Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

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

Seminal September 2020

v1.5 release reel: teleoperation interfaces and controller options on simulated humanoid and mobile robots.

TL;DR A modular MuJoCo framework that builds robot-learning tasks from interchangeable robots, grippers, controllers and objects, with nine standardized manipulation benchmarks. Soft Actor-Critic baselines solve three of the nine tasks and learn faster with operational-space control than with joint-velocity control.

What they do

robosuite is a simulation framework and benchmark for robot learning built on MuJoCo. It grew out of the SURREAL distributed-RL project and is now part of the ARISE Initiative. This version of the paper describes the v1.5 release: 10 robot models (including the GR1 humanoid, Spot and TIAGo), 9 grippers (including the Inspire dexterous hand), 4 bases, nine standardized manipulation tasks, and tools for collecting human demonstrations.

How it works

  • Modeling APIs. A Task merges one or more robots (each with gripper and base models), an arena and any objects into a single MJCF model for MuJoCo. Objects are loaded from XML or composed procedurally from primitive geoms, and a placement initializer samples non-colliding start states each episode.
  • Simulation APIs. Environments expose a Gym-style interface. Sensors return proprioception, RGB, depth, segmentation masks, force-torque readings and the task reward.
  • Controllers. A composite controller either splits the action vector across body-part controllers (OSC_POSE, OSC_POSITION, JOINT_POSITION, JOINT_VELOCITY, JOINT_TORQUE) or runs whole-body IK on end-effector targets. OSC and joint-position actions can also include stiffness, or stiffness and damping.
  • Devices. Keyboard, SpaceMouse and MuJoCo GUI teleoperation for collecting and replaying demonstrations.

Results

  • The nine tasks: block lifting, block stacking, pick-and-place, nut assembly, door opening, table wiping, two-arm lifting, two-arm peg-in-hole and two-arm handover.
  • SAC baselines on low-dimensional state (500 epochs, 500-step episodes, five seeds, about two days per run on 2 CPUs) solve three of the nine tasks: block lifting, door opening and two-arm peg-in-hole.
  • On block lifting and door opening, both Panda and Sawyer learn faster with OSC_POSE than with JOINT_VELOCITY. The authors recommend Panda with OSC as the standard benchmark setting.

Limitations

The paper has no limitations section. Its benchmark covers one algorithm (SAC) on state observations only. SAC makes slow progress on the other six tasks, which the authors say need better exploration over longer horizons. All results are in simulation.

Figures

From Figure 1: Robot environments procedurally generated with the robosuite APIs.
From Figure 1: Robot environments procedurally generated with the robosuite APIs.
From Figure 2: System diagram. A policy or a human with an I/O device sends actions to the environment; the robot's c...
From Figure 2: System diagram. A policy or a human with an I/O device sends actions to the environment; the robot's controller turns them into torques that MuJoCo executes, and sensors return observations, reward and metadata. The environment is instantiated from a task made of a robot model, an arena and object models.
From Figure 3: The Robot module. A robot is initialized with robot, base and gripper models and a controller, turns a...
From Figure 3: The Robot module. A robot is initialized with robot, base and gripper models and a controller, turns actions into torques at runtime, and exposes its specifications, proprioception and sensing at any point in the simulation.
Abstract

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.