Source-linked AI summary
robosuite: A Modular Simulation Framework and Benchmark for Robot Learning
Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martín-Martín, Abhishek Joshi, Kevin Lin, Abhiram Maddukuri, Soroush Nasiriany, Yifeng Zhu
TL;DR
Robot-learning research needs accessible, reproducible simulation and standardized evaluation because hardware access and reproducibility remain limited. robosuite addresses this with a modular MuJoCo framework combining procedural task construction, composite controllers, and benchmark environments. In its reported experiments, SAC solves three of nine environments, while operational-space controllers enable faster learning than joint-velocity controllers on the selected tasks.
Problem
Limited reproducibility and accessibility of robot hardware constrain research using data-driven robotics algorithms.
Method
robosuite combines modular MuJoCo task construction, standardized environments, high-quality controllers, and benchmarking experiments for robot learning.
Results
SAC solves three of nine environments, while both tested robots learn the selected tasks faster with operational-space than joint-velocity controllers.
Takeaways & Limitations
robosuite provides standardized tasks and interfaces intended to support rigorous, reproducible, and benchmarkable robot-learning research.
Abstract
from arXiv · showhide
robosuite is a simulation framework for robot learning powered by the MuJoCo physics engine. It offers a modular design for creating robotic tasks as well as a suite of benchmark environments for reproducible research. This paper discusses the key system modules and the benchmark environments of our new release robosuite v1.5.
1 Introduction
robosuite is a MuJoCo-powered, modular simulation framework and benchmark designed to support accessible, reproducible research in robot learning. Its v1.5 release combines procedural environment generation, controllers, multimodal sensing, demonstrations, and standardized tasks.
- Motivation and framework: robosuite uses MuJoCo to provide a modular framework and benchmark for data-driven robot learning research.The project aims to lower barriers to research at the intersection of AI and robotics.
- Modularity and procedural generation: Its modular design supports creating new simulation environments and tasks from combinations of robot models, arenas, and parameterized 3D objects.Procedural-generation APIs provide programmatic construction of new environments and tasks.
- Controllers: robosuite v1.5 includes composite controllers that coordinate body-part controllers across joint-space and Cartesian-space position, velocity, torque, inverse-kinematics, and operational-space control.The controller system is intended to provide realistic, flexible robot control interfaces.
- Benchmarks: The release provides standardized tasks of varied diversity and complexity together with benchmarking results for reproducible research.The manuscript presents the framework design, v1.0 benchmark tasks, and evaluations of relevant data-driven algorithms.
- Sensors and observations: The framework provides heterogeneous sensory signals, including physical states, RGB cameras, depth maps, segmentation masks, and proprioception.These modalities support diverse observation configurations for robot-learning experiments.
- Human demonstrations: Utilities support collecting human demonstrations with keyboards, 3D mice, and GUI cursor devices, then replaying and using those datasets for learning.Demonstration tools extend the framework beyond reinforcement-learning-only workflows.
2 System Modules
robosuite separates environment modeling from simulation interaction, combining modular task components, configurable robots, and controllers that translate actions into torques. Its Robot abstractions support diverse models and configurations while exposing state and episode-level behavior.
- System design: Modeling APIs construct simulation environments from robots, objects, and arenas, while Simulation APIs expose actions, observations, and rewards.Tasks are assembled into MuJoCo MJCF models and instantiated as runtime simulations.
- Environments: Environments provide Gym-style task interfaces with configurable rendering, observations, control frequency, and episode horizon.Each environment corresponds to a robotic task and accepts external action commands.
- Environments: Task instances combine an arena, robot models, and object models, with placement initialization sampling valid non-colliding episode starts.The arena defines the workspace, while objects may be loaded from assets or generated procedurally.
- Robots: Robot classes support single-armed, bimanual, and mobile-manipulation embodiments through robot, gripper, base, and controller components.Robot instances encapsulate relevant properties and directly set their initial state and control torques.
- Robots: The framework provides 10 commercially available robot models, including Panda, Sawyer, and IIWA, with plug-and-play combinations of robots, grippers, bases, and controllers.Multiple identical robot instances can be instantiated using unique IDs.
- Controllers: Composite controllers convert high-level actions into body-part-specific joint torques across operational-space and joint-space control modes.Starting in robosuite v1.5, each body-part controller translates supplied actions into torques for its assigned robot part.
3 Benchmark Environments
robosuite v1.5 provides nine standardized manipulation environments spanning single-arm and coordinated two-arm tasks, with randomized initial conditions and SAC benchmarking. The benchmarks show that controller choice affects learning efficiency, while SAC solves three environments and progresses slowly on the remainder.
- Environment Suite: Nine standardized environments cover lifting, stacking, pick-and-place, nut assembly, door opening, table wiping, and three coordinated two-arm tasks.The two-arm tasks are Handover, Lifting, and Peg-In-Hole; task objects, locations, or configurations are randomized at episode start.
- Benchmarking Setup: SAC benchmarks combine proprioceptive and object-specific observations across Panda and Sawyer robots with OSC POSE and JOINT VELOCITY controllers.Agents use low-dimensional physical states and standardized training settings across experiments.
- Benchmarking Setup: 500 epochs of 500-step episodes provide standardized training curves, with normalized rewards capped at 500 per episode.The experiments use identical algorithm hyperparameters and report evaluation-return means over training.
- Controller Ablation: Both robots solve Block Lifting and Door Opening faster with OSC POSE than with JOINT VELOCITY controllers.The authors hypothesize that operational-space control accelerates exploration in task space.
- Benchmarking Results: SAC solves three of nine environments—Block Lifting, Door Opening, and Two Arm Peg-In-Hole—while making slow progress on the others.The paper recommends Panda with OSC for standardized and fair comparisons across the nine environments.
4 Conclusion
robosuite provides a MuJoCo-based simulation framework and benchmark suite for robot-learning research, supporting standardized evaluation and reproducible research. Since its 2017 debut, the framework has been used across reinforcement learning, imitation learning, and sim-to-real transfer.
- Conclusion: robosuite combines a MuJoCo-based simulation framework with standardized tasks for rigorous evaluation and reproducible robot-learning research.The framework supports research and development of robot-learning solutions.
- Conclusion: Since 2017, robosuite has been used for reinforcement learning, imitation learning, and sim-to-real transfer.