This repository contains a number of challenging environments based on OpenAI's Gym hand manipulation tasks, all of which are challenging for current RL and trajectory optimisation techniques. The GIFs of example solutions for some of the environments come from a soon-to-be published custom trajectory optimisation technique I have developed.
Requirements: Mujoco-py and Gym
Once these are installed, run pip install dexterous-gym.
Alternatively, clone the repository and run pip install -e . from within the main directory.
import gym
import dexterous_gym
env = gym.make("EggCatchUnderarm-v0")
See dexterous_gym/__init__.py for full list of registered environments. See dexterous_gym/examples/test_all_envs.py to run all available environments with random actions.
Variants: "PenSpin-v0".
This is a simple modification of the OpenAI gym HandPen environment where all we do is change the observation space to remove any notion of a goal and then define a custom reward function. This reward function encourages the pen to be spun whilst remaining horizontal.
A majority of the environments are goal-based, and have a similar API to the openAI Gym manipulation environments (observations are dictionaries with "observation", "achieved_goal", "desired_goal"). In the case of the two object environments "achieved_goal" and "desired_goal" are also dictionaries with entries "object_1" and "object_2". All environments come with a standard dense-reward setting (where the reward is defined in terms of the positional and rotational distance between the achieved goal(s) and the desired goal(s) as well as a sparse setting, where the reward is -1.0 until the goal(s) are achieved (at which point the reward is 0.0).
These environments involve two fixed-position hands. The hand which starts with the object must find a way to hand it over to the second hand, since the goals are chosen such that only the second hand can move the object to the target position/orientation. The object can be an egg, a block or a pen.
These environments again have two hands, however now they have some additional degrees of freedom that allows them to translate/rotate their centre of masses within some constrained region. The aims of these tasks is to throw the object from one hand to the other, which can then move it to the desired position/rotation. The "Hard" versions involve a greater distance between the two hands, such that the objects must be thrown further. Again the object can be an egg, a block or a pen.
Similar to the HandCatchUnderArm environments but now the two hands are upright, and so the throwing/catching technique that has to be employed is different. TODO: add "harder" (i.e. further separation between hands) version of these environments, and run some tests with my trajectory optimisation algorithm to get some example trajectories.
These environments involve coordination between the two hands so as to throw the two objects between hands (i.e. swapping them). This is necessary since each object's goal can only be reached by the other hand. TODO: introduce "harder" and overarm versions of these environments. Also, note that the examples shown were generated when there was an error with the environment - need to re-run experiment now this is fixed.