Daniel M. Lofaro

Toward a user-guided manipulation framework for high-DOF robots with limited communication

This paper presents our progress toward a user-guided manipulation framework for High Degree-of-Freedom robots operating in environments with limited communication. The system we propose consists of three components: (1) a user-guided perception interface which assists the user to provide task level commands to the robot, (2) planning algorithms that autonomously generate robot motion while obeying relevant constraints, and (3) a trajectory execution and monitoring system which detects errors in execution. We have performed quantitative experiments on these three components and qualitative experiments of the entire pipeline with the PR2 robot turning a valve for the DARPA Robotics Challenge. We ran 20 tests of the entire framework with an average run time of two minutes. We also report results for tests of each individual component.

The Ach Library: A New Framework for Real-Time Communication

Correct real-time software is vital for robots in safety-critical roles such as service and disaster response. These systems depend on software for locomotion, navigation, manipulation, and even seemingly innocuous tasks such as safely regulating battery voltage. A multiprocess software design increases robustness by isolating errors to a single process, allowing the rest of the system to continue operation. This approach also assists with modularity and concurrency. For real-time tasks, such as dynamic balance and force control of manipulators, it is critical to communicate the latest data sample with minimum latency. There are many communication approaches intended for both general-purpose and real-time needs [9], [13], [15], [17], [19]. Typical methods focus on reliable communication or network transparency and accept a tradeoff of increased message latency or the potential to discard newer data. By focusing instead on the specific case of real-time communication on a single host, we reduce communication latency and guarantee access to the latest sample. We present a new interprocess communication (IPC) library, Ach which addresses this need, and discuss its application for real-time multiprocess control on three humanoid robots (Figure 1). (Ach is available at http://www.golems.org/projects/ach.html. The name Ach comes from the common abbreviation for the motor neurotransmitter Acetylcholine and the computer networking term ACK.).

Multi-process control software for HUBO2 Plus robot

Humanoid robots require greater software reliability than traditional mechatronic systems if they are to perform useful tasks in typical human-oriented environments. This paper covers a software architecture which distributes the load of computation and control tasks over multiple processes, enabling fail-safes within the software. These fail-safes ensure that unexpected crashes or latency do not produce damaging behavior in the robot. The distribution also offers benefits for future software development by making the architecture modular and extensible. Utilizing a low-latency inter-process communication protocol (Ach), processes are able to communicate with high control frequencies. The key motivation of this software architecture is to provide a practical framework for safe and reliable humanoid robot software development. The authors test and verify this framework on a HUBO2 Plus humanoid robot.

Humanoid throwing: Design of collision-free trajectories with sparse reachable maps

This work shows a method of creating trajectories to achieve end-effector velocity control for high degree of freedom position controlled, high-gain, robots. The focus of this work is throwing an object. It is shown that the full reachable area of the end-effector does not need to be known to achieve the desired velocity when a good collision model of the robot is available. The end-effector velocity (magnitude and direction) is specified as well as a duration of this velocity. A sparse map of reachable end-effector positions in free space and the corresponding poses in joint space is created using random sampling in joint space and forward kinematics. The desired trajectory in free space is placed within the sparse map with the first point of the trajectory being a known pose from the original sparse map. The Jacobian Transpose Controller method of inverse kinematics is then used to find the subsequent points in the trajectory. Each pose in the trajectory is checked against the collision model to guarantee no self-collisions. This method was tested on the 130 cm tall full size humanoid Jaemi Hubo and its virtual representation.

Humanoid pitching at a Major League Baseball game: Challenges, approach, implementation and lessons learned

Three different approaches of having a full-size humanoid throw the first pitch at a Major League Baseball game are tested and implemented. The approaches include kinematic mapping using a motion capture system to capture a humans throwing motion then mapping that to a full-size humanoid. The second method is a fully automated approach that uses the sparse reachable map to provide viable full body throwing trajectories to provide the end effector with the desired velocity. The third approach borrows from the animation industry. The key-frames of the desired trajectory are constructed by hand. The time between each key-frame is defined by the user. Interpolation methods are used to smoothly move between key frames while limiting the jerk. Each method is analyzed and tested in simulation and on physical hardware. The full-size humanoid used is the Hubo series robot. Based on the latter tests one method was chosen to successfully throw the ceremonial first pitch at a Major League Baseball game in April 2012.

Robot audition and beat identification in noisy environments

In pursuit of our long-term goal of developing an interactive humanoid musician, we are developing robust methods to determine musical beat locations from live acoustic sources. A variety of beat tracking systems have been previously developed, but for the most part they are optimized for direct audio input (no acoustic channel and no noise). The presence of an acoustic channel and noise typically degrades performance substantially. A robots motors, in particular, create nonstationary noise that can be difficult for a beat detection system to accommodate, Using an algorithm previously developed by the authors, we explore techniques for reducing the effects of the acoustic channel and noise on the system, enabling a humanoid to robustly follow music under realistic conditions.

Feasibility of cloud enabled humanoid robots: Development of low latency geographically adjacent real-time cloud control

This paper explores the feasibility and proposes a method on how to obtain high frequency real-time controllers operating in loop with physical robot hardware over a geographically adjacent cloud server architecture. Having the cloud in the loop has many purposes including increasing computation, decreasing power usage, reducing overall robot weight etc. Today when robots use the cloud in loop it is typically for sharing information, high level planning, and other non-realtime tasks. All of the balancing and stability algorithms (the heart of the humanoid) stay onboard the robot. What if we could run high frequency real-time loops over the cloud? The better question is how would we do that? As with any real-time system latency is a big factor in the application and operation frequency. This paper shows that with a geographically adjacent server approach it is feasible to obtain high frequency real-time control in loop with with a physical robot over the cloud. The feasibility of such a system running in-loop on humanoids is explored and tested.

DARPA Robotics Challenge: Towards a user-guided manipulation framework for high-DOF robots

Supervision and teleoperation of high degree-of-freedom robots is a complex task due to environmental constraints such as obstacles and limited communication, as well as task specific requirements such as using more than one end-effector at the same time. In this work we present a supervision and teleoperation framework that allows an operator to see the surroundings of a robot in 3D, make necessary adjustments for a dual or single arm manipulation task, preview the task in simulation before execution, and finally execute the task on a real robot. The framework has been applied to the valve turning task of the DARPA Robotics Challenge on the PR2, Hubo2+, and DRCHubo robots.

Humanoid throws inaugural pitch at Major League Baseball game: Challenges, approach, implementation and lessons learned

Three different approaches of having a full-size humanoid throw the first pitch at a Major League Baseball game are tested and implemented. The approaches include kinematic mapping using a motion capture system to capture a humans throwing motion then mapping that to a full-size humanoid. The second method is a fully automated approach that uses the sparse reachable map to provide viable full body throwing trajectories to provide the end effector with the desired velocity. The third approach borrows from the animation industry. The key-frames of the desired trajectory are constructed by hand. The time between each key-frame is defined by the user. Interpolation methods are used to smoothly move between key frames while limiting the jerk. Each method is analyzed and tested in simulation and on physical hardware. The full-size humanoid used is the Hubo series robot. Based on the latter tests one method was chosen to successfully throw the ceremonial first pitch at a Major League Baseball game in April 2012.

Low latency bounty hunting and geographically adjacent server configuration for real-time cloud control

This paper explores the feasibility and proposes a method on how to obtain high frequency real-time controllers operating in loop with physical robot hardware over a geographically adjacent and bounty hunting cloud server architecture. Having the cloud in the loop has many purposes including increasing computation, decreasing “on robot” power usage, reducing overall robot weight etc. Today when robots use the cloud in loop it is typically for sharing information, high level planning, and other non-real-time tasks. All of the balancing and stability algorithms stay onboard the robot. What if we could run high frequency real-time loops over the cloud? The better question is how would we do that? As with any real-time system latency is a big factor in the application and operation frequency. This paper shows that with a geographically adjacent and bounty hunting server approach it is feasible to obtain high frequency real-time control in loop with with a physical robot over the cloud. The feasibility of such a system running in-loop on humanoids and wheeled robots are explored and tested.