// the find
dvalenciar/robotic_arm_environment
Doosan robotic arm, simulation, control, visualization in Gazebo and ROS2 for Reinforcement Learning.
A ROS2 + Gazebo simulation of a 6-DoF Doosan arm (a0912/m1013) wrapped as a reacher-task RL environment: the arm chases a randomly-respawning target sphere. Aimed at people who want a ROS2-native sim to bolt RL algorithms onto, not at people who want a polished Gym environment out of the box.
The repo was just migrated off Foxy/Gazebo Classic (both EOL) onto Jazzy/Gazebo Harmonic, which is more maintenance than most hobby robotics repos bother with. The three-package split (doosan description/control, sphere target, RL glue) is clean, and each piece launches independently for debugging — you can spin up just the arm in RViz or just the sphere before wiring the full loop. It uses ros2_control/gz_ros2_control for joint control instead of a bespoke topic-publishing hack, which is the correct way to drive the arm in Gazebo.
There's no Gymnasium/Gym API wrapper — "custom RL environment" means you write your own adapter before stable-baselines3 or anything else can touch it. The only example given is running random actions; there's no actual trained policy, reward curve, or evidence the reacher task is solvable in a reasonable number of steps. The license is a non-standard "non-commercial use only" clause with no SPDX identifier, which will make some teams avoid it outright. The three ament test files per package are boilerplate copyright/flake8/pep257 linting checks, not tests of the simulation or environment logic, so there's no safety net if the migration broke something subtle in the physics or control loop.