Mobile robots are required to navigate freely in a complex and crowded environment in order to provide services to humans. For this navigation ability, deep reinforcement learning (DRL)-based methods are gaining increasing attentions. However, existing DRL methods require a wide field of view (FOV), which imposes the usage of high-cost lidar devices. In this paper, we explore the possibility of replacing expensive lidar devices with affordable depth cameras which have a limited FOV. First, we analyze the effect of a limited field of view in the DRL agents. Second, we propose a LSTM agent with Local-Map Critic (LSTM-LMC), which is a novel DRL method to learn efficient navigation in a complex environment with a limited FOV. Lastly, we introduce the dynamics randomization technique to improve the robustness of the DRL agents in the real world. We found that our method with a limited FOV can outperform the methods having a wide FOV but limited memory. We provide the empirical evidence that our method learns to implicitly model the surrounding environment and dynamics of other agents. We also show that a robot with a single depth camera can navigate through a complex real-world environment using our method.

In this paper, we propose a new methodology to embed deep learning-based algorithms in both visual recognition and motion planning for general mobile robotic platforms. A framework for an asynchronous deep classification network is introduced to integrate heavy deep classification networks into a mobile robot with no loss of system bandwidth. Moreover, a gaming reinforcement learning-based motion planner, a novel and convenient embodiment of reinforcement learning, is introduced for simple implementation and high applicability. The proposed approaches are implemented and evaluated on a developed robot, TT2-bot. The evaluation was based on a mission devised for a qualitative evaluation of the general purposes and performances of a mobile robotic platform. The robot was required to recognize targets with a deep classifier and plan the path effectively using a deep motion planner. As a result, the robot verified that the proposed approaches successfully integrate deep learning technologies on the stand-alone mobile robot. The embedded neural networks for recognition and path planning were critical components for the robot.

Designing an actuator system for highly dynamic legged robots has been one of the grand challenges in robotics research. Conventional actuators for manufacturing applications have difficulty satisfying design requirements for high-speed locomotion, such as the need for high torque density and the ability to manage dynamic physical interactions. To address this challenge, this paper suggests a proprioceptive actuation paradigm that enables highly dynamic performance in legged machines. Proprioceptive actuation uses collocated force control at the joints to effectively control contact interactions at the feet under dynamic conditions. Modal analysis of a reduced leg model and dimensional analysis of DC motors address the main principles for implementation of this paradigm. In the realm of legged machines, this paradigm provides a unique combination of high torque density, high-bandwidth force control, and the ability to mitigate impacts through backdrivability. We introduce a new metric named the “impact mitigation factor” (IMF) to quantify backdrivability at impact, which enables design comparison across a wide class of robots. The MIT Cheetah leg is presented, and is shown to have an IMF that is comparable to other quadrupeds with series springs to handle impact. The design enables the Cheetah to control contact forces during dynamic bounding, with contact times down to 85 ms and peak forces over 450 N. The unique capabilities of the MIT Cheetah, achieving impact-robust force-controlled operation in high-speed three-dimensional running and jumping, suggest wider implementation of this holistic actuation approach.

This paper presents a control system platform architecture developed for multi-degrees of freedom (DoFs) robots capable of highly dynamic movements. In robotic applications that require rapid physical interactions with the environment, it is critical for the robot to achieve a high frequency synchronization of data processing from a large number of high-bandwidth actuators and sensors. To address this important problem in robotics, we developed a control system architecture that effectively utilizes the advantages of modern parallel real-time computing technologies: multicore CPU, the Field Programmable Gate Array (FPGA), and distributed local processors. This approach was implemented in the fast running experiments of the MIT Cheetah. In such a highly dynamic robot, the required control bandwidth is particularly high since the MIT Cheetah's leg actuation system is designed to generate high force (output torque up to 100Nm) with high bandwidth (400Hz electrical, 120Hz mechanical) with minimal mechanical impedance for fast locomotive capability. On the integrated control system, a multi-layered architecture is programmed. Inspired by the MapReduce model and the pipelining method, more than 50 processes are operated in parallel, and major processes among them are optimized to achieve the maximum throughput. The proposed architecture enables the control update frequency 4 kHz. With this control system platform, we achieved a high-force proprioceptive impedance control [1], and a trot-running up to 6 m/s with a locomotion efficiency rivaling animals [2]. This control system architecture is well suited for the future trend towards real-time computing system and, thus can be a candidate for a future standard robot control platform.

In this paper, we introduce the design principles for highly efficient legged robots and the implementation of the principles on the MIT Cheetah robot. Three major energy loss modes during locomotion are heat losses through the actuators, losses through the transmission, and the interaction losses that includes all losses of the system interacting with the environment. We propose four design principles that minimize these losses: employment of high torque density motors, low impedance transmission, energy regenerative electronics and a design architecture that minimizes the leg inertia. We present the design features of the MIT cheetah robot as an embodiment of these principles. The resulting cost of transport (COT) is 0.51 during 2.3 m/s running, which rivals running animals in the same scale.

High speed legged locomotion involves high acceleration and extensive loadings of the leg, which impose critical challenges in actuator design. We introduce actuator dimensional analysis for maximizing torque density and transmission `transparency'. A front leg prototype developed based on insight from the analysis is evaluated for direct proprioceptive force control without force sensors. The vertical stiffness controlled leg was tested on a material testing device to calibrate the mechanical impedance of the leg. By compensating transmission impedance from commanded torque, the leg was able to estimate impact force. For the impact test, the mean absolute error as a ratio of full scale sensor force is 0.041 in the 3406 N/m stiffness experiment and is 0.049 in the 5038 N/m experiment. The results indicate that prescribed force profile control is possible during high speed locomotion.

This paper presents a soft robotic platform that exhibits peristaltic locomotion. The design principle is based on the unique antagonistic arrangement of radial/circular and longitudinal muscle groups of Oligochaeta. Sequential antagonistic motion is achieved in a flexible braided mesh-tube structure with NiTi coil actuators. A numerical model for the mesh structure describes how peristaltic motion induces robust locomotion and details the deformation by the contraction of NiTi actuators. Several peristaltic locomotion modes are modeled, tested, and compared on the basis of locomotion speed. The entire mechanical structure is made of flexible mesh materials and can withstand significant external impacts during locomotion. This approach can enable a completely soft robotic platform by employing a flexible control unit and energy sources.