBUF_LIBPROTOBUF libprotobuf)
 set(_REFLECTION grpc++_reflection)
 set(_PROTOBUF_PROTOC $<TARGET_FILE:protoc>)

mjpc/main.cc

@@ -47,6 +47,6 @@ int main(int argc, char** argv) {
 #endif
   absl::ParseCommandLine(argc, argv);
 
-  mjpc::StartApp(mjpc::GetTasks(), 6);  // start with manipulation bring
+  mjpc::StartApp(mjpc::GetTasks(), 14);  // start with quadruped flat
   return 0;
 }