Small updates in paper [ci skip]

paper/paper.md

@@ -38,21 +38,22 @@ bibliography: paper.bib
 Robot simulation plays a pivotal role in robotics and machine learning research, offering a cost-effective and safe means to develop, validate, and benchmark algorithms in various scenarios. With the growing complexity of robotic systems and the increasing demand for data-driven approaches in machine learning, there is a pressing need for versatile and efficient robot simulators that cater to the diverse requirements of researchers. In response to this demand, we introduce RobotDART, a high-performance and modular robot simulator designed to empower researchers in robotics and machine learning with a powerful and flexible simulation environment.
 
 # Motivation
-Simulators play a crucial role in robotics by providing a safe environment for testing without the risk of damaging real robots. Within the robotics community, existing simulators primarily prioritize seamless deployment on physical robots, which is evident in their design philosophy. One of the best examples is Gazebo \cite{}, which leverages ROS as an abstraction level so that users can seamlessly transition between simulated and real robots. Because of this design choice, Gazebo: (1) is asynchronous, which means that the simulator operates on its own clock and time-step analogous to real-world robots, (2) introduces a significant overhead, including network communication and serialization to ROS messages, and (3) is not designed for running multiple simulations concurrently. At a lower abstraction level, the robotics community has developed various physics libraries such as Mujoco [ref], Bullet [ref], and ODE [ref]. However, these libraries typically lack sensor abstractions, particularly for cameras, and do not come with pre-validated robot models, necessitating users to program their own simulators.
+Simulators play a crucial role in robotics by providing a safe environment for testing without the risk of damaging real robots. Within the robotics community, existing simulators primarily prioritize seamless deployment on physical robots, which is evident in their design philosophy. One of the best examples is Gazebo \cite{}, which leverages ROS as an abstraction level so that users can seamlessly transition between simulated and real robots. Because of this design choice, Gazebo: (1) is asynchronous, which means that the simulator operates on its own clock and time-step analogous to real-world robots, (2) introduces a significant overhead, including network communication and serialization to ROS messages, and (3) is not designed for running multiple simulations concurrently. At a lower abstraction level, the robotics community has developed various physics libraries such as Mujoco \cite{}, Bullet \cite{}, and ODE \cite{}. However, these libraries typically lack sensor abstractions, particularly for cameras, and do not come with pre-validated robot models, necessitating users to program their own simulators.
 
 With the emergence of reinforcement learning for robot control, there arises a need for simulators that can address novel requirements not adequately addressed by current simulators or "low-level" physics libraries. Specifically, simulators designed for reinforcement learning must encompass the following features:
+
 - beyond real-time simulation speed: Given the necessity to execute thousands or even millions of simulations, both controllers and simulators stand to benefit from operating at speeds faster than real-time;
 - overhead-free and fast reset: with the imperative to conduct numerous simulations efficiently, minimizing time introduced by the communication overheads  and the initialization processes becomes paramount;
 - synchronous operation: ensuring reproducible learning runs and effectively managing faster-than-reality simulations necessitates synchronization of simulation and controller time-steps;
 - thread safety: modern computers with their many of cores can be leveraged to execute many simulations in parallel but the simulator needs to be designed in accordance;
 - camera and depth sensor simulation: deep reinforcement learning has demonstrated promising results in learning vision-based policies through end-to-end learning; however, this demands the simulation of both depth and color cameras;
-- flexibility and modularity: the field of machine learning is characterized by rapid development, making it challenging to anticipate the diverse needs of researchers; for instance, techniques like privileged reinforcement learning [ref] and domain randomization [ref] necessitate direct access to various low-level components of the simulator to enable experimentation and innovation.
+- flexibility and modularity: the field of machine learning is characterized by rapid development, making it challenging to anticipate the diverse needs of researchers; for instance, techniques like privileged reinforcement learning \cite{} and domain randomization \cite{} necessitate direct access to various low-level components of the simulator to enable experimentation and innovation.
 
 None of these features correspond to the mainstreams robot simulators like Gazebo. RobotDART aims at filling this gap by providing a flexible simulator for data-driven robotics while being open-source and as simple as possible.
 
 # Design and Implementation
 
-RobotDART is built upon the Dynamic Animation and Robotics Toolkit (DART), a robust open-source physics engine known for its accuracy and efficiency in simulating rigid body dynamics and articulated systems. By leveraging DART's capabilities, RobotDART provides a solid foundation for simulating a wide range of robotic systems with realistic dynamics and interactions with the environment.
+RobotDART is built upon the Dynamic Animation and Robotics Toolkit (DART)\cite{}, a robust open-source physics engine known for its accuracy and efficiency in simulating rigid body dynamics and articulated systems. By leveraging DART's capabilities, RobotDART provides a solid foundation for simulating a wide range of robotic systems with realistic dynamics and interactions with the environment.
 
 One of the key strengths of RobotDART lies in its modular architecture, which allows researchers to easily customize and extend the simulator according to their specific research needs. RobotDART provides a collection of pre-built robot models, sensors, and environments, covering a diverse range of scenarios commonly encountered in robotics research. Researchers can seamlessly integrate their own robot models, sensors, and environments into RobotDART, thanks to its flexible plugin system and intuitive Python API.