Satellites and antennas 
  Satellite Network
  Specifications

On Mars Communication
  Radio Communications
  Radio Specifications

LMR Software and Control

DSP and Noise Correction

 Basic Specifications

The basic design will be based on the Mars Sojourner robot launched in 1996, with extensive modifications in the area of autonomy, artificial intelligence, and data processing.

It is assumed that the rover will generate about 16 W of power with solar cells, and will carry a battery-powered transponder to allow it to be found if the power shuts down.

Currently, we hope to improve the data transfer rate from the 9.6 Kbps on Sojourner’s UHF modem to 100Kbps between the LMR and the medium-sized rover.

Also, the processor will probably be around 100 to 300 MHz Pentium as opposed to the 8085-architecture processor on Sojourner.

So far the task subdivision between the LMRs, the larger rovers, and the base will be as follows:

• The LMRs will perform only the most basic data analysis for local navigation and path planning. The raw photographic data will be transmitted in bursts to the larger rovers. The LMR’s computer will run a data protection algorithm while transmitting, and it will also be responsible for building a height map of the immediately surrounding terrain and choosing a local path to avoid obstacles. Commands from the large rover will be uplinked to the LMR, and the computer will use those commands to plan its activities on a larger scale. However, it will have some autonomy in terms of the local navigation and motion so that it will move in a convoy with other LMRs while choosing its own path. This has the advantage of allowing it to be of some use even if communication with the central rover is temporarily interrupted.
• The medium-sized rover will lead the convoy and its computer will receive commands directly from the base through a system of relays. It will have much more power available for its computer and will handle the decision-making process of choosing small tasks for its LMRs. It will also receive raw data from the LMRs and transmit it to base using a data protection algorithm. Most importantly, it will handle the job of recognizing potentially interesting objects found by the LMRs. It will have the advantage of coordinating multiple LMRs to investigate “discoveries”. While the preliminary data processing for this will be done on the rover, the information will be sent back to base for a more extensive analysis.
• The base computers will perform the bulk of the data processing. They will use an appropriate interface with human operators, such as RCW in conjunction with the ASPEN planning and scheduling system (see below). The base will receive the photographic and telemetry data from all the rovers, and its computers will run analyses of the data combined from multiple sources. The task distribution between the rovers and LMRs will be handled automatically by a system based on the ASPEN concept that will use distributed net processing. This system will take the load off human operators by taking general goals and directives from humans and automatically generating conflict-free schedules of very low-level commands for the rovers. The execution of the commands, however, such as the navigation algorithms, will be left to the rovers. In addition, this system will have the ability of a human override in case of an emergency to allow an operator or engineer to have complete access to and control of each rover and LMR.

On a large scale, navigation will be determined by the larger rovers as they will calculate the best general path to the assigned destination waypoint. The path will be followed by the whole convoy. However, the navigation of the LMRs will be left to their own computers. The path planning algorithm will use data from stereo images from the rovers, satellite terrain maps, and a close-range laser hazard detector. In general, the terrain map will be based on the FIDO (Field Integrated Design and Operations) concept developed by the JPL for autonomous sample-return missions.

The navigation algorithms running on all the rovers allow a significant decrease in the need for low-level commands to be sent to the rovers for navigation. In particular, one may expect the rover to handle:

• Long distances between assigned waypoints, on the order of 100 m per command.
• Ability of convoys to navigate over the horizon.
• Use of panoramic images for navigation as well as sample selection.
• Relatively accurate automatic localization of the rover’s position (an error range of 3-5m).
• A continuous driving system that can perform path planning and localization while it moves and performs other scientific tasks. Extended operations will be permitted, lasting up to 20 hours of continuous work.

In order to facilitate long, autonomous travel, new algorithms for processing stereo data and generating local terrain maps have been devised that improve calculation speeds five-fold. This enhancement, along with the more powerful processor to be used on each LMR, will allow it to drive and steer continuously while calculating the optimal path as it moves.

Each new waypoint for the LMR will be on the order of 100 m distant. At each waypoint, the rover will use its stereo camera to obtain a panoramic picture of its surroundings spanning 120 degrees. The edge of the envelope of the image will be at a radius of approximately 20 m. The rover’s on-board computer will then combine the stereo images into a “mosaic” that shows the same terrain from different angles. The computer will create a composite height map from the different images, and, finally, use the information from the height maps to generate a three-dimensional terrain map that the path-planning algorithm can use to avoid obstacles and find its way to the next waypoint. Ideally, this process should be continuous and the image processing should be done as the rover drives on. While the rover is steering, it will use its more low-level obstacle avoidance mechanisms, such as laser detectors, to avoid running into obstacles in its immediate proximity that were unseen earlier. The path to each waypoint may be subdivided into small segments of about 5 m between subsequent stereo images. Localization may be performed at each step by taking a backward image and analyzing the relative motion of landmarks identified earlier. The waypoints will be determined by the central large rover and will correspond to the general path of the convoy.

The activities, scientific experiments, and other actions that the LMR is to perform durings its travel will be sent out as commands from the large rover. However, the LMR will have some degree of autonomy in responding to emergencies or unforeseen situations by changing its planned activities using an artificial intelligence planning and sequencing system known as CASPER (Continuous Activity Scheduling, Planning, Execution, and Replanning). This algorithm will be running on the LMR’s computer, and will receive general commands from the medium-sized rover leading the convoy. The commands may include activities to perform, scientific experiments to conduct, samples to search for and extract, and interactions with other LMR. The CASPER system will generate a sequence of low-level commands that can be directly fed to the rover’s computer. However, if unexpected events (emergencies, unforeseen discoveries of interesting samples, unplanned obstacles, etc.) occur, the CASPER system will use an algorithm known as iterative repair to autonomously decide the best course of action. This algorithm sequentially resolves conflicts in the LMR’s planned operations and the dynamically changing resource constraints (such as an unexpected power failure), rover state constraints, and other scientific requirements in a way that does minimal damage to the ability of the rover to fulfill the goals sent to it by the medium-sized rover. The CASPER algorithm responds dynamically to changes in the rover’s environment by continuously replanning the rover’s actions. For instance, if, on the way to its next waypoint, the rover’s stereo cameras detect an unforeseen, large obstacle that was not shown in satellite maps of the area, the algorithm will automatically begin to follow the boundary of the obstacle. If this will lead the rover closer to sites that were planned for much later in the rover’s schedule, the CASPER algorithm will automatically rearrange the order of the rover’s activities to use this opportunity to perform some experiments ahead of time.

The CASPER system will include a dynamic path-planning algorithm that will allow the rover to navigate towards its waypoint while being able to handle unforeseen obstacles. The algorithm is specifically tailored to minimize the amount of unnecessary motion the rover needs to perform (to save energy), and it is designed to be sufficiently parsimonious in its use of computing resources to be able to run locally on the rover’s computer. Taking a terrain topographical map based upon satellite imagery and the rover’s stereo cameras as input, this algorithm will model the obstacles as polygons and generate a sequence of steps of two major types: motion towards the goal (waypoint) and motion following the boundary of a large obstacle. Typically, the rover is in the mode of moving towards the goal while avoiding small obstacles, and the motion is updated at each step by the algorithm as new data arrives from the stereo cameras. When large obstacles are encountered, the rover will switch to following their boundaries to arrive at a place where the path planner can find a new path towards the goal.

The algorithm is essentially the same as the RoverBug algorithm that the JPL is developing for the 2003 and 2005 sample return missions that will necessarily require a high degree of rover autonomy since the missions will be unmanned. The actual path-planning algorithm relies on the terrain map of the surrounding area to create a local tangent graph (LTG) in the visible region of ground. The tangent graph is a set of line segments that are tangent to all the obstacles at given vertices and that ultimately connect the rover’s current position to its destination (which may be a locally-generated segment of the path to the waypoint). Each line segment extends from one obstacle vertex to the vertex of another obstacle in such a way that at each vertex the line segment touches the vertex but does not traverse the inside of the obstacle. At each vertex of a particular obstacle, numerous line segments originate and run to the vertices of neighboring obstacles, without going through the interior of any obstacle. The set of all such line segments in the visible region is the local tangent graph, and the problem of navigating towards the goal (in the cases where a route of approaching the goal exists and boundary-following is not necessary) is reduced to the problem of calculating the shortest path from the rover to the goal on the local tangent graph. By the way the graph is constructed, all paths lie outside the obstacles, and so the rover can physically follow them. In fact, the rover’s computer calculates the shortest path to the goal and the rover then follows that path. (This calculation is very similar to solving the mathematical Traveling Salesman Problem in which the shortest path between points on a graph must be determined.) When the edge of the local terrain map is reached, the process is repeated.

However, there are situations in which the above-described algorithm will fail. In particular, if the rover encounters a large obstacle that completely blocks its path to the goal and for which a complete map is unavailable, the local tangent graph will not contain a path that will sequentially take the rover closer to its goal. Thus, the rover must use another technique to overcome this problem, namely boundary circumnavigation. The rover will switch to this navigation mode if it encounters an obstacle around which there is no path in the local tangent graph. The rover’s computer calculates the shortest path around the boundary of the object in order to avoid such unnecessary waste of power and time as tracing out the complex indentations of the obstacle. Generally, the trajectory of the rover should be a convex hull around the object, and, in addition, the path-planning algorithm should continously search for shortcuts. As soon as this technique has placed the rover at a spot where its new LTG’s have paths that take the rover closer to its target, the boundary circumnavigation will cease and the rover will revert to following the shortest path in its LTG. In some cases, the rover may return to its starting place in a loop around an obstacle, which means that the obstacles surround the target in such a way that it is unreachable. Otherwise, it may be proven that the alogrithm will generate optimal paths to the target given the limitations of the knowledge of the obstacles in the rover’s vicinity.

The path-planning and navigation algorithm can be very effective, but it depends very heavily on knowledge of the rover’s exact position in relation to its surroundings at all times. The simplest method of counting the revolutions of the wheels and the steering after some intial position can accumulate errors very quickly due to wheel slipping and other inaccuracies. Thus, a reliable localization technique is necessary to periodically fix the rover’s position in its map. This localization is performed by a separate algorithm that uses the above-described method as a first estimate. It then compares that estimated position and the local terrain map to stored terrain maps of the surrounding area. It recursively examines possibilities of how the rover could be located to generate the terrain map around as in fact observed and quickly zeroes in on the rover’s true position by comparing the positions of the obstacles around the rover from its point of view to the terrain map that is stored in memory.

LMR Coordination and Cooperation

The distribution of tasks between the various rovers in the convoy will be handled automatically, with the possibility of an emergency human operator override if communication with the rovers is maintained. The scientific goals and objectives are provided by humans at the base using such tools as RCW (Rover Control Workstation). The central rover will use its artificial intelligence capabilities to continually generate the best plan for the accomplishment of the assigned goals. The software used for actually implementing the decision-making system will most likely be a derivative of the ASPEN (Automated Scheduling and Planning Environment) system, which is a framework for using all information at the rover’s disposal to automatically update its plan of low-level steps towards accomplishing its goals. The ASPEN system uses an iterative repair algorithm, as described above, to resolve conflicts between various scheduled activities and constraints posed by the environment and by the state of the rover. However, there are several different possible modes of organizing the autonomous decision-making process and splitting responsibilities between the centralized main computer and the less powerful computers on the LMRs. The approaches include a completely centralized one, a distributed processing method, and a method based upon the “auction” model of completely localized planning by competing LMRs.

• The traditional approach utilizes a single centralized planner. This is an algorithm that resides either on the base’s computers or on the large rover and gathers raw data from all the LMRs, which it processes to produce low-level commands that are sent back to the LMRs. This approach has the advantage of minimizing the amount of processing that needs to be done by the LMRs, whose computing resources are naturally very limited. In addition, this approach is generally the simplest conceptually, since all conflicts in a tentative plan of execution are resolved in one place and all possibilities can be quickly checked. However, it has several major disadvantages. First, it requires a large amount of raw data (images, unsorted instrument readings, etc.) to be transmitted continuously by the LMRs, since they will not have the ability to analyze it on their own. Such a transmission rate may well exceed the capabilities of the communications system on the LMRs, which has its own severe power limitations. More importantly, the LMRs would be completely dependent upon the functionality of the central rover’s computer and the maintenance of a reliable communications link. Any disruptions in communication with the LMRs, as are likely to occur in a real mission, would leave the LMRs completely disabled and unable to perform any independent activities.
• At the other extreme is what is known as a contract net protocol. In this case, the central planner generates a set of goals to accomplish. However, the planning process of accomplishing the goals is completely moved to the LMRs’ computers. Most importantly, the method of distributing the tasks to the LMRs will be based on a bidding system where each LMR is given a set of tasks and then information on its own constraints and position to calculate the efficiency with which it could perform the given tasks. The centralized planner then simply assigns each task to the rover that gave the highest “bid”, or prediction of its ability to carry out that particular task. The generation of lower-level commands is done entirely by the rovers’ computers. A major advantage of this system is its flexibility in responding to new environments individually for each LMR. Each LMR operates basically independently, which makes them less vulnerable to failures of the centralized computer and does not overburden the communications system. It also takes into account individual factors concerning the states of the LMRs in deciding  the distribution of tasks. However, it requires the on-board computers on the LMRs to conduct much of the data analysis and processing for generating a sequence of low-level commands from its assigned high-level objective. Since, the computers on the LMRs are severely constrained by the limited power supply and other factors, this approach may not be practical on the mission.
• A third methodology uses an intermediate approach to the task distribution problem, known as the distributed planning approach. Like the centralized approach, this algorithm uses a central computer to generate the overall goals and objectives for the rovers to accomplish. This is based on data collected from all the rovers and takes into account the variations in the states of the rovers and the resources available to them. The accomplishment of each goal, however, is left to each individual rover. In other words, each LMR will be assigned a general goal (in terms of navigation, scientific experiments, etc.) and it will use its own computing resources to find the best way to accomplish that goal. This approach combines some of the advantages of both the other approaches. On the one hand, centralized global decision-making for large tasks does not overburden either the communication or the computing resources of the LMRs. On the other hand, this model gives each LMR enough autonomy to respond flexibly to unanticipated changes in its state or environment and automatically replan its activities. Also, the LMRs will not be completely disabled by loss of their link to the central rover or the base. The local decision-planning and navigation that will be conducted by each LMR will still be computation-intensive, but not unreasonably so given the significant improvement in the computing power that will be allocated to each LMR. Thus, this decision-making model seems to represent the optimal choice for most instances during this mission. The LMRs will be given autonomy in performing their own navigation, which will allow them to separate from the convoy to perform separate scientific tasks, and to operate in conditions where communication with the convoy or the base may not be reliable. At the same time, the centralized goal-choosing process will make the whole procedure more transparent to human operators, who may choose to interfere and reassign tasks at any time. The major data analysis will still be done by the central computer, based on the minimally-processed data sent by the LMRs.

References

R. Sherwood, T. Estlin, S. Chien, G. Rabideau, B. Engelhardt, A. Mishkin, B. Cooper , "An Automated Rover Command Generation Prototype for the Mars 2001 Marie Curie Rover," SpaceOps 2000, Toulouse, France, June 2000.

S. Chien, G. Rabideau, R. Knight, R. Sherwood, B. Engelhardt, D. Mutz, T. Estlin, B. Smith, F. Fisher, T. Barrett, G. Stebbins, D. Tran , "ASPEN - Automating Space Mission Operations using Automated Planning and Scheduling," SpaceOps 2000, Toulouse, France, June 2000.

S. Chien, A. Barrett, G. Rabideau, T. Estlin, "Three Coordinated Planning Methods for Cooperating Rovers," Third International  Symposium on Intelligent Automation and Control, World Automation Congress, Maui, Hawaii, June 11-16, 2000.

R. Volpe, T. Estlin, S. Laubach, C. Olson, B. Balaram, "Enhanced Mars Rover Navigation Techniques," IEEE International Conference on Robotics and Automation, San Francisco, CA, April 2000.

T. Estlin, G. Rabideau, D. Mutz and, S. Chien, "Using Continuous Planning Techniques to Coordinate Multiple Rovers," IJCAI99 Workshop on Scheduling and Planning meet Real-time Monitoring in a Dynamic and Uncertain World, Stockholm, Sweden, August 1999.

T. Estlin, T. Mann, A. Gray, G. Rabideau, R. Castano, S. Chien and E. Mjolsness, "An Integrated System for Multi-Rover Scientific Exploration," Sixteenth National Conference of Artificial Intelligence (AAAI-99), Orlando, FL, July 1999.

G. Rabideau, R. Knight, S. Chien, A. Fukunaga, A. Govindjee, "Iterative Repair Planning for Spacecraft Operations in the ASPEN System," International Symposium on Artificial Intelligence Robotics and Automation in Space (ISAIRAS), Noordwijk, The Netherlands, June 1999.

T. Estlin, S. Hayati, A. Jain, J. Yen, G. Rabideau, R. Castano, R. Petras, S. Peters, D. Decoste, E. Tunstel, S. Chien, E. Mjolsness, R. Steele, D. Mutz, A. Gray, T. Mann, "An Integrated Architecture for Cooperating Rovers," International Symposium on Artificial Intelligence Robotics and Automation in Space (ISAIRAS), Noordwijk, The Netherlands, June 1999.

P. Backes, G. Rabideau, K. Tso, S. Chien, "Automated Planning and Scheduling for Planetary Rover Distributed Operations,"
Proceedings of the IEEE Conference on Robotics and Automation (ICRA), Detroit, Michigan, May 1999.

Author: Artem Gleyzer (agleyzer@mit.edu)