Free Software and High-Power Rocketry: The Portland State Aerospace Society

Free Software and High-Power Rocketry:
The Portland State Aerospace Society

James Perkins    Andrew Greenberg
Jamey Sharp    David Cassard    Bart Massey
Portland State Aerospace Society
Computer Science Department
Portland State University
Portland, OR USA  97207--0751
`james@psas.pdx.edu`    `andrew@psas.pdx.edu`
`jamey@cs.pdx.edu`    `dcassard@psas.pdx.edu`    `bart@cs.pdx.edu`
`http://psas.pdx.edu`

Copyright 2003 James Perkins, Andrew Greenberg, Jamey Sharp, Dave Cassard, Bart Massey
First published in Proceedings of the 2003 USENIX Annual Technical Conference (USENIX '03).
Permission is granted for the noncommercial reproduction of the complete work for educational or research purposes only.

Abstract

The Portland State Aerospace Society (PSAS) is a small, low-budget amateur aerospace group. PSAS is currently developing medium-sized sub-orbital rockets with the eventual goal of inserting nanosatellites (satellites that weigh less than 10 kg) into orbit. Because achieving orbit requires a navigation system able to guide the rocket along an orbital trajectory, PSAS is pioneering an open source and open hardware avionics system that is capable of active guidance.

In this paper, we describe how free software and open hardware have dramatically changed the capabilities of amateur aerospace groups like PSAS. We show how we have applied existing and custom free software to the avionics and ground support systems of a sub-orbital sounding rocket, and discuss what further work must be done.

We conclude that the sophistication and complexity achieved by current amateur avionics projects---which are beginning to challenge the distinction between amateur and professional---would not be possible without the use of free software.

1 Overview

This paper details the role of free software and open hardware in our group, the Portland State Aerospace Society (PSAS). First we introduce the field of rocketry, and how PSAS is making contributions to the field. This is followed by a history of the group's rocket development and an overview of our current project. We then cover the details of our system, including requirements, flight software, ground software, and the project management and collaboration tools. Finally, we discuss our future work, and draw conclusions about the applicability of free software to projects like ours.

2 Introduction to Rocketry

Rocketry has a rich history of significant contributions made by amateur groups. Indeed, most national space programs can trace their roots back to small, active groups of amateurs working on rocketry during the early part of the 20th century [Win83].

Amateur rocket motor classifications are categorized by their total impulse, which is total force multiplied by burn time. ``A'' motors have up to 2.5 Newton-seconds (Ns) of total impulse, and each successive letter of the alphabet doubles the impulse range of the previous letter [Tri].

Today, rocketry includes a wide spectrum of participants and can generally be divided into four categories.

Model rocketry involves the smallest and most common types of rockets, with motors made of pressed black powder that are smaller than 320 Ns (``H''), and bodies built out of balsa wood and cardboard or phenolic (resin) tubes. Most model rockets weigh less than a kilogram and are recovered with a small parachute or streamer. Since model rockets generally stay below 460 m (1,500 ft), they avoid most government regulation in the US.

Hobby rocketry begins from the upper limits of model rocketry. Motors delivering up to about 10,240 Ns (``M'') of thrust are common. Motors are typically made of composite fuel (ammonium perchlorate in a HTPB plastic binder), and are sometimes clustered or staged to achieve greater power. Some high-powered hobby rockets can reach more than 6 km (20,000 ft), so launches in the US are government regulated and require launch waivers. Most hobby rockets use commercially available, single board avionics systems that sense the best time to eject parachutes [alt], although some custom-built avionics packages have been much more sophisticated.

Amateur rocketry usually implies a certain level of innovation and customization, such as custom motors, custom avionics and metal airframes. This innovation may be inspired by a lack of suitable commercial solutions or a desire to ``do it yourself''. The current altitude record for an amateur group is 85 km (280,000 ft) with a 327,680 Ns (``R'') motor [rrs].

Professional rocketry is the most familiar category. It includes organizations such as NASA and the European Space Agency (ESA), along with their private contractors such as Lockheed Martin, Boeing, and Orbital Sciences. Over time, the line between amateur and professional rocketry has blurred. Some industry observers discriminate by financial gains, technical expertise, or sheer altitude achieved, but there is no common standard. Indeed, as technologies such as computational power and integrated sensors become cheaper and more available, the capabilities of amateur groups have begun to catch up to some professional projects.

3 The Portland State Aerospace Society

The Portland State Aerospace Society (PSAS) was founded in 1997 by two students at Portland State University (PSU) in Portland, Oregon to provide an aerospace-based, systems-level educational design project. PSAS has launched one hobby and two high-powered amateur rockets. The group has grown to include more than three dozen people, including high school, undergraduate and graduate students, as well as engineers from local industry. Current PSAS projects are focused on taking small, manageable steps towards the the distant vision of inserting nanosatellites (satellites that weigh less than 10 kg) into orbit.

Is it possible for amateur rocketry groups to achieve orbit? None have, to date. In most countries, the regulatory hurdles are at least as much of a challenge for an amateur group as the substantial technological and financial issues. The US Commercial Space Launch Act of 1984 requires any rocket with over 890,000 Ns (``S'') of impulse lasting for 15 seconds or more to meet stringent safety requirements set by the FAA, NASA, and Commercial Space Transportation Board (CSTB) [csl, sls]. These requirements include an extensive safety analysis, which can take years and cost hundreds of thousands of dollars. Furthermore, vehicles which actually achieve orbit must comply with international ``space law'' as laid out in various international agreements [un].

While fulfilling the goal of achieving orbit may be beyond the ability of our small group, the enabling technologies needed to get there by an amateur group are now readily obtainable: inexpensive computational power, sophisticated sensors, high-power actuators, and the availability of robust, open source software for engineering and logistical support.

3.1 Toward Amateur Active Guidance

To achieve orbit at minimum cost, a rocket must follow an ``orbital trajectory'' that minimizes both the aerodynamic drag in the lower atmosphere and the time to get to orbit [Wer92]. To follow such a trajectory, the rocket must be able to measure its current trajectory, compare against the planned trajectory, and actively correct for errors. This ability to follow a trajectory is called ``active guidance'', and the authors know of no amateur group that has yet achieved this. To bridge this gap in amateur rocketry technology, PSAS has chosen to work on open source and open hardware high-powered amateur rockets that are capable of active guidance.

There are many meanings of active guidance. Here we use active guidance in the classic rocket sense: a guidance computer on-board the rocket measures the rocket's current position, heading and course. The guidance computer then activates a steering mechanism in order to guide the rocket along a predetermined path.

To determine the rocket's position, attitude (orientation) and trajectory (flight path), a rocket's flight computer uses data from one or more sensors:

GPS receivers provide absolute position at a slow (~1 Hz) rate.
Inertial Measurement Units (IMUs) provide relative linear and rotational acceleration, velocity and position using accelerometers and gyroscopes at a faster rate (~1 kHz).
Magnetometers provide attitude (orientation) by comparing the local magnetic field to a 3D map of the Earth's magnetic field.
Pressure sensors allow altitude to be computed from atmospheric pressure models.
Optical sensors, such as star and horizon trackers, may be used to compute attitude.

Each of these sensors have different signal and noise characteristics. Kalman filtering [CC99] is a signal processing algorithm for combining noisy sensor data that provides guarantees on the optimality of its estimate. This estimate can then be used as an approximation of the rocket's true position, attitude and trajectory.

To steer the rocket, some mechanism must apply a force to the rocket during flight. Steering mechanisms vary widely:

Small fins are common, but are ineffective above about 25 km (82,000 ft).
Reaction Control Systems use small rocket motors to adjust the heading of the rocket, but are usually large, heavy systems that are ineffective in the lower atmosphere.
Thrust Vector Control changes the angle of the main motor's thrust, for example by gimballing the main motor nozzle, by independently throttling multiple clustered motors, by putting small movable vanes in the path of the exhaust, or by injecting extra oxidizer into the edge of a fuel-rich exhaust stream [SB00].

To ``close the loop'', estimates of position, attitude and trajectory are used to calculate steering commands. The most difficult part of active guidance may be getting the signal processing in the control loop correct. Many famous rocket failures, such as the first launch of Ariane 5, were due to simple errors in design or implementation of the navigation system algorithms [ins].

3.2 Related Work

Many amateur groups exist today, but none that the authors know of are currently working on active guidance. Guided amateur designs do exist, but all either track the sun or are used to simply stabilize the rocket in flight. None can actually be used to follow a planned trajectory. An excellent example of a guided, but not actively guided, rocket is the MARS Rocket Society's gyroscope-controlled gimballed-nozzle rocket [mar].

3.3 Launch Vehicles

Because we strongly emphasize safety, reliability and new functionality for every launch, PSAS has launched only four times over its six year history. In a field with failure rates as high as 30% at some events, we have not yet lost a vehicle. Further, each new launch has demonstrated new airframe or avionics functionality which justified the time and expense of performing a launch.

Table 1: Comparison of PSAS Launch Vehicles

Name Size Dia. Weight Ns Altitude Cost Design CPU OS

LV0 1.8 m 10.0 cm 5.4 kg J 0.3 km $500 4 mo. PIC16C Interrupt-driven

LV1 3.4 m 11.0 cm 19.5 kg M 3.6 km $2000 18 mo. PIC17C Interrupt framework

LV2 4.0 m 13.6 cm 46.0 kg P 23.0 km $10000 >21 mo. 5x86 Linux / RTLinux

3.3.1 Launch Vehicle No. 0

The first PSAS project began with four volunteers and a simple hobby rocket dubbed ``Launch Vehicle No. 0'' (LV0). The team modified a commercial kit of cardboard tubes and balsa wood with fiberglass and epoxy resin and added a simple avionics system.

While the airframe took only a weekend to complete, the avionics system took two people a few months of building and testing. The avionics system had just a few components: a 1 MHz 8 bit RISC microcontroller; a micro-electro-mechanical (MEMs) accelerometer; a 426 MHz 1 W amateur television (ATV) transmitter; a monochrome video camera; and a commercial logging altimeter. The interrupt-driven microcontroller firmware was written in assembly language. Accelerometer samples were downlinked using the ATV audio channel at 300 bps. Telemetry data was then logged on a 386 DOS laptop.

LV0 was launched on June 7, 1998 (Table 1). Not surprisingly, things went wrong. The airframe was all but unstable and a short circuit in the wiring harness erased system memory. However, the received images and data proved the soundness of the overall system concept.

3.3.2 Launch Vehicle No. 1

LV0's success attracted more volunteers: about a dozen people designed and built the next generation rocket, Launch Vehicle No. 1 (LV1). The airframe was made of carbon fiber over a PVC and aluminum core, and was built by one person over 3 months, with little engineering analysis.

The LV1 avionics system was arguably the most advanced amateur rocket avionics package in the world in 1999. Its subsystems included:

A custom flight computer board with an 8 bit 33 MHz RISC microcontroller, 1 MB of non-volatile SRAM for data logging, pyrotechnic ignition circuitry, and interface circuitry to other subsystems including temperature and pressure sensors.
An inertial measurement unit (IMU): X, Y, and Z axis linear accelerometers and yaw, pitch, and roll rate gyroscopes. These sensors produced a ``six degree of freedom'' measurement which was numerically integrated to calculate the rocket's 3D position and velocity.
A commercial 12-channel GPS receiver.
A color video camera and microphone that transmitted over a 426 MHz ATV transmitter.
A 913 MHz 1 W transmitter that sent 19.2 kbps telemetry data encoded by the flight computer.
A 146.43 MHz (2 m) amateur radio receiver with a DTMF decoder. Decoded tones were sent to an independent microcontroller that could fire the recovery system or send commands to the flight computer in an emergency.

The flight computer's assembly-language, interrupt-driven executive sampled all sensors, logged data, and transmitted low-pass filtered telemetry. The firmware also used the GPS, IMU and pressure sensors to determine the rocket's altitude and thus when to deploy the recovery system.

LV1 required extensive ground support due to its size and complexity. A surplus pneumatic lifter was modified into a launch tower with a 6 m (20 ft) launch rail. Launch control software automatically performed the countdown and launched the rocket via a 903 MHz wireless link to a small microcontroller-run relay board.

Another Linux-based ground computer captured, logged and displayed live telemetry data with a custom GTK-based application. This helped the range safety officer decide if the flight was proceeding as planned: if not, the emergency radio could be used to deploy the recovery system (parachutes) even if the flight computer failed.

LV1 was launched on April 11, 1999 and again (as LV1b) on October 7, 2000 (Table 1). Half of the $2,000 development cost came from an IEEE/AT&T ``Student Enterprise'' grant.

LV1's complexity led the dozen-plus volunteers to divide into airframe, avionics and logistics teams. Each team had their own mailing list, web page and FTP folders. However, formal design methods were mostly ignored and the web site was updated infrequently. Designs were implemented without review, which caused schedule slips and and two ``scrubbed'' launches.

3.3.3 Launch Vehicle No. 2

The next generation launch vehicle needed the flexibility, modularity and extensibility that LV1 lacked. A few of the lessons learned include:

Real-time Operating Systems (RTOS's) hand-built in assembly language are not a good idea for a complex, multi-volunteer effort. An off-the-shelf RTOS with better networking and free, easily learned development tools is much more appropriate.
The navigation software needs serious computational power: a CPU with hardware floating point support is highly desirable.
Subsystems in the avionics system must be easily added, removed or swapped.
The airframe itself must be flexible and expandable to handle unforeseen requirements.

The design of LV2 took more than a year of careful coordination between the avionics and airframe teams. The design process was much more formal than had previously been tried. A white paper was written on the various design options for the avionics system. The airframe team used finite-element analysis to predict the performance of different airframe structures. To fund the project, PSAS applied for and won the 2000 Oregon Space Grant, a $10,000 NASA-sponsored small grant program.

The LV2 airframe uses a flexible and modular design to facilitate swapping out entire subsystems such as the avionics, recovery or propulsion (motor) unit. We have also separated the avionics system from the payload module, enabling LV2 to fly other academic or amateur rocketry payloads. The resulting airframe is made of cylindrical aluminum modules covered by a fiberglass aeroshell. Simulations predict a maximum altitude of 23 km (75,000 ft) (Table 1).

Based on a computational complexity analysis of the navigation software, we decided to use a flight computer with a floating point unit (FPU), better support for multitasking, and a modern development toolchain. However, a single-processor avionics system was unappealing because of the high-rate, I/O-intensive tasks many of the sensors required: controlling a high speed analog-to-digital converter and running a closed loop motor controller, for example.

A multi-node common bus solves many of these problems by enabling a larger, more powerful central flight computer to communicate with many smaller microcontrollers. This allows a ``smart sensor'' and ``smart actuator'' approach that frees up the central processor to perform higher-level calculations and supervisory tasks.

We chose the Controller Area Network (CAN) bus as our intra-rocket multi-node bus. The CAN bus is an automotive bus developed by Robert Bosch, GmbH which is quickly gaining acceptance in both the industrial and aerospace markets. CAN is a multi-master, losslessly-arbitrated serial bus that can be run up to 1 Mbps. The CAN bus includes packet-level checksums and tolerance of node errors (including logic that forces a node off of the bus if it is causing errors). Perhaps the most interesting aspect of CAN is its message-based identification of packets: instead of node addresses, the CAN bus identifies and prioritizes the messages based on an 11 bit identifier. Each message, such as GPS location, or IMU inertial data, is broadcast on the bus with a unique message ID [can].

For the central flight computer we have selected the PC104 form-factor. This allows us to use standard off-the-shelf parts instead of taking the time and effort to make our own. The flight computer consists of:

A Jumptec MOPS520 PC104+ board, a 133 MHz 5x86 processor (the AMD SC520) with 64 MB of SDRAM, typical PC ports, and a CAN interface.
A carrier board for a 128 MB CompactFlash hard disk.
A PCMCIA carrier board for a Lucent Orinoco 802.11b card.

Considerable difficulty was encountered in developing the wireless telemetry system for LV1. We thus wanted a commercially available long-range high-speed bi-directional wireless telemetry system for LV2: such a system could be used for telemetry as well as by the rocket, launch tower and ground computers. After some research, we gravitated toward the Amateur Radio Relay League (ARRL) 802.11b standard. ARRL 802.11b is the IEEE 802.11b 2.4 GHz spread spectrum standard, but operated under the FCC amateur radio regulations (FCC Part 97) instead of the low-power, unlicensed regulations of the 2.4 GHz ISM band (FCC Part 15). Running under ARRL 802.11b allows us to use up to 100 W of radiated power. Our current system uses a standard PC-Card (Lucent Orinoco), a 1 W bidirectional power amplifier, and high gain (+12 dB) helical ground antennas [arr].

The smaller sensor/actuator CAN nodes use the Microchip PIC18F458, an 8 bit, 40 MHz RISC flash-memory microcontroller with a built-in CAN protocol unit. There are currently five CAN nodes:

The LV1 IMU.
A SigTech Navigation MG5001 OEM GPS receiver.
A power interface node to control system power and track battery charge.
A recovery node---a battery-backed up CAN node with high voltage pyrotechnic firing circuitry. Like LV1, it has a 2 m amateur radio receiver which decodes DTMF tone sequences sent as emergency commands.
An ATV node, consisting of a color video camera, ATV transmitter, power amplifier and an overlay board that displays textual vehicle status information along with NTSC flight video.

The CAN nodes have proven to be useful general-purpose building blocks: new nodes are frequently prototyped atop a generic ``misc CAN node'' design.

4 Software Overview

Figure 1: LV2 System Software

The software system is divided into three major areas:

Flight Software
- Flight Computer Software flies in the rocket, and runs on a single-board computer running Debian GNU/Linux (Section 5.1).
- Avionics Firmware flies in the rocket, and runs on multiple independent microcontroller nodes (Sections 5.2--5.3).
Ground Software
- Launch Control Software is used by the flight control officer to sequence the rocket launch. Launch Tower Software manages launch tower electronics (Section 6.1).
- Telemetry Display Software is used by the flight control officer and spectators to view the rocket's status. (Section 6.2).
- Uplink Software is operated by the range safety officer at launch control and sends emergency commands through the 2 m uplink (Section 6.3).
Collaboration Software runs on the (Portland State University) hosted PSAS server. The PSAS server is the locus for a variety of services including mailing lists, collaborative web pages and CVS repositories (Section 7).

4.1 Functional Requirements

There are a number of mandatory requirements for a successful rocket launch. In order of priority:

Safety. The risk of death, injury, and property damage must be minimized
Reliable Recovery. The airframe must be safely recovered. The principal software constraint is to deploy the recovery parachute only at apogee, when vertical airspeed is at a minimum.
Flight Data Recovery. The purpose of every flight is to collect new data: data recovery is thus essential.
Telemetry Downlink. Real time sensor data must be easily monitored by the launch controller to decide whether or not to override the flight computer via the emergency uplink.
Radio Dropout Tolerance. Radio links are notoriously unreliable. The software must be able to tolerate link failures.
Recovery Assistance. Position information must be available to the recovery teams tracking the rocket after the parachutes have opened. High winds or malfunction can carry the rocket kilometers away during ascent or descent.
Power Management. Intelligent power management is crucial to reliable performance. Power use also constrains peak altitude. Lithium batteries store energy at a density of ~300 WHr/kg. Thus, during a standard flight profile each watt used on the rocket adds 30 g (1.1 oz) of battery weight. Simulations predict that the altitude- to-weight ratio is -335 m/kg. Thus every watt used on the rocket means a reduction of 10 m (33 ft) of altitude.

4.2 Logistical Requirements

The volunteer nature of our group imposes some special project management requirements:

Minimize Coordination Overhead. In our project, collaboration is entirely ad-hoc. Project synchronization and access control mechanisms must reflect this.
Minimize Training. Well-known development methods are necessary to enable immediate contributions by our new volunteers.
Limit Costs. Funding is scarce: we must take advantage of any available cost reductions.

4.3 Choosing a Common Platform

Our requirements for a low-cost, reliable and flexible platform discourage the use of non-free, proprietary and unfixable software and suggest using well-documented, open-standards-based free software. GNU/Linux not only runs on the limited hardware we have available, but also enables the use of thousands of free software applications available for UNIX environments.

We have standardized on the Debian GNU/Linux distribution for all of our development, collaboration and rocket systems in order to reduce the amount of time spent maintaining and installing software [deb]. Thanks to Debian's Advanced Package Tool (apt), our systems generally require little maintenance effort while remaining relatively secure. This enables us to focus on our development efforts instead of struggling with platform and tool configuration. It also reduces the time necessary to configure development environments for our new volunteers.

4.4 From Soft to Hard Real-time

The term ``real-time'' refers to applications that require guaranteed bounds on the time between the occurrence of an event and the software's response. Soft real-time requirements allow for occasional missed deadlines under high load, but hard real-time applications must meet all deadlines under all conditions [Lab99].

Linux is designed to optimize throughput, not to guarantee hard real-time response. Processing may be delayed by mutexes, interrupt locks, high interrupt load, paging, and other causes. The scheduling latency of a Linux task may be improved using the sched_setscheduler() system call to identify it as a ``real-time'' task. Other tricks include using mlock() or mlockall() to avoid paging delays, and enabling full kernel preemption using various kernel patches [Gal95].

While these approaches improve mean response time, the system delays are still unbounded. One approach to regions of code requiring hard real-time response is to move them below the user API into or below the kernel, where the code paths and potential interruptions to them can be fully understood. For example, interrupt handlers might be used to schedule Linux kernel threads, limiting code analysis to the interrupt handler path.

For full control over response time, the entire flight computer application could be run on a hard real-time operating system. There are several light-weight commercial offerings, such as Wind River's VxWorks operating system, the QNX Microkernel and Jean Labrosse's MicroC/OS-II. However, commercial RTOS's generally require specialized programming knowledge, often support a narrower array of devices than free software, and can be expensive.

Fortunately, the best of both worlds may be found in real-time operating systems that run the entire Linux kernel and all user processes in the lowest-priority thread. One such offering is FSMLabs Inc.'s RTLinux [fsm]. The hard real-time elements of our software can be implemented using the real-time kernel primitives: everything else runs as Linux user processes.

The PSAS flight computer application has a variety of requirements. Some components have no real-time requirement, while others have soft or hard real-time requirements. For the next launch, our application's real-time requirements are soft: launch and apogee must be detected and handled within a second, and data must be logged and transmitted to the ground as soon as possible. However, the introduction of navigation algorithms on future launches will require us to move to hard real-time, and we plan to begin using RTLinux with our Fall 2003 launch.

5 Flight Software

Table 2: Processors used in the LV2 Avionics System

System Name Arch. Bits Speed RAM Flash MMU? OS

Flight Computer AMD 5x86 32 133 MHz 64,000 KB 128,000 KB Y Linux

GPS Receiver ARM7TDMI 32 40 MHz 128 KB 1,000 KB N eCos

CAN Nodes PIC18 8 40 MHz 2 KB 32 KB N PicCore

The LV2 flight system uses three significantly different kinds of processors (Table 2). This necessitates three significantly different software architectures: processes running in a 32 bit pre-emptive multitasking RTOS; threads running in a light-weight POSIX-compatible RTOS; and tasks running in a custom interrupt-driven framework. In this section we review these software architectures. Note that we distinguish firmware from software: firmware consists of small, hardware-oriented programs stored in nonvolatile memory and considered read-only by the local processor.

5.1 Flight Computer Software

A key element of the application software running on the 5x86-based flight computer (see Section 3.3.3) is its message-passing architecture. This software is currently written in C, and runs in a Debian GNU/Linux environment with a Linux 2.4 kernel. Messages are passed across the wireless link, across the CAN bus and to the nonvolatile data log (Figure 1).

Every component of the PSAS avionics system has different response speed and latency characteristics. In order to avoid holding up the whole application, each device is monitored and controlled by one or more asynchronous execution tasks. Coordination of these tasks is accomplished by passing messages between them, initially by using a message server task. We have been through three flight computer software designs with different message-handling schemes: ``Muxer'', ``Renegade'' and ``FCFIFO''. After introducing these designs, we will discuss some other flight-computer software elements of interest: the CAN Bus driver and our network configuration.

5.1.1 Muxer

Figure 2: Muxer Flight Computer Software Design

The first design was implemented in C using POSIX threads, TCP/IP sockets, and device I/O processes. The processes passed messages by connecting to a server process called Muxer, which broadcasted each message received to every other task.

Muxer's server thread began by initializing a shared buffer. It then listened for TCP/IP connections. The central data structure was a variable-sized shared queue of messages. When a client connection was accepted, the server thread spawned client queue reader and writer threads. The new client's queue reader thread slept, waiting for new messages to arrive in the queue. The client's queue writer waited for an incoming message and enqueued it. Queue access was serialized by a mutex, and semaphores were used to wake sleeping queue readers when new messages were waiting for delivery. The last client to read a message from the queue deallocated that queue element.

The Muxer design included a client library to ease creating a TCP/IP connection to the Muxer service. Clients were mostly other Linux processes on-board the flight computer or (via the wireless link) on the ground (Figure 2). A shell script first started the Muxer server process, then each of the client processes. The client processes would open and initialize their I/O device. They would then use the Muxer client library to open a connection to the Muxer server thread.

There were some difficulties with this design:

Excessive wake-ups. The star message topology with N clients resulted in each new message waking up N threads, which wrote N messages to the network stack, which woke up N processes to read the data.
Message synchronization. Messages were of variable size, making mid-reception synchronization to a message stream a challenge.
Debugging. Debugging POSIX threads was difficult. Threads would sometimes damage the environment of their neighbors.
Opacity. The queue structure, its mutex and use was not easily understood by the team.
Wrong protocol. TCP/IP was a poor choice for wireless communications, incurring multi-megabyte queuing and retries in the network layer. Ultimately a downlink process was added to bridge the messages into a UDP protocol for the wireless downlink.

A CAN bus reader was implemented and Muxer was demonstrated to work. Nonetheless, when changes were introduced that broke Muxer, no one was able to find the time to sort out why. Team members eventually realized that solving the flight computer design problem by starting from the message handling structure and then moving out from there was not working.

5.1.2 Renegade

Figure 3: Renegade Flight Computer Software Design

One team member began working on an independent effort, and titled it the ``Renegade'' design. Renegade continued the theme of a central Muxer task dispatching each message to all asynchronous I/O-handling clients. However, the architecture was implemented as a UNIX process with UNIX pipes to child I/O processes (Figure 3).

The top-down, application-oriented and simplified design was an improvement. Most importantly, needed rocket functionality was quickly completed:

CAN bus messages were read and distributed successfully to all I/O devices.
Intertask communication became independent of networking.
Wireless dropouts were handled efficiently rather than causing backups.
Fixed-sized messages eliminated synchronization issues.
Debugging with GDB and UNIX signals was trivial.
The Linux kernel design could be relied on for optimized intertask performance.

Renegade did have shortcomings.

The lack of message filtering and the broadcast architecture still led to excessive wake-ups.
The inherited pipe structure made it difficult to restart I/O processes from the shell.
UNIX pipes were difficult to inject data into for testing and debugging.
No throughput tests were ever attempted.

After some consideration of Renegade's shortcomings, we once again decided to redesign the flight computer software to correct them.

5.1.3 FCFIFO

Figure 4: FCFIFO Flight Computer Software Design

The third and completed design is called FCFIFO. It combines the independent processes from the Muxer design with the conventional UNIX interprocess communications of Renegade. In addition, it eliminates the central multiplexer (Figure 4). On startup, each process opens a uniquely named pipe to read messages from, then opens the named pipes of any processes it needs to communicate with. By eliminating the central multiplexer, task switch and copying overhead are minimized.

A simple Linux user-mode process-oriented application will satisfy the June 2003 soft real-time requirements. While named pipes do not guarantee a response time, they do aid in maximizing throughput. They also allow processes to run asynchronously, because they can buffer a few messages if the reading process falls behind. As components transition to RTLinux hard real-time tasks, this architecture will be ideal: named pipes are the preferred means of communication between Linux and RTLinux tasks.

Where it is advantageous for activities to occur independently, the application implements these via separate processes.

Logger takes messages on its named pipe and stores them in the nonvolatile log file.
CAN Reader opens the CAN bus device and blocks until it reads a message either from the CAN bus or from its named pipe. It passes messages to the Logger and networking for storage, sensor information to Smart, and a few other events to Sequencer.
CAN Writer opens the CAN bus device for writing and waits for incoming CAN messages on its named pipe. It then sends those messages out onto the CAN bus.
Fc2Net waits on its named pipe for a message, then broadcasts the message to the wireless network using our UDP application protocol.
Net2Fc initializes a UDP server port, then waits for a message. If a command comes from the ground via the wireless network, Net2FC passes the command on to other relevant processes.
Smart evaluates raw sensor data and converts it into a more useful form. (We discovered that this is a difficult component to name. After much thought, we decided it is the ``Smart Module Assembling Real-time Tasks''.) For example, GPS and IMU data are processed to obtain position, velocity, and acceleration information. Smart also detects events such as rocket launch, apogee, and touchdown.
Sequencer decides when the system should transition from one rocket flight state to another, and generates the commands needed to set up the CAN nodes for the new state.
Manager is analogous to the UNIX init process. It starts the whole set of named pipe children, restarts children that die unexpectedly, and manages a graceful application shutdown.

5.1.4 Linux CAN Bus Driver

CAN communication is performed through a GNU GPL 82527 CAN chip driver for Linux, written by Arnaud Westenberg and adapted to our the flight computer's 82527 interface [lcb]. The driver allows applications to communicate with the CAN bus through the /dev/can0 device node.

5.1.5 Network Configuration

During development, the flight computer communicates with development machines using its built-in Ethernet or local IEEE 802.11b wireless. In flight configuration it uses the ARRL 802.11b hardware described in Section 3.3.3. Standard Linux PCMCIA and Orinoco drivers configure and integrate the network hardware. However, a simple custom driver is needed to switch the system between IEEE and ARRL 802.11b modes: the 2.4 GHz bidirectional power amplifier can be turned on and off via two general purpose I/O pins on the flight computer.

The wireless network card is configured to use channel zero (2.400 - 2.450 GHz, which is within the amateur 13 cm band), IBSS ad-hoc networking mode, a fixed IP address from a private network address range, and the maximum power output available from the card. Enabling IBSS ad-hoc networking prevents the card from participating in auto-configuration via a management device such as a wireless access point. The telemetry viewing computers at launch control are the only other device similarly configured, to limit performance-robbing traffic collisions.

All network transmissions between the rocket and ground control use UDP broadcast packets. This provides a low-overhead transmission path with no packet retries. TCP was briefly considered for rocket communications, but TCP requires a reasonably error-free communications medium. Given the extreme 23 km distance of the rocket , antenna geometry effects, attenuation due to the motor's exhaust plume, and tracking error of the ground antennas, it cannot be assumed that the wireless link can support TCP communications.

The telemetry and control protocol is symmetric, time-stamped, and is logged on each end. This allows us to measure communications loss through post-flight data analysis. Periodic general status messages from the rocket allow the ground software to resynchronize key parameters rapidly even when communication is intermittent. A data payload checksum helps validate the quality of the received data, although invalid data frames are also logged to aid in link quality analysis.

5.2 CAN Node Firmware: ARM7TDMI

The Global Position System (GPS) receivers on-board LV2 are commercially available boards that use the Zarlink GP4020 GPS correlator chip [gpsb], a custom ASIC with a 12-channel correlator and an ARM7TDMI microcontroller core. The receivers come with closed-source proprietary firmware which runs the correlators, processes the satellite data, and transmits position and velocity information over a standard asynchronous serial bus.

Unfortunately, the navigation algorithms in commercially available GPS receivers fail to cope with the high dynamics of LV2 and will lose satellite lock when the vehicle exceeds 515 m/s, 18 km in altitude, or 8 g's of acceleration. Because LV2 will exceed these limits and a locked GPS receiver is a critical part of our navigation system, we have started a separate project: developing free firmware for GPS receivers which use the GP4020 chip. Dubbed GPL-GPS [gpsa], the firmware is based on Clifford Kelley's ``OpenSource GPS'' project [kel] and will allow the receivers to stay locked during highly dynamic flights. The open firmware will allow us to implement our own sophisticated processing techniques, including integrating the GPS and IMU into a GPS-aided inertial navigation system [FB99]. The firmware may also allow us to:

Implement local differential GPS corrections via a similar GPS receiver on the ground in a known, static position. Enabling a local differential base station can yield positioning accuracies in the 1 m range.
Use multiple GPS receivers to determine attitude (orientation) by comparing the position of their antennas.
Keep GPS lock on the satellites by integrating data from the inertial measurement unit to aid the correlator tracking loops.

We chose to use an existing RTOS in order to speed up firmware development, and because a multi-threading abstraction will simplify the GPS firmware design. GPL-GPS requires an RTOS with extremely low latency, POSIX compliance to facilitate porting, a reasonable free or open-source license, ability to run in 1 MB of flash memory and 128 KB of RAM, and a reasonable effort level (estimated as a time multiplier) for us to implement with our hardware. Based on our research (Table 3), we selected Red Hat's eCos (embedded Configurable Operating System) [Mas03] as the best fit to our criteria. We are currently working on the alpha release of the GPL-GPS project, which includes a port of eCos and Kelley's OpenSource GPS to the GP4020. We hope to have initial results by September 2003.

Table 3: Embedded Operating Systems Comparison

Name RT POSIX Free Small Effort

RTLinux Y Y Y N 1x

µcLinux N Y Y N 2x

MicroC/OS-II Y N N Y 4x

µITRON Y N Y Y 4x

eCos Y Y Y Y 4x

Custom Y N Y Y 10x

5.3 CAN node Firmware: PIC18

The Microchip PIC18F458 8 bit RISC microcontrollers [pic] used on the CAN nodes (see Section 3.3.3) have 1.5 KB of RAM, 32 KB of flash memory. and a long list of peripherals. The latter include an 8-channel 10 bit analog to digital converter, various serial interfaces including CAN, timers and an in-circuit debugging port. The PIC18 is used as a small and simple interface between sensors and the CAN bus, implementing a rudimentary ``intelligent sensor and actuator'' network. Because of the extremely tight memory constraints and the simple functionality of the firmware, we have written a custom C-language interrupt-driven framework called ``PicCore''.

Unfortunately, there are few free software applications for PIC18 development. For example, there is no GCC cross-compiler for the PIC18 family. The GDB debugger does not support Microchip's ``ICD2'' in-circuit debugging tool. Our current development environment consists of Microchip's no-cost Windows-based integrated development environment (MPLAB IDE) [pic], graciously donated copies of HI-TECH Software's PICC-18 C compiler [C18], and Microchip's ICD2 debugger/programmer. Our failure to get MPLAB to run under emulators such as WINE unfortunately require us to have Windows-based development environments for the PIC18 developers. Developing our own PIC18 development tools would exceed the resources of our project, so for now we have resigned ourselves to operate with two development environments. We are currently searching for equivalent hardware that has better free software development support.

6 Ground Systems

The software and hardware components on the ground (as shown in Figure 1) have several objectives:

Maintain the safety of all participants and bystanders.
Initiate the launch sequence (with means of emergency abort).
Receive and record telemetry data from the vehicle during flight,
Receive and record video broadcast from the vehicle during flight,
Initiate manual recovery procedures if the flight software appears to be failing.

Due to our small budget, our volunteer team has always relied on the kindness of strangers for available computing hardware. As a result, we need to be able to run our software on as many hardware and software configurations as possible. We have chosen Java as our ground systems language because Java byte-code enables us to run our code on any sufficiently powerful platform. Java's automatic memory management and simple GUI framework make ideal tools for user interfaces and data visualization. Java is well known: many of the developers in our group are familiar with it.

6.1 Launch Control Software

The launch control software is a Java application which steps through an automated launch sequence. The sequence of events is coordinated with the rocket and launch tower via ARRL 802.11b wireless links: the launch tower may be several kilometers away from launch control for safety reasons. Safety systems include manual interlocks and a rocket-controlled interlock triggered on flight computer diagnostic information. These interlocks prevent accidental launches in the event of software or hardware failure.

The launch tower computer is the same as the rocket's flight computer: a PC104 stack with 5x86 processor, 802.11b card, power supply and nonvolatile flash memory. The controller even has a CAN bus: a CAN-based relay board controls launch tower hardware such as strobe warning lights, sirens, and the rocket motor ignition relay.

6.2 Telemetry View Software

When the rocket is at its peak altitude of 23 km (70,000 ft), it is almost impossible to visually ascertain what is happening. The ``Rocketview'' telemetry display software is used by flight personnel and interested bystanders to observe the rocket's status. The flight controller uses telemetry information to decide whether the flight sequence is proceeding according to plan: if not, the range safety officer can manually deploy the recovery system using the emergency 2 m radio uplink.

The Rocketview application displays data received from the rocket via ARRL 802.11b wireless link. The display includes 6 strip charts (X, Y, and Z position as well as roll, pitch and yaw attitude), several text fields, and a free-form console log for miscellaneous messages from the flight computer. The Rocketview application was modeled after a GTK-based C program used for LV1. Because Rocketview is written in Java, it can take advantage of Java's graphics capabilities and of such open source components as the JFreeChart charting tool [jfr].

6.3 Uplink Software

The uplink system is an emergency backup communication link. The link is manually activated by the flight controller in case the flight computer fails to deploy the recovery system. The uplink computer runs a Java application which communicates via serial port with a Yaesu 50 W 2 m amateur transceiver. Upon user command, the application keys the Yaesu transmitter and generates a series of DTMF tones (the familiar touch-tone telephone tones) which are received by the recovery CAN node on the rocket. Besides emergency commands, the uplink software can also generate a small number of CAN bus commands for diagnosis purposes.

6.4 Video Capture System

Amateur TV signals received from the color camera on the rocket are recorded on tape and displayed live on a monitor. The flight personnel and bystanders can thus observe the flight from the rocket's point of view. The rocket uses a video overlay board to to display position and flight computer status information. Because the ATV transmitter is higher power and lower bandwidth than the ARRL 802.11b transmitter, this display functions as a backup telemetry downlink.

7 Collaboration Software

Before the LV2 project, PSU's Electrical and Computer Engineering department hosted a web site and Majordomo list server [maj] for PSAS activities. The site was primarily used to promote recruitment and showcase progress. No centralized version control was used and only one or two people updated the web site. Nearly all engineering documentation was passed back and forth between individuals using private email and removable computer media. This collaboration model was ill-suited to the size and complexity of the LV2 project: it was too difficult for team members to share their efforts quickly and effectively.

To resolve these problems, PSU agreed to host a donated Pentium-class computer. By co-locating a PSAS-owned computer, we gained flexibility in service provisioning and the flexibility to set up and experiment with collaboration software. Four of the team members with significant system administration experience have administrative privileges. Anyone on the team can get OpenSSH [ssh] shell accounts for purposes including software development and document preparation. PSU creates a backup of this system nightly using Amanda [ama].

Mailman [mai], with its web interface, makes maintaining the public announcement, team coordination, and site administration mailing lists easy. In particular, mail list subscribers handle their own subscription and mailing list preferences. Although less than ideal, CVS [cvs] has made it possible to share and back up work on software and other development documents, even across different operating systems.

The PSAS web site (http://psas.pdx.edu) still serves a promotional purpose. However, it now also serves as an ongoing project development notebook, with meeting minutes, task lists, specifications, team work areas, back-of-the-envelope calculations and diagrams online. This is accomplished using TWiki [twi]. TWiki is a Perl CGI-based Wiki Wiki Web implementation by Peter Thoeny and others, comprising a web-based collaboration platform designed for corporate and academic intranets. Using TWiki, any team member can add or alter site content using a web browser. TWiki automatically generates hypertext links, allows for attaching multimedia content, supports website search, can be configured to automatically notify subscribers of changes and can remind users of upcoming or overdue action items.

Begging for installation of tools from University staff, or working through one or two authorized web secretaries using static HTML, would severely constrain our entire project. Free software collaboration tools make it possible for team members to contribute and collaborate at their own pace, which dramatically improves our team's productivity.

8 PSAS Contributions

To date, amateur groups building custom avionics have developed private software of little use to other groups. In contrast, the PSAS software design is modular, based on inexpensive and open hardware, licensed under the GNU General Public License, and available for download from our web site. Any interested team is welcome to use and build on our work.

We strongly encourage other amateur groups to use our TWiki site, mailing lists and CVS tools to collaborate with us on our projects: advancing the state of the art through wider community collaboration.

A few of our free software and open hardware projects available for use are:

Gerbertiler. A Perl script which tiles together printed circuit board layout files. Tiling together many small boards into one large board significantly reduces the one-time fee associated with having a single board produced.
Misc. CAN node. A 4 x 7 cm board for prototyping PIC18F458/CAN node applications. Complete schematics, board layout files and working code are available.
PicCore. An interrupt-driven framework, peripheral APIs and network drivers for the PIC18F4xx family of microcontrollers, written using HI-TECH Software's PICC-18 C compiler.
GPL-GPS. As discussed in Section 5.2, GPL-GPS is a project to create free firmware for any Zarlink GP4020-based GPS receiver board.
LV1 IMU. A design for an inertial measurement unit that costs less than $150 to build. Complete schematics, board layout files and CAN node interface software are available.
CAN Bus Driver. Modest changes to Arnaud Westenberg's Linux CAN Bus Driver (Section 5.1.4) customize it to our COTS flight computer board and are being contributed back to that project. More importantly, we plan to create and contribute an RTLinux CAN Bus Driver derived from that work.
Navigation Algorithms. The signal processing algorithms for approximating system location (Section 3.1) may be adaptable to other combined inertial and GPS navigation applications.

9 Next Steps and Future Work

In late June or early July 2003, we are planning a low-altitude launch to flight test the avionics system to 6 km (20,000 ft) in central Oregon. This test will verify the basic operation of the critical hardware and software systems during flight. Systems to be tested include:

Avionics. The flight computer software including crude navigation algorithms, the ARRL 802.11b telemetry link, the emergency uplink system, and critical sensors and actuators such as the IMU and recovery system.
Launch Tower. Launch tower computer, launch software, umbilical cord and launch safety systems.
Ground Computers. telemetry display and uplink software.

In September 2003, we are planning to build on the June results and launch to 23 km (75,000 ft) in the Black Rock Desert of Nevada. By the September launch, we hope to:

Migrate the flight computer software to RTLinux.
Improve the sensor suite, including a next generation IMU, the first generation of the GPL-GPS, and a magnetometer.
Improve the navigation algorithms, including real-time GPS-aided inertial navigation routines to calculate position, attitude and trajectory.
Upgrade the ARRL 802.11b telemetry link to use forward error correction algorithms to improve the quality of data received and provide an ability to confidently reconstruct some amount of lost data.
Integrate the GPL-GPS receivers into the rocket and base station, creating a local differential GPS system.

After these two launches we will begin a parallel development effort to:

Develop a steerable hybrid motor system, probably using liquid injection thrust vector control.
Further tune the navigation algorithms in the avionics system through simulations and experiments with ducted fans.

By late 2004 or early 2005 we hope to begin the integration of these two systems and begin the first of many test flights with an active guidance system in place. Instead of trying to fly a fully actively guided rocket in one step, we plan to continue our careful, incremental approach to development. For example, by increasing the gain of the control system while removing fin area, we can build confidence in the navigation system as the rocket becomes more and more unstable because of the lack of fins.

By late 2005 or early 2006 we hope to have an actively guided, possibly staged, rocket lifting research projects to high altitudes, if not to the edge of space.

10 Conclusion

The PSAS software system employs integrated computing, from 8 bit and 32 bit microcontrollers through common PC laptop and server hardware. The system hosts a wide variety of visualization, device control, software development, promotional and coordination roles. Data flows through real-time serial and parallel buses as well as wireless links, requiring only straightforward programming based on network abstractions. Required tasks range from hard real time signal processing in a preemptive multitasking environment to coordinating a caravan of two dozen people to the Nevada desert. Our budget for development equipment is quite small, and our team of contributors vary widely in skill level and available time.

With minor exceptions, free software has helped us tackle all of these technological and logistical problems. Commercial systems could not have scaled so widely, cost so little, offered so many choices, offered free support and training, and integrated these many domains together so well. Our group could not have grown as quickly, or achieved our current level of sophistication, without free software and open hardware.

Availability

For complete technical details on our projects, including launch videos, technology overviews, white papers, source code, schematics, and board designs, please visit http://psas.pdx.edu/.

Acknowledgements

Throughout this paper, we have acknowledged the important debt we owe to developers of free software and previous amateur aerospace groups who have made our project possible. We are extremely grateful for the invaluable encouragement and financial support from Portland State University's Computer Science and Electrical and Computer Engineering Departments. We would like to thank the IEEE, AT&T and the NASA/Oregon Space Grant, all of which have provided significant financial support; also the companies which have donated materials and software such as HI-TECH Software, ANSOFT, and Microchip. Finally, we would like to thank Usenix for supporting our efforts to bring free software to the amateur aerospace community, and Keith Packard for providing excellent counsel on our paper.

References

[alt]: John Coker's Rocket Pages: Altimeter Comparison. Web document. URL http://www.jcrocket.com/altimeters.shtml accessed April 2, 2003 07:00 UTC.
[ama]: Amanda, The Advanced Maryland Automatic Network Disk Archiver. Web document. URL http://www.amanda.org/ accessed April 6, 2003 10:00 UTC.
[arr]: Amateur Radio Relay League High-Speed Digital Networks. Web document. URL http://www.arrl.org/hsmm/ accessed April 6, 2003 10:00 UTC.
[C18]: HI-TECH Software PICC-18 Compiler. Web document. URL http://www.htsoft.com accessed April 6, 2003 23:00 UTC.
[can]: Robert Bosch GmbH CAN home page. Web document. URL http://www.can.bosch.com/ accessed April 6, 2003 10:00 UTC.
[CC99]: Charles K. Chui and Guanrong Chen. Kalman Filtering. Springer-Verlag, 1999.
[csl]: United States Code, Title 49 Transportation, Chapters 701 and 703. Web document. URL http://uscode.house.gov/title_49.htm accessed April 7, 2003 06:19 UTC.
[cvs]: CVS. Web document. URL http://www.cvshome.org/ accessed April 6, 2003 09:30 UTC.
[deb]: Debian. Web document. URL http://www.debian.org/ accessed April 6, 2003 09:30 UTC.
[FB99]: Jay A. Farrell and Matthew Barth. The Global Positioning System and Inertial Navigation. McGraw-Hill, 1999.
[fsm]: FSMLabs - The RTLinux Company. Web document. URL http://www.fsmlabs.com/ accessed April 6, 2003 10:00 UTC.
[Gal95]: Bill O. Gallmeister. POSIX.4: Programming for the real world. O'Reilly and Associates, Inc., 1995.
[gpsa]: GPL-GPS home page. Web document. URL http://gps.psas.pdx.edu/ accessed April 6, 2003 23:00 UTC.
[gpsb]: Zarlink Semiconductor GP4020 GPS Receiver Baseband Processor. Web document. URL http://products.zarlink.com/product_profiles/GP4020.htm accessed April 6, 2003 10:00 UTC.
[ins]: Ariane 5 - Flight 501 Failure; Report by the Inquiry Board, European Space Agency. Web document. URL http://www.ima.umn.edu/~arnold/disasters/ariane5rep.html accessed April 7, 2003 03:00 UTC.
[jfr]: JFreeChart. Freely available source code. URL www.jfree.org/jfreechart/index.html accessed April 7, 2003 00:48 UTC.
[kel]: OpenSource GPS Project. Web document. URL http://home.earthlink.net/~cwkelley/ accessed April 6, 2003 23:00 UTC.
[Lab99]: Jean J. Labrosse. MicroC/OS-II: The Real Time Kernel. Miller Freeman, 1999.
[lcb]: Linux CAN-bus Driver for the Intel(c) 82527 and Philips sja1000 controllers. Web document. URL http://home.wanadoo.nl/arnaud/, accessed April 7, 2003 07:21 UTC.
[mai]: Mailman, the GNU Mailing List Manager. Web document. URL http://www.list.org/ accessed April 6, 2003 09:30 UTC.
[maj]: Majordomo. Web document. URL http://www.greatcircle.com/majordomo/ accessed April 6, 2003 09:30 UTC.
[mar]: MARS Amateur Rocketry Group. Web document. URL http://www.mars.org.uk/ accessed April 6, 2003 23:00 UTC.
[Mas03]: Anthony J. Massa. Embedded Software Development with eCos. Prentice Hall Professional Technical Reference, 2003.
[pic]: Microchip, inc. Web document. URL http://www.microchip.com/ accessed April 6, 2003 23:00 UTC.
[rrs]: Reaction Research Society Space Shot. Web document. URL http://www.rrs.org/Projects/Launches/Space_Shot/space_shot.html accessed April 2, 2003 07:00 UTC.
[SB00]: George P. Sutton and Oscar Biblar. Elements of Rocket Propulsion. Wiley-Interscience, December 2000.
[sls]: Spacelawstation.com: U.S. Space Law. Web document. URL http://www.spacelawstation.com/uslaw.html accessed April 7, 2003 06:20 UTC.
[ssh]: OpenSSH. Web document. URL http://www.openssh.org/ accessed April 6, 2003 09:30 UTC.
[Tri]: Tripoli Rocketry Association. Motor size classifications. Web document. URL http://www.tripoli.org/motors/motor_classes.html accessed April 3, 2003 19:00 UTC.
[twi]: TWiki - A Web Based Collaboration Platform. Web document. URL http://www.twiki.org/ accessed April 6, 2003 09:30 UTC.
[un]: United Nations Office for Outer Space Affairs; International Space Law. Web document. URL http://www.oosa.unvienna.org/SpaceLaw/spacelaw.htm accessed April 7, 2003 02:00 UTC.
[Wer92]: James R. Wertz, editor. Space Mission Analysis and Design. Microcosm, Inc., October 1992.
[Win83]: Frank H. Winter. Prelude to the Space Age: The Rocket Societies 1924-1940. Smithsonian Institution Press, 1983.

This document was translated from L^AT_EX by H^EV^EA.