GzDRL
GzDRL is an open-source reinforcement learning framework for robotics that supports training directly within the Gazebo simulation environment. It employs a C++ backend derived from EnvPool to implement batched parallel environments with GIL release, balancing high-throughput training with compatibility with the Gazebo and ROS ecosystems.
Timeline (last 90 days)
GzDRL is an open-source framework that attempts to complete reinforcement learning training directly inside Gazebo.
GzDRL uses an explicit 'action → physics → observation' execution sequence, turning simulation stepping into a deterministic and controllable process rather than relying on asynchronous message passing.
GzDRL's performance foundation comes from a C++ backend derived from EnvPool.
GzDRL supports batched parallel environments, allowing environments to be created in Python while releasing the GIL during execution.
GzDRL supports pure CPU training, enabling reinforcement learning experiments without a large NVIDIA GPU.
GzDRL offers researchers features such as reproducible state transitions, batch vectorization, multi-agent environments, and domain randomization.
GzDRL achieved a throughput of 78.6k environment steps per second on a single workstation CPU.
The author conducted five independent PPO training runs using the same random seed, producing completely identical learning curves with 100% matching policy checkpoint hash values.
The author directly deployed a trajectory-tracking policy trained in simulation to a physical quadrotor without any fine-tuning.
All Facts (9)
GzDRL is an open-source framework that attempts to complete reinforcement learning training directly inside Gazebo.
50%UnverifiedGzDRL uses an explicit 'action → physics → observation' execution sequence, turning simulation stepping into a deterministic and controllable process rather than relying on asynchronous message passing.
50%UnverifiedGzDRL's performance foundation comes from a C++ backend derived from EnvPool.
50%UnverifiedGzDRL supports batched parallel environments, allowing environments to be created in Python while releasing the GIL during execution.
50%UnverifiedGzDRL supports pure CPU training, enabling reinforcement learning experiments without a large NVIDIA GPU.
50%UnverifiedGzDRL offers researchers features such as reproducible state transitions, batch vectorization, multi-agent environments, and domain randomization.
50%UnverifiedGzDRL achieved a throughput of 78.6k environment steps per second on a single workstation CPU.
50%UnverifiedThe author conducted five independent PPO training runs using the same random seed, producing completely identical learning curves with 100% matching policy checkpoint hash values.
50%UnverifiedThe author directly deployed a trajectory-tracking policy trained in simulation to a physical quadrotor without any fine-tuning.
50%