Planning and scheduling

Deciding what to do next and in what order, from local path selection to multi-sol activity plans and onboard science target selection.

Planning software has to turn goals into commands that a vehicle can execute safely when no one can watch it. Two features of the environment make that hard. Light-time delay and contact windows mean a rover is usually commanded once per sol and must act on its own between uplinks [1]. The processors are radiation-hardened parts a generation or more behind terrestrial hardware, so every planner competes with the rest of flight software for a few MIPS and a few megabytes. On Mars Exploration Rover (MER) the flight computer was a 20 MHz RAD6000 with 128 MB of RAM [7]. On the Earth Observing One spacecraft the onboard planner was allocated 4 of 8 MIPS and 40 MB of 256 MB [12].

Four problems, four families of method

The corpus separates planning into problems that differ in horizon, in who runs them and in how much has flown.

Local and global path planning. The oldest family scores candidate motions against a map built from range data. The Stanford Cart built a sparse three-dimensional model of its obstacles and drove around them without a human in the loop [2]. Potential-field methods moved collision avoidance inside the servo loop [3]. Incremental replanners such as D* repair a path as the map changes [6]. Sampling-based planners grow trees through the state space [4], and later analysis showed that the classic variants converge to a non-optimal path unless modified [5]. Flown systems took a conservative subset. MER ran GESTALT, an arc-scoring local planner that evaluated 23 forward arcs, 23 backward arcs and 2 point turns on a 10 by 10 m grid of 20 cm cells, at about 70 s per cycle; with stereo and visual odometry the median net driving speed fell to 0.6 cm/s against a 5 cm/s peak [7]. The Field D* global planner was later integrated with GESTALT, using under 1 MB for a 50 by 50 m cost map and adding 3 percent to drive-step time in testbed runs [8]. Details are on Autonomous navigation.

Ground and onboard activity scheduling. Activity planning places science, engineering and communication activities on a timeline subject to resource and flight-rule constraints. Ground tools such as APGEN, which accumulated fifteen years of operational use [23], are being succeeded by the open-source Aerie system [24]. Onboard, the first flight demonstrations were the CASPER planner on Earth Observing One, which had to be cut from a 12 MB inherited image to 5 MB [12], and the Autonomous Sciencecraft Experiment built around it [13]. The Remote Agent prototype had integrated planning, a reactive executive and model-based mode identification, but was a ground prototype [31]. CLEaR, which couples CASPER to a task-description-language executive, was exercised on rover testbeds [14]. Mars 2020 flew the first rover onboard scheduler, a greedy planner designed for schedule generation on the order of a minute and 15 to 20 invocations per sol [15]. It was rolled out incrementally as Simple Planner during Perseverance operations [17], after a verification campaign on the flight software [39] and with ground tooling to build the constraint-based plans it consumes [16]. MEXEC, an integrated onboard planner and executive, reported an in-flight demonstration [18], and its extension to the Endurance lunar rover concept is a design study [35]. CADRE, a lunar multi-rover technology demonstration, describes a leader-elected planning, scheduling and execution architecture [19]. Details are on Activity planning and sequencing.

Onboard science target selection. Here the planner chooses what to observe. AEGIS selects ChemCam targets from Navcam imagery on Curiosity: in flight over sols 1343 to 2830 it ran 237 times, with typical runs under 500 s on the RAD750, and placed targets on outcrop or rock over 90 percent of the time against under 40 percent for blind targeting [20]. A SuperCam version has since been deployed on Perseverance [21]. EO-1 flew the earlier spacecraft-level analogue, selecting observations from onboard image analysis [13]. Details are on Onboard science autonomy.

Manipulator motion planning. Arm planning starts from resolved-rate control of joint velocities through the Jacobian [28] and configuration control of redundant arms [29]. Optimizers such as CHOMP refine a naive path by covariant gradient descent [27]. Flown practice is more restrained: Curiosity arm operations were planned on the ground and executed as sequences, with placement accuracy measured over the first 200 sols [25]. Sampling-based arm planning for Perseverance sequencing has been proposed and tested in simulation with path simplification that reduces viapoints [26]. An end-to-end autonomy architecture for dexterous manipulation was demonstrated under DARPA ARM-S on a ground testbed [30]. Details are on Manipulation planning.

What changed and what it enabled

Autonomous driving moved from a minor fraction of distance to the default mode. By mid-June 2004 Spirit had driven 3405 m of which 1253 m were autonomous, and Opportunity 1264 m of which 224 m [7]. Curiosity logged 21,318.5 m over 738 attempted drives in its first seven years [9]. Perseverance’s Rapid Traverse Campaign drove 5063.4 m in 24 driving sols across a 31 sol window, with 94.8 percent of the distance path-planned by autonomy, against under 7 percent for MER and MSL AutoNav across their missions [10]. The enabling changes were faster computing for perception, with Perseverance processing stereo with no routine pause per step [10], and onboard global localization on a repurposed Ingenuity co-processor, which ran in 32 s on Mars and removes the position uncertainty that otherwise grows at an assumed 5 percent of distance [11].

On the scheduling side, the shift is from a plan fixed on the ground to a plan the vehicle adjusts. The EO-1 experience showed the cost: the CASPER telemetry packet was limited to 248 bytes, so engineers logged the choice made at each decision point rather than the options [12]. The Mars 2020 scheduler gives the ground a plan with flexible start times and lets the rover repair it as activities run long or short [15]. A systems engineering retrospective on that planner records what was not completed before launch [22].

Where each family stands

Capability	Status	Evidence
Local hazard-avoiding arc planning	Flown on MER, MSL, Mars 2020	[7][9][10]
Global path planning (Field D*)	Integrated into MER AutoNav flight software, tested on ground testbed	[8]
Onboard global localization	Flown on Perseverance	[11]
Onboard activity scheduling	Flown on EO-1 and Perseverance	[12][17]
Integrated planner and executive (MEXEC)	In-flight demonstration reported	[18]
Onboard science targeting	Flown on MSL and Mars 2020	[20][21]
Remote Agent planning	Ground prototype	[31]
Risk-aware terrain classification and path planning	Simulation and ground data	[32]
Learned terrain classification, labeled datasets	Dataset published, not flown	[33]
Autonomous navigation over Europa analogue terrain	Field analogue	[34]
Sampling-based and optimizing arm planners	Simulation and ground test	[26][27][30]

Path planning with learned or probabilistic components is the least flown area. A risk-aware tool motivated by wheel damage on Curiosity combined terrain classification and path planning and was evaluated offline [32]. AI4Mars supplies roughly 326,000 crowdsourced segmentation labels for terrain-aware driving but is a dataset, not a flight capability [33]. A full stack of LiDAR SLAM, terrain mapping and sampling-based planning with wheel-on-limb settling checks was run on a Europa analogue terrain [34]. A JPL and Carnegie Mellon position paper lists resource-constrained computing and hardware limits as the standing path-planning challenges for planetary robots [36].

What is not established

Planning and scheduling durations on flight processors are reported for few systems. The EO-1 report gives allocations and code sizes, not measured CPU utilization or replan latency, and the hierarchical planning approach it proposes for heap exhaustion was not in the flying version [12]. The Mars 2020 scheduler’s runtime figures are design expectations, not measurements [15]. The AEGIS comparison with blind targeting is a notional baseline reconstructed after the fact, not a controlled trial, and the paper’s abstract and body disagree on the headline fraction [20]. The MER results cover only the first months of a multi-year mission and carry no error bars or obstacle detection rates [7]. No held source quantifies the science return attributable to onboard scheduling on Perseverance against ground-only planning. Planners for dual-arm space manipulators and for sampling-based arm motion have been shown in simulation, and no held source reports them in flight [26][27].