«The Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 May 13, 2005 Submitted in partial fulfillment of the requirements for the ...»
This is strong evidence that TEMPEST plans enable battery charge management, and that deviation from the plan resulted in severe battery discharge and other sub-optimal effects like camera sun blinding. It also suggests an instability in the control of sun-synchronous routes. As the rover gets further behind, the collection of solar energy is progressively more difficult to achieve, and the required speed to catch up to the plan schedule, and hence required locomotion power, increases. If the robot is required to get back on schedule before completing a full circuit, schedule delays of a certain duration will be unrecoverable. At some point, the power required for driving and survival outweighs the incoming solar power, and leads to a continual discharge of batteries. The batteries must sustain the recovery for its duration, or be fully discharged. Alternatively, if the rover is not required to catch up to the plan schedule before the completion of the circuit, the recovery speed could be adjusted to enable the system to sustain itself. Lastly, if shadowing is not expected to interfere with lighting, the vehicle could potentially stop and rotate in place to follow the sun in a survival holding pattern, and reacquire the sun-synchronous circuit on the following day.
5.7 Discussion The sun-synchronous field experiments plainly illustrated the utility of mission-directed path planning. Sun-synchronous planning would have been difficult for humans to do by hand. Large-scale terrain obstacles prevented an idealized, circular sun-synchronous navigation path. Determining the timing to achieve the optimal sun angle balance on the irregular path would have entailed extensive trial-and-error, or a hierarchical approach that first selects the route between Via Point goals and then searches over different schedules on the route solution. Mission-directed path
planning provided a number of notable benefits for rover operations:
Sophisticated Reasoning: TEMPEST reasoned about how best to time the route, despite the elongated shape of the traverse. It minimized the overall effect of inevitable solar array mispointing by balancing the lead and lag of the sun with respect to the solar array at various points in the traverse. The plan for Experiment 1 proved highly effective in maintaining battery energy during the entire traverse. Executing the plan closely maintained the batteries well above the minimum bus voltage for the entire 24 hours. More interestingly, Experiment 2 illustrated the danger of substantial deviation from the TEMPEST-derived plans. Operational delays allowed the sun to be biased well to the front of
the rover, and caused the batteries to discharge below the nominal bus voltage. Re-acquiring the plan later in the mission allowed the batteries to re-charge to initial voltages.
Use of Intermediate Via Points: Intermediate Via Point goals enabled human operators to steer the route away from terrain hazards that were not represented on the low-resolution elevation model. Furthermore, the designation of intermediate goals is analogous to hierarchical approaches - prior path selection to enable time-optimal trajectory planning as in . However, by selecting points rather than paths, TEMPEST was far more free to consider the coupling between route, timing and resources in selecting a plan.
Natural Integration with Local Navigation: The cooperation and mutual abstraction between TEMPEST and the Local Navigator was natural. The Local Navigator module reliably avoided rocks while moving toward TEMPESTdefined waypoints. It had no knowledge of the global path, resource usage, or timing requirements. Meanwhile, TEMPEST solved for the mission-directed plan to avoid large-scale obstacles and to manage resources, but had no knowledge of local obstacles below the resolution of the terrain map.
The experiments also highlighted a number of future challenges for TEMPEST.
Brittleness to Unanticipated Conditions: The software provided no capability for planning under uncertainty, contingency planning, or automated re-planning, and was therefore brittle to unanticipated problems. In Experiment 2, a planner capable of reasoning under uncertainty might have anticipated the possibility of schedule delays and biased the plan ahead in time to avoid crippling sun angles. Re-planning might have compensated for the delays by inserting Charge actions to replenish the battery, taking more direct routes between Via Point goals, or electing to eliminate Via Points as necessary to save time to get back on schedule.
Lack of Richness in Mission Specification: At the time of the experiments, TEMPEST was limited in how missions could be specified to it. For example, it only allowed goals to be specified in terms of position, and in the case of the final goal, minimum battery energy. Plans must often meet constraints on battery energy for other goals, or fall within allowable time ranges. Specific activities might also be assigned to goal positions, and might be governed by geometric or other constraints (e.g. goal must be achieved at least one hour before dark). TEMPEST only modeled a few operational constraints for actions, (e.g. maximum slope, minimum sun elevation angle). Other restrictions on line-of-sight to communications relays and enforcing sun-in-camera stayout zones would have been a major benefit for Hyperion.
Poor Computational Performance: This preliminary form of TEMPEST required significant computational and memory resources. Planning in a four-dimensional state space (x, y, time, energy), Experiment 1 took 5 hours 48 minutes to plan on a laptop with a 400 MHz Pentium II with 128 MB of RAM.
Exclusively Offline Operation: TEMPEST was run exclusively offboard Hyperion and in an offline mode for the field experiments. Slow planning and lack of re-planning capability removed the incentive to configure TEMPEST for online operation.
These lessons motivated substantial improvements in domain richness, re-planning, performance and in online operations. Follow-on field experiments on Hyperion and a new rover, Zoe, demonstrate how mission-directed path planning enable far greater navigational autonomy and improve mission planning for rover exploration.
The sun-synchronous navigation planning problem is not yet solved. It would be interesting in future work to develop planning that could specify a sun-synchronous circuit given only high-level goals like target areas for exploration and total traverse distance. It would also be intriguing to quantify the sensitivity of path cost on schedule delays, as a way of bounding the delays from which recovery is possible. Finally, this thesis did not examine the behavior of plans in more desperate situations. For example, would TEMPEST display a “tacking” strategy, as in sailing, to travel in the direction of the sun?
6. Robotic Astrobiology
Astrobiology is the branch of biology concerned with the emergence and survival of life in the universe, the effects of outer space on living organisms, and the search for extraterrestrial life. Since Earth is the only known place to harbor life, a primary activity of astrobiologists is to characterize the extreme habitats on Earth, and to identify the mechanisms used by organisms to survive under conditions that mimic the extremes of other bodies in the Solar System. A second focus is to develop enabling technologies for remote life detection so that future missions to the Solar System will be equipped to find extraterrestrial life if it exists. The Life in the Atacama (LITA) project seeks to develop technology to enable robotic astrobiology for NASA, and at the same time to conduct useful Earth science in the Atacama Desert of northern Chile .
In a vein of the LITA research program, the TEMPEST planner was further developed to support autonomous, widearea scientific investigations. In the Sun-Synchronous Navigation project, TEMPEST demonstrated its capacity to select routes, coordinate route scheduling, and to manage battery energy, albeit in an off-line mode. The LITA project motivates a much higher standard for online operations, richer representations for the planning problem, and planning in support of science goals.
6.1 Life in the Atacama The Atacama is one of the driest places on Earth, and has long been known to support very little life. The LITA project conducted two field experiments, in 2003 and 2004, involving rover field testing and biological investigation, and will culminate in 2005 in a multi-week integrated field demonstration. In that final trial, a team of scientists in the United States will direct a robotic search for life in Chile. To stress robot autonomy, scientists and engineers will have limited communications bandwidth, and will only be allowed to transmit commands and receive telemetry once per day. Their mission will be to characterize the presence and distribution of microscopic life over tens of kilometers of travel in an effort to better understand the limitations of life in the Atacama ecosystem.
6.2 Navigational Autonomy for Science In sharp contrast with the Sun-Synchronous Navigation project, whose principal aim was navigation, the goal of the LITA project is to develop a rover and software that enable remote, autonomous scientific investigation. However, LITA’s scientific approach continues to stress long-distance navigational autonomy. It has baselined a strategy to characterize regional habitats by conducting widely separated detailed surveys and faster periodic surveys along traverses between detailed survey sites. Scientists will analyze data sets from previous days and, by correlating their findings with orbital data of the region, will select new targets to test hypotheses on patterns of life. Most days will be spent traversing a minimum of hundreds of meters and often several kilometers in pursuit of new goals. Navigational autonomy must reliably transport the rover to target sites for this investigation to be meaningful.
The sun geometry for the Atacama desert favors a solar array that is horizontally mounted on a rover. This configuration eliminates the coupling between driving direction and incoming solar energy, which makes the planning problem at once easier and yet less interesting.
The strategy also depends on interleaving science observations with traverses in a single day. These phases cannot be adequately planned independently - resource availability changes throughout the day, as does sun geometry that might have a bearing on navigation. Navigation plans must incorporate the position, time and resource requirements of science activities to ensure that objectives are globally feasible.
Finally, the LITA field demonstration will not tolerate planning once per day. As was shown in the Arctic, simple delays can cause a plan to become infeasible. Rather than forcing the robot to abandon a day’s activities in the event of a schedule delay or other unforeseen event, re-planning to adjust to updates in state might allow the rover to accomplish most or all of its goals.
6.3 Atacama Desert The Atacama Desert contrasts dramatically from the Canadian Arctic of earlier TEMPEST field experiments. At 21 ° S latitude, continual sun was replaced by a day/night cycle closer to common experience. Solar radiation was more intense in the Atacama, however, and provided a flux of roughly 1000 W/m2 during the field experiments in 2003 and
2004. The Atacama is far more dry - streams were completely absent, and compared to the Arctic, cloud cover was rare1, leading to more reliable solar power for rovers. A common element between the Arctic and the Atacama was the rarity of large-scale plant life. The Atacama is a rocky, salty environment that supports life in small pockets. The result is a landscape whose appearance mimics the surface of Mars, a boon for local navigation designed for extraterHowever, surprisingly, humidity levels at the near-coastal sites rose dramatically at night, due to heavy fog banks originating from the Pacific, which sometimes extended well into the daylight hours.
restrial use. The Atacama presents a greater challenge to navigation that the Arctic - the terrain is more varied, and rough terrain is more common. Mid-scale terrain features, too small to be represented in maps deriving from orbital data, yet too large to be perceived by traditional local sensing, are very common.
The next two sections provide results from TEMPEST tests from the 2003 and 2004 field seasons respectively. The objectives for each season were different, and developments between the campaigns resulted in an entirely new robot and vastly updated software for the later season. Therefore, each section introduces the objectives, rover and software used in tests prior to presenting results.
6.4 Field Experiment 2003 6.4.1 Objectives The principal objective with respect to robot autonomy was to enable fully autonomous driving of at least 1 kilometer, in preparation for integrated science experiments in coming years. Where in the Arctic, short manual interventions to send plan actions to the rover were commonplace, the new system had to operate completely without human involvement. Integrating TEMPEST with an executive and a plan monitor was essential to this performance jump, but resulted in a far more complex system. Specific to planning, the objective was to characterize TEMPEST and overall system behavior in terms of plan quality, the reasons behind re-planning, and plan stability.