What it does Roboschool was OpenAI’s open-source physics simulator for reinforcement learning, offering drop-in replacements for standard Gym MuJoCo environments (Ant, HalfCheetah, Humanoid, etc.) plus extras like a Pong-playing robot and flag-capture humanoids. It wrapped a custom Bullet physics build with Qt5 rendering and shipped with pre-trained “agent zoo” demos you could run out of the box.

The interesting bit The project’s real goal wasn’t replication—it was escape velocity from MuJoCo’s trajectory-centric, fragile tasks. The roadmap sketched out multiplayer games, vision-and-joint combined observations, even verbal command following. That none of this materialized is almost more telling than what did; it’s a snapshot of where RL robotics research thought it was headed in 2017.

Key highlights

• 13 environments including classic control (inverted pendulum), locomotion (Hopper, Walker, Ant), and humanoid flag-run tasks
• Custom Bullet physics build with Qt5/OpenGL rendering, requiring boost-python3 and assorted system dependencies
• Pre-trained agent zoo with runnable demos (humanoid flag-run, three-agent race)
• Explicitly deprecated September 2019; OpenAI recommends PyBullet Gym environments instead
• Code archived as-is, no updates expected

Caveats

• Installation from source is genuinely involved: custom Bullet build, boost-python3 wrangling, PKG_CONFIG_PATH debugging, potential -fPIC or NVIDIA hardware rendering workarounds
• Python 3.5/3.6 only; Python 2.7 support described as “non-trivial amount of work”
• No candidate images available in the repository

Verdict Worth a look if you’re studying the history of open-source RL tooling or need to reproduce a 2017-era paper. Everyone else should follow OpenAI’s own advice and use PyBullet. This is a time capsule, not a foundation.