// the take
robosuite is a MuJoCo-based simulation framework for robot manipulation research, built by Stanford SVL and now also used by UT RPL and NVIDIA's GEAR group. It gives RL and imitation-learning researchers standardized benchmark tasks (Lift, Stack, NutAssembly, TwoArmHandover, etc.) plus modular APIs for composing robots, controllers, and environments, rather than a production robotics stack.
v1.4 migrated off the old mujoco-py to DeepMind's official MuJoCo Python bindings, removing a dependency that caused a lot of install headaches for years. The controller layer (joint velocity/torque, IK, OSC, whole-body) is decoupled from the robot model, so swapping a Panda for a humanoid GR1 doesn't mean rewriting the control code. It ships actual benchmarking results tied to standardized tasks instead of just handing you an empty Gym-style shell, so algorithm comparisons are apples-to-apples. Demonstration collection and replay plug directly into robomimic, so the teleop-to-training data pipeline already exists instead of being left to the user.
The real dependency surface is heavier than the README implies once you add MuJoCo plus the optional photorealistic rendering (ray tracing, Isaac Sim integration) — this is not a quick pip install. The repo's directory tree is dominated by per-robot mesh assets (STL/DAE/OBJ files), which bloats clone size and makes forking or vendoring awkward. All benchmark numbers are sim-only; sim-to-real transfer gets a docs link but no evidence in the README itself, so generalization to real hardware is unverified here. The jump from v1.1 through v1.5 rewrote core infrastructure repeatedly (backend swap, composite controllers, whole-body control), which is a sign the API isn't stable if you're building something meant to outlive a single version.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →