«The Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 May 13, 2005 Submitted in partial fulfillment of the requirements for the ...»
The degree to which routes are stable affects the degree to which the execution of the plan can be predicted. In a scenario where TEMPEST is used as an offboard planning tool, route stability would enable an engineering team to prevalidate TEMPEST plans without examining a wide range of contingency cases. Figure 6-8a) through f) illustrates plan route stability for the field experiment. Each frame depicts the route stability for a re-planning sequence in pursuit of a single goal, beginning with an initial plan and continuing with a number of re-plans. In each case, the horizontal axis shows the fraction of the re-plan waypoints that are identical to the initial plan (Fi). The vertical axis shows the fraction of the re-plan waypoints that are identical to those from the previous plan (Fp). The markers on the traces correspond to results from specific re-plans. The traces begin at the enlarged markers, the first re-plan, and proceed in chronological order. It follows that all traces begin on the line Fi = Fp, since for the first re-plan, the previous plan is also the initial plan.
Figure 6-9: Arrival Time Stability: Changes in arrival time in re-plans correlate well with deviations from the previous plan during execution.
Observe that for all but one trace, the endpoint falls generally left and above the starting point. One can infer that for these cases, re-plans are initially unstable but grow gradually more stable as plan execution progresses. This seems to make intuitive sense. With the greater freedom that comes with a large distance between start and goal, ISE finds a
ROBOTIC ASTROBIOLOGYnumber of plans of similar cost but with differing routes. Subtle changes in initial conditions may cause substantial route variations. However, as the distance to the goal shrinks, the freedom is reduced, leading to greater stability.
The exception is the plan sequence from April 25 (Figure 6-8d), whose first re-plan shares fewer than 10% of the initial route's waypoints. Successive re-plans deviate even more from the initial plan at first, but then return to match about 40% of the remaining plan. Figure 6-3a may help clarify what is happening in this case. The plans seem to alternate between two general routes over the last 1/2 of the traverse. Plan 2 (the initial plan shown) takes the right fork, Plan 3 (the first re-plan) the left. In the first half of the route, Plan 2 and Plan 3 are almost entirely distinct, but run very close to each other. In later planning instances, the plans settle on a variation of the right fork, increasing the fraction of the plan that is identical to Plan 2.
A planner exhibits arrival time stability when re-plans, in response to time deviations from an original plan, result in similar deviations in goal arrival time. Experiments indicate that TEMPEST planning are stable with respect to time.
Figure 6-9 plots arrival time slip (vertical axis) against plan schedule slip (horizontal axis) for re-plans generated on April 20 through April 26. Each marker corresponds to a different re-plan instance. The dashed line falls where schedule slip exactly matches goal arrival delays. Re-plans falling above the dashed line are less direct then their predecessors, while re-plans below the line are more direct. Aside from a few outliers, the data seems to suggest a strong correlation between operational delays and schedule slips.
6.6 Field Experiment 2004 6.6.1 Objectives In anticipation of full science operations for the 2005 field experiment, the principal autonomy goal for 2004 was to integrate science activities into operations. For TEMPEST this meant representing human-designated science goal actions within Mission Specifications, reasoning about the time and resource consumption of these activities in the scope of the global traverse plan, and enforcing temporal and energy constraints imposed on the completion of goals.
6.6.2 Zoe Rover A new rover, Zoe, was developed to better integrate science instruments and to incorporate the lessons learned with Hyperion (see Figure 6-10). Its mass was 180 kg and its dimensions were 2.7 m long by 1.7 m wide, on par with the mass and size of Hyperion. It was far more capable of ascending steep slopes and crossing over rough terrain, and is designed mechanically to drive at a higher average speed. Enhanced computing onboard Hyperion enabled the Local Navigator to reliably avoid obstacles up to 1.0 m/s (3.6 km/hr), though system-level tests documented in this thesis were run at 0.5 m/s (1.8 km/hr).
Zoe’s power configuration was more capable of collecting and storing solar energy than Hyperion. Zoe utilized a smaller solar array than Hyperion’s (2.4 m2) for supplying current to loads on the system and to charge batteries. The array’s solar cells were triple junction cells, which provided a nominal efficiency of 24%, a substantial improvement over Hyperion’s. The net effect between the size reduction and the efficiency enhancement was anticipated to be a 64% increase in solar power. Zoe’s principal batteries for system-level tests were lithium-ion cells, designed for a maximum capacity of roughly 1340 W-hr of charge. The impact of these power upgrades was that in driving, Zoe was even more power-rich than Hyperion. However, with the addition of power-hungry science instruments enabled during science activities, it was not obvious whether the previous system would have been sufficient to sustain the rover in daily operations.
Increased navigational autonomy was a secondary goal of LITA 2004. To achieve a greater level of space relevance, rover state estimation no longer relied on GPS, but instead on a combination of wheel odometry, rate gyros, a sun tracker to enable absolute measurements of vehicle attitude, and a novel non-linear smoothing algorithm to update the position estimate history as new sun measurements are taken. This approach holds promise for the future, but was not functioning nominally during 2004 system-level tests. The impact to TEMPEST planning was that position state errors could no longer be maintained to below the resolution of the position state representation.
ROBOTIC ASTROBIOLOGYIn contrast to Hyperion, Zoe had a substantial suite of science instruments, and could collect several types of measurements autonomously by the end of the 2004 field experiment. A stereo panoramic imager (SPI) comprised three cameras mounted on a pan/tilt head on the Zoe mast, enabling mono- and stereo panoramic image data sets. A foreoptic for Zoe’s visual/near-infrared spectrometer (VNIR) was also mounted on the pan/tilt head, providing spectral data for regions of the panorama. Beneath the rover, a Fluorescence Imager (FI) could automatically deploy to activate and image the fluorescence of geologic or biotic materials. Optical cameras whose field of views covered the workspace of the FI provided context for fluorescence measurements.
6.6.3 Software Architecture Hyperion’s autonomy architecture was re-designed to better enable integrated science and navigation planning and execution and fault recovery. Zoe incorporated several new modules - a Rover Executive (RE), a replacement for Hyperion’s Mission Executive, to coordinate the planning and execution of mission plans; a Goal Manager (GM) for pre-processing of goal specifications sent to TEMPEST and goal elaboration following TEMPEST planning; and an Instrument Manager (IM), an executive process to coordinate the execution of measurement activities.
The Rover Executive (RE) was developed under the IDEA architecture , and coordinated mission planning and execution. Unlike the previous Mission Executive, the RE aggregated logical models of each anticipated event and possible state transitions. It called upon the EUROPA planner  to perform both deliberative and reactive
scale temporal constraint checking on plans generated by the GM and TEMPEST, maintained a set of timelines to enforce temporal constraints on all planned activities, and enacted simple plan execution monitoring and recovery actions for common vehicle faults.
The RE coordinated the execution of plans encoded on the timelines. Drive actions were passed on to the Local Navigator as 18 m wide by 6 m deep goal regions similar to previous field experiments. Charge actions were executed directly by the RE by waiting the required duration. Science activities were passed to the IM, which managed the deployment of instruments and the execution of measurement sequences. For all activities, the RE monitored the execution timing of activities relative to the plan timelines. If either the execution completed prior to the earliest allowable time, or did not complete by the latest allowable time, the RE requested a re-plan from the GM and TEMPEST. The RE was not given the flexibility to anticipate the failure of an action to take early recovery steps.
The Goal Manager (GM) was designed as a pre- and post-processor for TEMPEST planning. As in 2003, goals for Zoe were specified exclusively by human operators. Under sequential goal planning described in Section 4.3.4, TEMPEST expected a fully-ordered sequence of goals. First, the GM elaborated goal activities, based upon goal activity parameters (e.g. size of panorama, number of pan-tilt steps), to predict appropriate time and energy allocations, and to convert coarse representations into more detailed specifications needed for execution.
Furthermore, the science and engineering team could not be expected to predict how many of the goals in the Mission Specification could feasibly be achieved within daylight hours, or which goal to remove if achieving the entire set was infeasible. Given a goal sequence, rewards assigned to each goal, and the latest time by which the mission must be completed, the GM used TEMPEST domain models and approximate energy constraints to estimate and select the highest-reward subset of goals achievable within the allotted time. The resulting subsequence of goals was sent to TEMPEST for planning. The GM returned elaborated plans to the RE for execution. Originally, the GM was intended to reduce the goal set in the event TEMPEST could not find a feasible plan. Time constraints in the software development schedule prevented this feature from being incorporated into the system.
6.6.4 Planning Approach Table 6-2 summarizes the planning parameters used in the LITA 2004 experiments. To accommodate science activities, TEMPEST reinstated sequential goal planning, but unlike in the Arctic or in LITA 2003, with goal actions.
TEMPEST was never intended to solve the general planning and scheduling problem typically solved by classical AI approaches (see Section 1.5). Because TEMPEST’s principal role is to solve for traverse plans that satisfy temporal and resource constraints, the critical parameters for a goal action are its position, duration and resource (energy) consumption. For every science activity requested in a Mission Specification, the GM provided TEMPEST with duration and energy consumption upper bounds. For each of these activities, TEMPEST created a new science action using a
ROBOTIC ASTROBIOLOGYGeneric Science action template, and added it to the Action Set. Other details, for example the hardware units designated for the activity or warmup and calibration procedures, were irrelevant to TEMPEST and left out of plan requests.
In preparing for possible 24-hour autonomy experiments in 2005 and beyond, one objective was to enable TEMPEST planning that would guarantee sufficient battery charge at the end of each to survive at low power overnight. Without knowing the completion time of a day’s mission, it is difficult to assign goal battery energies that would enable overnight survival. To accommodate this, TEMPEST was augmented to enable time-bounded sequential goal planning, as described in Section 4.3.6. All Mission Specifications included an additional goal (with a null action) whose position was co-located with the final requested goal - an “End of Day” goal. To this goal, human operators assigned a legal time bounds corresponding to sundown and the battery energy required for night survival from that time. The time-bounded goal planning mechanism restricted plans to terminate at the position, within the time bounds and at the energy for the End of Day goal. The completion times and battery energies for requested goals remained free. With no specialized delay action in the Action Set, Charge actions would allow TEMPEST to insert delays into plans. To enable longer delays without added penalty, the LITA 2004 Action Set included additional, longer Charge actions.
The End of Day planning strategy, as stated above, did not result in the desired behavior. Ideally, plans would timeefficiently achieve all goals through Drive and Generic Science actions, and then loiter at the final goal position using Charge actions until satisfying the End of Day time bounds and energy. However, because TEMPEST planned in backward-chaining order, Drive actions were favored over most of the search, and Charge actions were only included in optimal plans to meet the conditions of the start time interval of the Mission Specification. This resulted in plans that began with Charge actions to accomplish the delay, followed by fast-as-possible Drive and Generic Science action sequences to achieve all the goals. Operationally, this embodied a risky strategy that required flawless execution of the traverse and science activities to meet the overall objective.
To remedy this situation, the objective function used in LITA 2003 was augmented with a third term that imposed “reward pressure.” Human operators selected a reward pressure constant, expressed in units of power per unit reward (Watt-hours/reward), to be used over all planning segments. At each step in the search, the additional reward pressure cost was defined as the pressure power multiplied by the duration of the action multiplied by the reward remaining in future goals. The effect was to lightly penalize adding loiter actions towards the end of plans (the beginning of the search), but heavily penalize adding loiter actions towards the beginning of plans.