]>
Towards a Better GPS Protocol Eric Steven Raymond Thyrsus Enterprises
esr@thyrsus.com
2004 Eric S. Raymond 1.0 04 January 2005 esr Initial draft. 1.1 11 February 2005 esr Corrected SGPS example, thanks to Kevin Niehage for the bug report. 1.2 25 April 2005 esr Specify UTC. Fix time-uncertainty units. Vertical course angle changed to climb/sink rate. 1.3 16 November 2006 esr Added GPSVID and GPSGP, changed to mandate ISO8601 dates. 1.4 21 November 2006 esr Fixed timestamp to Zulu time. Specified signed latitude. 1.5 25 February 2009 esr Fixed GPSGP so the response isn't identical to the send, avoiding problems on multidrop lines. Added SGPS revision field to $GPVID. Went back to requiring checksums, because you just know it was going to bite someday otherwise. Changed sentence length limit to 255. The NMEA 0183 protocol used by GPS units might best be described as "layer upon layer of cruft". In this paper, we examine the problems and consider a cleaner design.
What's Wrong with NMEA 0183, and Why Fix It? The protocol used by GPS devices to report to computers is a small subset of NMEA 0183. NMEA stands for "National Marine Electronics Association", and the features GPSes use for reporting time/position/velocity information are a small part of a protocol originally designed for communication between parts of complex marine navigation systems. Thus the full protocol includes support for depth sounders, LORAN, and many other things irrelevant to a modern GPS. The lowest level of NMEA 0183 is quite sensibly designed. The protocol consists of sentences, each led by a dollar sign and an identifying text tag, followed by multiple comma-separated textual fields, ended by an asterisk, a checksum, and LF/CR. This is a simple, clean format with good extensibility, easy to parse and generate. It is well adapted to its job, which is to pass small amounts of numeric and status information. The textual format makes it easy to log NMEA sessions, edit them, and play them back — a substantial advantage in developing talker and parser software. Unfortunately, the good news ends there. The design of the upper layers of NMEA 0183 is patchy, kludgy, and replete with the kind of errors that arise from growth by accretion rather than forethought. Here are some of the more obvious problems: NMEA timestamps usually (e.g in GPBWC, GPBWR, GPGBS, GPGGA, GPGLL, GPGXA, GPTRF, GPZTG) report time-of-day only. The exceptions (GPRMC, GPTRF, GPZDA) report only two digits of year and no century. Time precision is unspecified, usually to the second though some devices report a fractional decimal part to millisecond precision. It is not possible to get a time/position/velocity report in a single sentence. Some sentences (GPRMC) report time and 2D position and velocity, some (GPGGA) report time and 3D position, some (GPVTG) report velocity only. As a result, the API for a protocol client is complicated by the necessity of maintaining separate age indications for 2D position, 3D position, and velocity, NMEA sentences have at least three kinds of validity indicators — mode (GPGSA only), status (GPGLL, GPGGA), and the Active/Void field (GPRMC, GPGLL). And that's before we get into the FAA extensions in late revisions of the protocol. Interpreting these status bits is a black art involving knowledge of undocumented and often vendor-specific quirks. There is no standard way of indicating that part of a time/position/velocity report is invalid (e.g. because the device does not have a fix of the required quality). Many devices simply report 0 for invalid fields, which doesn't sound bad unless you're near the zero-zero point in the Bay of Benin — or at sea level. It is also not generally possible to distinguish between information a GPS is not yet reporting but will return on a good fix (such as altitude) from information it will never report (such as, on many units, local magnetic variation). As least one messy bit in NMEA 0183 was an adaptation to machines with only small amounts of expensive RAM: the fact that satellite status may show up in a sequence of as many as three sentences to be processed serially. On modern machines, RAM buffers are cheap. It makes more sense to ship a single long sentence and decrease code complexity in the receiver by not requiring a stateful parser. Position accuracy estimates are not easy to compute from NMEA reports. Reporting a measurement without giving its 95% confidence interval is bad practice. For modern GPS devices, even the small piece of NMEA directly concerned with GPS capabilities is seriously over-complex. Whereas older GPS devices included elaborate facilities for waypoint tracking and navigational computation, newer ones are designed under the assumption that they are connected to a general-purpose computer that is more powerful and flexible at these things; thus, the GPS only needs to be a time/position/velocity oracle. NMEA 0183 is in general very loosely specified and poorly documented. Its problems are compounded by the fact that it is a proprietary specification, jealously guarded by IP lawyers. As a result of these problems, implementing NMEA 0183 talker software is far more complex than need be, and the protocol tends to introduce latencies that vary in an unpredictable way. The latter is a serious problem for aviation and other high-precision GPS applications, and probably provided a technical reason that one major GPS vendor (Garmin) dropped NMEA 0183 support entirely in 2004 in favor of a tighter binary protocol (we refrain from speculating on other less creditable motives for this move). How To Do Better The critique above immediately suggests several ways to improve a protocol for GPS reports; Keep the low-level syntax, because it's not broken. It has all the advantages of textual protocols. Going to a more tightly-packed binary format might look attractive at first glance, but the gain in information would be marginal at best. Textual formats already use 7 out of 8 bits per byte and encode variable-length numeric fields more efficiently than binary; also they avoid endianness issues. Add to the syntax standard ways of indicating that either (a) the GPS cannot now ship valid data for the field, or (b) the GPS will never ship data for this field. Include a full timestamp, to millisecond precision or better, with every sentence. Every timestamp should be in the same standard form and should include a full date with century. Report the uncertainty corresponding to a 95% confidence interval on a standard normal distribution for each measurement field. Design the protocol around a single core sentence that reports time/position/velocity and the uncertainties in same. Make it an objective of the design for an informal specification to fit on a single page. Informal specification: SGPS Here, then, is a proposed informal specification for SGPS, Simple GPS Protocol. The protocol consists of sentences, each led by a dollar sign and an identifying sentence tag, followed by multiple comma-separated textual fields, ended by an asterisk, a CRC32 checksum, and LF/CR. Sentences are at most 255 characters long, counting the trailing CR/LF. A field that is empty indicates that the talker does not have valid information for this field but promises to report it in the future. A field consisting of a question mark (?) indicates that the talker does not expect to ever ship valid information for this field. The first field of every SGPS report sentence is a full timestamp in the format of the W3C profile of ISO 8601, with the timezone required to be Zulu (UTC) time. GPTPV The core sentence of SGPS has the following layout: The sentence tag is GPTPV, standing for Time/Position/Velocity. The first field is the required timestamp. The second field is the uncertainty of the timestamp in (fractional) seconds. The third field is signed latitude in degrees. Encoding must be decimal degrees, not degree/minute/second. The fourth field is signed longitude in degrees. Encoding must be decimal degrees, not degree/minute/second. The fifth field is horizontal uncertainty in meters (95% confidence). The sixth field is altitude in meters. The seventh field is vertical uncertainty in meters (95% confidence). The eighth field is speed over ground in meters per second. The ninth field is speed-over-ground uncertainty in meters per second (95% confidence). The tenth field is course over ground in degrees from true north. The eleventh field is uncertainty of course over ground in degrees (95% confidence). The twelfth field is climb/sink in meters per second. The thirteenth field is uncertainty of climb/sink in meters per second (95% confidence). The fourteenth field is an FAA mode indicator. These fourteen fields completely describe the position and velocity of an object and the associated uncertainties. The FAA mode field is added to satisfy a U.S. regulator's requirement. Here is an example: $GPTPV,2005-02-11T04:40:51.231Z,?,49.45,-123.12,2.3,70.1,52.0,01.0,02.1,23.1,0.6,,,8,A*31 2005-02-11T04:40:51.231Z, Time (Feb 11 04:40:51 EST 2005) ?, Timestamp uncertainty will never be reported 49.45, Latitude (- sign indicates latitude south) -123.12, Longitude (- sign indicates longitude west) 2.3, Meters of horizontal uncertainty of position 70.1, Altitude, meters above sea level 52, Uncertainty of altitude 0.01, Speed, meters/sec 0.02, Speed uncertainty 23.1, Course over ground relative to true North 0.6, Course uncertainty in degrees. , Climb/sink not reported , Climb/sink uncertainty not reported A FAA mode indicator A (Auto). 31 Checksum. GPSVU A second sentence describes GPS satellite status. The sentence tag is GPSVU, standing for Satellite View Update. The first field is the required timestamp. The second field is a count of satellites. The remainder of the sentence fields are groups of four, one for each predicted position of a visible satellite. Each group has the following four elements: The PRN or satellite ID. Elevation in degrees Azimuth, degrees Signal-to-noise ratio in decibels. If this satellite was used in the last fix, suffix this field with a '!'. Here is an example: $GPSVU,2005-02-11T04:40:51.231Z,11,03,03,111,00,04,15,270,00,06,01,010,00,13,06,292,00,14,25,170,00,16,57,208,39!,18,67,296,40!,19,40,246,00,22,42,067,42,24,14,311,43!,27,05,244,00*40 GPVID A third sentence identifies the device. It is GPVID for Version ID, and the fields are as follows: The sentence tag is GPVID, standing for Vendor ID. The first field is the required timestamp. The second field is the SGPS revision level. The third field is the vendor name. The fourth field is the device name or model number. The fifth field is a chipset designation. The sixth field may be empty or a subtype ID, typically a firmware revision level. All fields must consist of US-ASCII text not containing commas. The total length of the sentence must not exceed the old NMEA maximum of 82. Here is an example: $GPVID,2006-11-17T12:29:37Z,1.0,Haicom,H204S,SiRF-II,231.00.00*5C GPGSP With the addition of a fourth sentence, $GPSGP, transition to the new protocol would be easy. It would have two forms: $GPSGP,1: directs the receiver to emit GPPVT and GPSVU only, if it is not already doing so. $GPSGP,0: directs the receiver to return to NMEA-classic mode, if it is capable of doing so. Example: $GPGSP,1*4E An SGPS-conformant receiver is required to respond with $GPSGP,timestamp,x,y where x is 1 or 0 reflecting the command, and y is 1 or 0 reporting its new mode. Other listeners can distinguish GPGSP responses from requests by checking whether field 1 contains an IS8601 timestamp; an easy way to check this is to look for the trailing Z. Other considerations Finally, SGPS-compliant receivers are required to respond to the requests $GPPVT, $GPVSU, $GPVID, and $GPGSP (without arguments) with the corresponding report based on most recent available data. Could This Be Adopted? Astute readers will already have noted that the SGPS sentences might be sold as a minor extension to NMEA 0183. first supplementing and eventually obsolescing the half-dozen or so sentences emitted by most modern GPSes. The only fields reported in the SGPS set that cannot be trivially derived from data already computed for NMEA reports are (a) Climb/sink, and (b) GPSTPV uncertainty fields. None of these should be difficult to derive.