Deep Reinforcement Learning for Robotic Grasping from Octrees
The second major release adds new GraspPlanetary environments alongside a new mobile manipulator and several models. The control was redesigned and fully integrated with ros2_control while supporting a new control technique via MoveIt 2 Servo. Furthermore, most modules and the Docker setup have undergone major refactoring to improve the overall quality of the code.
See [email protected] for detailed changes.
A minor release with some small additions. This release mainly focused on providing compatibility with latest releases of all dependencies, while also simplifying few things. No local changed were done to the environments/algorithms/agents that would affect the functionality (although updated dependencies might influence some of the quantitative results, e.g. changes within the physics engine).
See [email protected] for detailed changes.
This version corresponds to everything described in my Master's Thesis. At this stage, the source code also contains many variants of different approaches that I have tried over the duration of the project, with a large degree of configurability. Many of these were quick-and-dirty fixes and some approaches do not bring any improvements, so the source code is not very clean.
If I have some time for it, there might be a 2.0.0 release that refactors and cleans everything up, splits the project into several modules and improves generalisation to different robots. Re-write of several parts in a lower-level language (Rust or C++) would also be considered. This epic will be tracked in https://github.com/AndrejOrsula/drl_grasping/issues/85.
Sim2Real (runtime and interfaces with controllers) is not included with this release. It is kept in sim2real branch to serve as an inspiration in case anyone needs it. It is not very safe, so I believe that everyone should just code their own version that they 100% understand if they ever want to apply it on a real robot that could potentially damage something (or someone :confused:).