performance_guarantees

This repository accompanies the paper "Guarantees on Robot System Performance Using Stochastic Simulation Rollouts"

Summary of Repository

• experiments.py: Each figure in the paper is associated with an experiment. We use this file to run each experiment and save the data necessary for plotting the results. Each experiment requires a .yaml config file in order to run. A config file for each experiment is found in the experiments/config_files directory. When an experiment is run the results as well as the associated config file are saved in a timestamped folder in the experiments/runs directory.

• plot_experiments.py: This file is used to plot the results of each experiment to generate the figures in the paper.

• task.py: This file defines the Task class which is a wrapper class providing basic simulation functionality such as environment creation and random action sequence sampling.

• traj_utils.py: This file provides functions for (i) rolling out a trajectory, (ii) evaluating multiple plans in parallel on a Gymnasium environment.

• traj_utils_hand.py: This file is similar to traj_utils but with functions specifically designed for use with shadow hand environment. Gymnasium handles the shadow hand environment slightly differently than it does the other MuJoCo environments. Ideally this will be merged with traj_utils.py.

• bound_utils.py: This file contains functions for bounding various measures of risk, chance constraint hypothesis testing, and baseline code for assessing and visualizing bound validity.

• analyze_bounds.py: This file provides the function for the experiment that assesses validity of the risk bounds when considering a single plan.

• analyze_chance.py: This file provides a function for the experiment that assesses validity of the chance constraint hypothesis testing framework.

• analyze_fixed_multihyp.py: This file provides a function for the experiment that assesses validity and necessity of the multi-hypothesis correction given a set of candidate plans.

• cem.py: This file defines the CEM class which implements the cross entropy sampling-based solver.

• spline_hand.py: This file defines the Spline class which implements the spline sampling-based solver from MJPC for the shadow hand.

• generate_multihyp_plans.py: This file is used for generating plans for the shadow hand using the spline solver. These plans are then used for testing the multi-hypothesis corrected bound.

• vis_rollouts.py: This file is used to visualize rollouts of the plans generated for the shadow hand.

• dist_cartoon.py: Used to generate a visualization of the different risk measures on an example distribution.

Our Changes

In this section we detail the changes we made to either the gymnasium or the gymnasium_robotics python libraries. Most changes were made in order to enable (i) domain randomization or (ii) multi-processing (not just multi-threading).

In /gymnasium_robotics/envs/shadow_dexterous_hand/manipulate.py we made the following changes:

1. in __init__() for BaseManipulateEnv class, we set the following variables

   randomize_initial_position=False
   randomize_initial_rotation=False
   

2. In the MujocoManipulateEnv class we redefined the method _sample_goal(self) as

   def _sample_goal(self):
   	goal = np.concatenate([[1, 0.87, 0.2], [1, 0, 0, 0]])
   	return goal
   

In /gymnasium_robotics/envs/shadow_dexterous_hand/manipulate_touch_sensors.py we made the following changes:

1. in __init__() for MujocoManipulateTouchSensorsEnv class, we set the following variables

   randomize_initial_position=False
   randomize_initial_rotation=False
   

In /gymnasium_robotics/__init__.py we registered a new environment HandManipulateBlockEggPen_BooleanTouchSensors-v1. This environment is defined in the file gymnasium_robotics\envs\shadow_dexterous_hand\manipulate_block_egg_pen_touch_sensors.py. This file merely randomizes which object .xml file is used in the model_path argument of the class initialization.

To randomize friction and density we created the files

• /gymnasium_robotics/envs/assets/hand/manipulate_block_touch_sensors_rand.xml
• /gymnasium_robotics/envs/assets/hand/manipulate_egg_touch_sensors_rand.xml
• /gymnasium_robotics/envs/assets/hand/manipulate_pen_touch_sensors_rand.xml

The default values are Density=1000, friction=“1 0.005 0.0001”. More information about the MuJoCo .xml files is available here.

In /gymnasium_robotics/envs/shadow_dextrous_hand/manipulate_touch_sensors.py we changed the __init__() function for the class MujocoManipulateTouchSensorsEnv. We set randomize_initial_position=False and randomize_initial_rotation=False.

In /gymnasium_robotics/utils/mujoco_utils.py we added the function robot_set_obs(model, data, joint_names, pos_values, vel_values) which is called by _robot_set_obs(self, obs) in the /gymnasium_robotics/envs/shadow_dexterous_hand/manipulate.pyfile.

In manipulate.py we changed the compute_reward() function to have squared position and orientation terms in a weighted sum.