// the find
AndrejOrsula/drl_grasping
Deep Reinforcement Learning for Robotic Grasping from Octrees
A research codebase for training robot arms to grasp objects using deep RL with octree-based 3D observations. The core claim is that octrees are more data-efficient than raw point clouds or depth images for sim-to-real transfer. This is a master's thesis project from 2022, backed by an IROS paper, aimed at robotics researchers who want to reproduce or extend the work.
The octree CNN feature extractor is the genuinely interesting part — using O-CNN for hardware-accelerated 3D convolutions on sparse voxel structures is a sound choice that avoids the quadratic scaling of dense 3D convolutions. Domain randomization coverage is thorough: geometry scale, surface friction, inertial properties, camera pose, textures, and lighting are all randomized per-episode, which explains the zero-shot sim-to-real transfer results. The Docker setup ships with prebuilt 25GB images including datasets, which is the right call for a project this dependency-heavy — local install is a nightmare of ROS 2 + Gazebo + custom forks. The curriculum learning implementation for grasp difficulty is sensible and incrementally increases task hardness based on success rate.
No parallel environments is a fundamental throughput bottleneck — SAC/TQC need millions of steps and a single Gazebo instance collecting them serially means training takes days on decent hardware. The pretrained agents submodule is explicitly broken as of v2.0.0 and commented out in the README, so you can't evaluate pre-trained models without digging through old releases. The dependency chain (ROS 2 Galactic, Gazebo Fortress, a fork of gym-ignition, a fork of O-CNN) means you're locked to Ubuntu 20.04 and anything touching these forks will need manual maintenance — Galactic reached EOL in November 2022, two months before the last commit. DreamerV2 support is marked experimental and restricted to RGB observations, making the model-based path a dead end for the repo's main octree contribution.