PaperlayerView exact revision on arXiv

Source-linked AI summary

PyRobot: An Open-source Robotics Framework for Research and Benchmarking

Adithyavairavan Murali, Tao Chen, Kalyan Vasudev Alwala, Dhiraj Gandhi, Lerrel Pinto, Saurabh Gupta, Abhinav Gupta

arXiv:1906.08236v1cs.ROcs.AIcs.CVcs.LG

TL;DR

Robotics research remains difficult to enter because researchers face hardware costs and software expertise barriers, while high-level code is often hardware-dependent. PyRobot wraps ROS in a shared, hardware-independent API for robot control. The paper reports its use with 10 LoCoBots in a university robotics course and presents it as a foundation for a broader research and teaching ecosystem.

  • Problem

    Robotics remains hard for new researchers because hardware setup entails cost and development time, and high-level applications lack a common hardware-independent API.

    Source: §1 · p. 1

  • Method

    PyRobot provides a lightweight, high-level interface on ROS with hardware-independent APIs for controlling different robots.

    Source: §1 · p. 2, §2 · p. 2

  • Results

    10 LoCoBots were used with PyRobot for homework assignments and projects in CMU's Spring 2019 Robot Autonomy course.

    Source: §7 · p. 8

  • Takeaways & Limitations

    The authors envision PyRobot's unified API and LoCoBot's standard hardware supporting shared algorithms, models, and real-robot datasets, and faster collaboration on robotics applications.

    Source: §7 · p. 8

    The pushing demonstration depends on stereo-depth quality across backgrounds; the authors recommend a floor with non-uniform texture for best performance.

    Source: §5.5 · p. 6

Abstract

from arXiv · show

This paper introduces PyRobot, an open-source robotics framework for research and benchmarking. PyRobot is a light-weight, high-level interface on top of ROS that provides a consistent set of hardware independent mid-level APIs to control different robots. PyRobot abstracts away details about low-level controllers and inter-process communication, and allows non-robotics researchers (ML, CV researchers) to focus on building high-level AI applications. PyRobot aims to provide a research ecosystem with convenient access to robotics datasets, algorithm implementations and models that can be used to quickly create a state-of-the-art baseline. We believe PyRobot, when paired up with low-cost robot platforms such as LoCoBot, will reduce the entry barrier into robotics, and democratize robotics. PyRobot is open-source, and can be accessed via https://pyrobot.org.

1. Introduction

Shared platforms and datasets have helped lower entry barriers and standardize evaluation in AI, while robotics still lacks an accessible research platform and common API. PyRobot is proposed as a high-level, hardware-independent ROS interface to address these barriers.

1. Introduction

  • Robotics remains difficult for new researchers because hardware setup carries financial and development costs, and low-cost robots alone have not supplied a research ecosystem.The paper notes that robot costs have become comparable to a GPU, yet entry remains difficult without research platforms and a self-sustaining ecosystem.Source: §1 · p. 1
  • ROS and MoveIt! require broad expertise across planners, kinematics, and low-level controllers, motivating a lightweight interface for AI practitioners, students, and hobbyists.The paper argues that many new users lack the expertise or time to learn the full stack.Source: §1 · p. 1
  • High-level robotics code remains hardware-dependent, and the community lacks a common platform and API for sharing code, datasets, and models.URDF partly encapsulates hardware details for downstream ROS services, but the paper says this does not solve hardware dependence for high-level AI applications.Source: §1 · pp. 1–2
  • PyRobot provides a lightweight, high-level ROS interface with hardware-independent mid-level APIs and manipulation and navigation examples.The paper also describes libraries for hand-eye calibration, tele-operation, trajectory tracking, and SLAM-based navigation.Source: §1 · p. 2

2. PyRobot Framework

PyRobot wraps ROS in a Python framework with shared robot-control APIs, aiming to make robotics accessible to beginners while preserving access to lower-level components. Its design supports common commands across robots and open collaboration around hardware and software.

2. PyRobot Framework

  • PyRobot wraps ROS mid-level features and low-level C++/C controllers and drivers, while exposing shared utilities for control, kinematics, planning, and visual SLAM.Users can still access different hierarchy levels, including low-level velocities and torques that bypass a planner.Source: §2 · p. 2
  • Beginners can start commanding a robot in a few lines of code without learning ROS or the underlying software and firmware stack.The framework is intended to lower the learning curve for new users.Source: §2 · p. 2
  • The Sawyer can be commanded in the same manner as LoCoBot, although PyRobot uses robot-specific configuration files for parameters such as joint names, ROS topics, and frames.Each robot configuration also specifies settings such as planner configuration and inverse-kinematics tolerance.Source: §2 · p. 2
  • Affordable, assemblable LoCoBot hardware paired with open software is presented as a basis for researchers to collaborate across system layers and share real-robot applications.The paper contrasts real-robot software goals with simulation's usefulness for testing and experiments.Source: §2 · p. 3

3. Supported Hardware and Simulators

PyRobot supports LoCoBot and Sawyer robots, as well as Gazebo simulation. The described hardware spans a mobile manipulator and a collaborative arm, with LoCoBot code also runnable in Gazebo.

3. Supported Hardware and Simulators

  • LoCoBot is a low-cost mobile manipulator with a five-DOF arm, 0.2 kg working payload, and 0.55 m maximum reach.The platform pairs the arm with either a Kobuki or Create2 base; its Realsense D435 camera is mounted on a pan-tilt attachment.Source: §3 · p. 3
  • PyRobot interfaces with the Sawyer, a seven-DOF collaborative arm from Rethink Robotics, through its Intera SDK.Source: §3 · p. 3
  • PyRobot supports Gazebo, and LoCoBot and LoCoBot-Lite can run the same code in Gazebo and on the real robot.Source: §3 · p. 3

4. PyRobot Controllers

PyRobot implements basic differential-drive controllers and interfaces with MoveIt! and Movebase; the evaluation reports base-control performance and LoCoBot arm repeatability.

4.1. Accuracy of Base Control

  • Base controllers comprise sampled-sequence DWA, proportional rotation–translation segments, and trajectory-tracking LQR built around bicycle-model dynamics.Trials measured commanded-versus-achieved translation and heading errors using Vicon across linear, rotational, and combined motions.Source: §4.1 · pp. 3–4
  • LoCoBot generally has lower base-control errors than LoCoBot-Lite, LQR and proportional control generally outperform DWA, and LQR is more effective at closing the loop on odometry.Table 1 reports translation and rotation errors across motion types and controllers.Source: §4.1 · pp. 4–5, Table 1 · p. 4

4.2. Repeatability Tests for Manipulator

  • Arm repeatability was 0.33–0.58 mm across tested poses, with extended poses less repeatable than torso-near poses and z-axis variability higher under gravity.Tests used 10 repetitions per pose without payload; Sawyer and UR5 are reported at 0.1 mm.Source: §4.2 · p. 5, Table 2 · p. 5

5. High-Level AI Applications

The section demonstrates PyRobot across SLAM-based navigation and learned or heuristic applications in navigation, grasping, and pushing. These examples combine robot sensing with PyRobot interfaces to execute tasks.

5. High-Level AI Applications

  • The section presents several high-level AI applications implemented through the PyRobot API.Source: §5 · p. 5

5.1. Visual SLAM

  • ORB-SLAM2 supplies camera pose and trajectory for robot state estimation, while depth observations support dense-map reconstruction for navigation planning.Source: §5.1 · p. 5

5.2. Navigation via SLAM and Path Planning

  • Movebase uses a visual-SLAM occupancy map to build a safety cost-map, plan collision-free paths, and replan when newly observed areas make a path infeasible.Source: §5.2 · p. 6

5.3. Learned Visual Navigation

  • CMP maps the current camera image and goal location to one of four macro-actions, which PyRobot executes through its base position-control interface.Source: §5.3 · p. 6

5.4. Grasping

  • The grasping pipeline converts an image-space grasp and gripper orientation into a 3D target using camera parameters and depth, then approaches and closes the gripper.The model was trained on data from people’s homes and is described as robust to varied objects and backgrounds.Source: §5.4 · p. 6

5.5. Pushing

  • The pushing algorithm filters and clusters an RGB-D point cloud, randomly selects a cluster and push-start point, then pushes toward the cluster center.Its depth-sensor dependence makes pushing quality sensitive to background; non-uniform floor texture is recommended for best performance.Source: §5.5 · p. 6

6. Related Work

Related work covers layered robotics software, low-cost robot learning, and open-source manipulators. It also describes the challenges of benchmarking robotics and the role of shared research resources.

Robotics Software Design

  • Robotics software includes hardware-agnostic motion-planning libraries and ecosystems for kinematics, distributed processes, and real-time robot control.Source: §6 · p. 6

Low-cost Mobile Manipulators

  • Prior low-cost robot learning examples include block stacking with a cheap inaccurate 6-DOF robot, visual grasping on an earlier LoCoBot, and a compliant low-cost arm.Source: §6 · pp. 6–7

Open Source Manipulators

  • Open-manipulator efforts include Raven, an open-architecture surgical research robot, and Robotis’s project for building low-cost robots with custom kinematics and design.Source: §6 · p. 7

Benchmarking in Robotics; Research Ecosystems in AI Fields

  • Robotics benchmarking remains difficult across varied hardware, objects, and environments, while AI fields have benefited from common tasks, datasets, and code that support replication and building on prior work.The paper notes that robotics has task-specific metrics and object datasets but no benchmarking consensus.Source: §6 · p. 7

7. Discussion

The discussion presents PyRobot as a high-level, hardware-independent framework and argues that pairing it with LoCoBot can support research and teaching. It describes a course deployment and a shared research ecosystem as examples of that potential.

Discussion

  • PyRobot provides a high-level, hardware-independent API, and the authors expect its pairing with LoCoBot to reduce robotics entry barriers and support a research and teaching ecosystem.Source: §7 · pp. 7–8

PyRobot for robotics instruction

  • Ten LoCoBots were used for homework and projects in CMU’s Spring 2019 Robot Autonomy course.Source: §7 · p. 8

PyRobot as a research ecosystem

  • The authors expect PyRobot’s unified API and LoCoBot’s standard hardware to help researchers share algorithms, models, and real-robot datasets, collaborate, and iterate faster.Source: §7 · p. 8

A. Code Listings

The code listings show PyRobot API examples for visual navigation, grasping, and pushing. They illustrate robot setup, perception inputs, and successive commands for base or arm motion.

A. Code Listings

  • The listings import PyRobot's Robot interface and show camera, gripper, and arm setup commands alongside the task-specific control calls.The examples include closing the gripper and setting camera pan and tilt.Source: A. Code Listings · pp. 11–12
  • The visual-navigation example sets a policy goal, maps each selected action to a relative position, and commands the base to execute it.The listed actions include forward motion and rotations by ±dw.Source: A. Code Listings · p. 11
  • The grasping example converts an image-space grasp into a robot-workspace pose, moves through pre-grasp and grasp poses, then closes the gripper.It also sets pre-grasp and grasp heights and includes arm and camera reset positions.Source: A. Code Listings · p. 11
  • The pushing example obtains a point cloud, computes a pre-push point and push point, then moves the gripper down and horizontally to push the object.The horizontal displacement is defined as twice the vector from the push point to the object center.Source: A. Code Listings · p. 12

Related papers

Loading 1906.08236v1…