Perseverance

NASA/JPL-Caltech/MSSS. Public domain (NASA / US government work).

Overview

Perseverance caches rock cores for later return to Earth. Chassis, mobility system and sky crane landing architecture are inherited from Curiosity; payload, avionics and entry sensing were reworked.

Specifications

Parameter	Value
Mass	1025 kg
Length	approximately 3 m
Wheel diameter	52.5 cm
Distance advanced per wheel revolution	1.65 m
Maximum wheel speed	4.2 cm/s, approximately 150 m/h
Arm	5 joints, 2.1 m
Power source	MMRTG, approximately 110 W at launch, approximately 45 kg

Values from [6].

Mission profile

Parameter	Value	Source
Landing	18 February 2021, Jezero crater	[13], [19]
Landing point	about 2 km from the western delta front	[13]
Jezero crater diameter	45 km, on the western edge of Isidis Planitia	[19]
Traverse through the crater floor campaign	more than 5 km, seven abrasion patches	[13]
Distance driven as of October 2024	more than 30 km	[19]
Samples collected	24 rock and regolith samples and one atmospheric sample
Sample depot	ten tubes deposited at Three Forks

The mission page carries totals the retrieved literature does not: more than 30 km driven as of October 2024 and 24 rock and regolith samples plus one atmospheric sample [19], against the more than 5 km and seven abrasion patches recorded for the crater floor campaign in the mission overview paper [13]. It also records the July 2024 Cheyava Falls sample from Sapphire Canyon, validated in September 2025 as containing a potential biosignature.

Commissioning ran on a fixed schedule: cruise to surface flight software transition on sols 5 to 8, Remote Sensing Mast deployment on sol 2, robotic arm deployment on sol 12, first drive on sol 14 [13]. Ingenuity was deployed over sols 36 to 43 and flew its technology demonstration campaign over sols 44 to 73. By the end of the crater floor campaign the rover had traversed more than 5 km and created seven abrasion patches [19].

Mobility

The suspension is the Curiosity rocker-bogie with six driven and four steered wheels. The wheels were redesigned against Curiosity’s damage record: 52.5 cm diameter aluminum with curved titanium spokes and cleats, one revolution advancing the rover 1.65 m [6]. The skin is thicker and the grousers more numerous and less aggressive than Curiosity’s, spreading the load an embedded rock applies over more material.

Drive rate	Value	Source
Maximum wheel speed	4.2 cm/s, approx 150 m/h	[6]
Top speed, flat hard ground	approx 152 m/h	[3]
Curiosity autonomous rate, stop-to-compute	15 to 18 m/h
Perseverance autonomous rate, computing while driving	approx 110 m/h

The gap between the two autonomous rates is a compute architecture result, not an actuator one: the actuators are the same class as Curiosity’s and the maximum commanded wheel speed differs by only a few percent [3], [6].

Power and energy

An MMRTG of the same class as Curiosity’s supplies approximately 110 W at launch, declining a few percent per year, buffered by two lithium-ion rechargeable batteries; the generator masses approximately 45 kg [6].

Awake and idle load exceeds MMRTG output, so battery state of charge rises only while the rover sleeps [4]. Every plan interleaves sleep with activity, and the ground scheduler enforces a minimum state of charge below which the predicted energy value may not fall; an activity that would violate it is not scheduled. Actuator preheats are scheduled as explicit setup activities preceding the activity they enable, and are themselves energy consumers. Sampling activities carry a prerequisite abrasion whose added time and power must be budgeted before the proximity science it enables [13].

Thermal

Perseverance uses the Curiosity thermal architecture: a pumped fluid loop through the warm electronics box that distributes MMRTG waste heat and rejects excess, with electrical heaters on components outside the box. Perseverance-specific plate temperature requirements distinct from Curiosity’s are not published, so they are not restated here. One flight thermal figure is on record from the sampling chain: tubes are sealed at approximately 40 C [2]. Sealing temperature is one of the three parameters that govern the leak rate achieved, alongside tube wall thickness and diametric tooth interference.

Compute and avionics

Element	Specification	Source
Main computer	BAE RAD750, PowerPC 750, up to 200 MHz	[6]
Main memory	256 MB DRAM, 2 GB flash, 256 kB EEPROM
Redundancy	two, one prime one backup
Vision Compute Element	separate RAD750 plus Computer Vision Accelerator Card	[3]
Accelerator	FPGA implementing the image processing pipeline in hardware
Stereo correlation and visual odometry, 1280 x 960	seconds on the VCE, approximately one minute on the main computer

The Vision Compute Element processes navigation imagery independently of the main computer, so the main computer continues to run the drive while imagery from the previous step is still being reduced [3]. The same VCE ran the Lander Vision System during entry, descent and landing and was converted to full-time surface navigation afterwards. Curiosity’s single 133 MHz RAD750 takes 47 s per visual odometry stereo pair including data product writing, which bounds its drive rate to the stop-and-compute figure above [3].

Autonomy

Terrain-relative navigation was added to entry, descent and landing. During parachute descent the Lander Vision System images the ground, matches those images against an onboard map built from orbital reconnaissance, estimates position relative to a hazard map, and commands the descent stage to divert to a safe point [3]. The image matching completes within the available seconds because the FPGA accelerator performs it. Landing site selection therefore moves from a pre-launch judgment about ellipse-scale terrain to an in-flight decision at hazard scale, which is what made the broken terrain near the Jezero delta front an admissible target [13].

Surface autonomy is the ENav algorithm, which performs orientation-sensitive hazard assessment rather than treating the rover as a point or a disc [3]. Evaluating clearance against vehicle orientation over each candidate patch is a substantially larger computation than an orientation-free check, and is affordable only on the accelerator. Orientation matters because the rocker-bogie determines wheel contact angles, and it is the contact angle distribution across the six 52.5 cm wheels that decides whether a given rock is climbed or driven into [6].

Communications

Link	Rate
X-band DTE, 34 m DSN antenna	160 bps or better
X-band DFE, 34 m DSN antenna	500 bps or better
X-band DTE, favorable geometry	800 bps or better
X-band DFE, favorable geometry	3000 bps or better
UHF rover to orbiter	up to 2 Mbps

Link rates from [6].

Bulk return is by UHF relay through the Mars Relay Network of MRO, Odyssey, MAVEN and the ESA Trace Gas Orbiter [6]. No open equivalent of the DESCANSO telecommunications article exists for Mars 2020, so channel plans, coding and Proximity-1 configuration are not stated here; the Curiosity entry describes the Electra relay conventions the network imposes on both rovers.

Payload and instruments

The 2.1 m five-joint arm carries a turret with the rotary-percussive corer, PIXL, SHERLOC with its WATSON imager, and the gaseous Dust Removal Tool [6], [13]. Turret instruments are placed within a few cm of the natural or abraded surface. Mast instruments are Mastcam-Z and SuperCam. MEDA provides meteorology and RIMFAX ground-penetrating radar.

Instrument	Key parameters	Source
SuperCam mast unit	6.11 kg, 383 x 201 x 163 mm, 27 W peak	[14]
SuperCam body unit	4.44 kg, 221 x 157 x 205 mm, 43 W peak with thermo-electric coolers, 12 W idle
SuperCam total	10.80 kg, 70 W
SuperCam LIBS laser	1064 nm, up to 14 mJ, more than 10 MW/mm2; requirement above 12 mJ across -30 to +10 C	[14][15]
SuperCam Raman laser energy	above 9 mJ over -30 to +10 C	[15]
SuperCam microphone	100 Hz to 10 kHz, sampled at 25 kHz or 100 kHz
SuperCam mast unit operating range	-30 to +10 C
SuperCam Raman	same laser frequency doubled to 532 nm, time-gated detection	[14]
SuperCam footprint	1.5 mm close in, approx 5 mm diameter at 7 m standoff
SuperCam transmission spectrometer	535 to 853 nm at 12 cm⁻¹ FWHM, Raman shift 105 to 7070 cm⁻¹
SHERLOC laser	248.6 nm pulsed neon-copper, at least 9 µJ per pulse, 80 Hz, 40 µs pulses, peak power under approximately 200 mW	[16]
SHERLOC mass	9.73 kg total: 6.83 kg turret assembly, 2.90 kg body assembly, 0.44 kg calibration target
SHERLOC turret survival heating	128.8 W peak; decontamination 80.4 W
SHERLOC spot / working distance	approx 100 µm / approx 48 mm
SHERLOC map	up to 7 x 7 mm at better than 100 µm per pixel, under 10 min
Mastcam-Z zoom	4:1, 26 to 110 mm focal length	[17]
Mastcam-Z field of view	25.6 x 19.2 deg to 6.2 x 4.6 deg
Mastcam-Z pixel scale	283 to 67.4 µrad/pixel
Mastcam-Z detector / baseline	1648 x 1214 CCD / 24.4 cm stereo baseline, 2.3 deg toe-in
Mastcam-Z closest focus	0.5 m at 26 to 50 mm, 1.0 m at 50 to 110 mm
Mastcam-Z mass / heater draw	1.38 kg per camera head, 1.47 kg camera plate / up to 72 W
MEDA air temperature	150 to 300 K, ±1 K, 0.1 K resolution	[18]
MEDA pressure	1 to 1400 Pa, ±20 Pa over 400 to 1200 Pa
MEDA wind	0.5 to 30 m/s expected at Jezero, 15 deg direction resolution

SuperCam block diagram. The mast unit carries the laser, telescope, focus mechanism, infrared spectrometer, Remote Micro-Imager and microphone; one optical fiber plus power and data lines runs down the mast to the body unit, where a demultiplexer splits returned light between the ultraviolet and violet reflection spectrometers and the intensified transmission spectrometer used for Raman. The split keeps the cooled detectors inside the warm electronics box and leaves only optics and laser at mast temperature. Source: [14]. CC BY 4.0.

SuperCam occupies almost exactly the ChemCam volume and mass, and its optical demultiplexer and ultraviolet and violet reflection spectrometers are nearly identical to ChemCam’s; the third reflection spectrometer was replaced by a time-gated intensified transmission spectrometer to enable Raman [14], [15]. Two mass-driven changes distinguish it: the thermo-electric coolers were designed in from the start rather than added late, giving a lighter cooling system, and the body unit spectrometers are titanium rather than beryllium. The telescope carries a periscope mirror that directs Raman laser light to the sample. SuperCam’s body unit sits next to the Remote Sensing Mast side of the rover to minimize the length of the optical fiber carrying signal from the mast unit, inside an instrument and electronics bay of 1181 x 1106 mm [14].

SuperCam flight hardware. (a) Mast unit: laser with its heater at left, electronics box, and the telescope carrying the periscope mirror that directs the frequency-doubled Raman beam; the unit stands on insulating feet to limit conduction into the mast. (b) Body unit: fiber bundles enter through the demultiplexer to the reflection spectrometers and the transmission spectrometer, with the thermo-electric cooler finned block at lower left. (c) Calibration target assembly. Source: [14]. CC BY 4.0.

SHERLOC scans with an internal mirror rather than moving the arm, which removes sub-millimeter arm positioning from the mapping loop [16]. Laser peak power is held under approximately 200 mW to avoid photochemically altering the organics being measured, which is why the design uses 40 µs pulses at 80 Hz rather than fewer, harder pulses. WATSON is a copy of the MSL MAHLI camera reused on shared SHERLOC infrastructure, so its 1600 x 1200 detector, 2.1 cm to infinity working distance and 365 nm ultraviolet illumination carry over unchanged [16].

Mastcam-Z adds zoom, focus and filter wheel mechanisms to each camera head [17]. The 24.4 cm baseline and 2.3 deg toe-in are fixed, so ranging performance varies with focal length across the 4:1 zoom range, and closest focus moves from 0.5 m at the wide end to 1.0 m above 50 mm. Heater draw to bring the camera heads above their operating limits reaches 72 W, which makes imaging an energy-scheduled activity like any actuated one [4].

MOXIE occupied a slot in the rover body and demonstrated in situ resource utilization. It draws Martian atmosphere through a HEPA dust filter, compresses it with a scroll pump, heats it to 800 C, and passes it over a nickel-based catalyzed cathode in a solid oxide electrolysis stack, where CO2 decomposes and a scandia-stabilized zirconia electrolyte passes oxygen ions to the anode [1]. The reference operating point is 55 g/h of atmospheric intake producing 6 g/h of oxygen at better than 98 percent purity [13]. The 800 C operating temperature is the compromise between electrolysis efficiency and damage to temperature-sensitive materials. The binding constraint is coking: at insufficient mass flow for a given current, carbon deposits on the cathode, raising cell resistance by reducing active area and potentially fracturing it, so each run is planned to stay inside a safe voltage zone bounded by the two Nernst potentials. Measured oxygen purity exceeded the 99.6 percent recommended for propellant and breathing use, with the residual impurity traced to CO2 crossover driven by the cathode-to-anode pressure differential [1]. The MOXIE chassis measures 23.9 by 23.9 by 30.9 cm, and being carried inside a rover imposed the volume, thermal and power limits that constrained the design [1]. A system scaled several hundred times, at 2 to 3 kg/h, would supply the approximately 31 t of oxygen in the 50 t propellant load of a six-person Mars ascent vehicle.

Sampling and caching

The sampling chain is three cooperating mechanisms [2]. The external five-joint arm presses the corer against the rock. Inside the rover body the Adaptive Caching Assembly holds the tube inventory, and a small internal Sample Handling Arm moves tubes between the bit carousel, the inspection stations, the sealing station and storage [13]. The corer houses either an abrasion bit or a coring bit, exchanged from the bit carousel by the handling arm.

Parameter	Value	Source
Abrasion patch	2 to 16 mm deep, 45 mm diameter	[13]
Dust-cleared area after gDRT gas puff	40 mm diameter
Rock core	13 mm diameter, up to 76 mm long
Sample tubes carried	43: 38 for rock or regolith, 5 witness tubes	[2]
Tube material	Ti-6Al-4V
Tube coatings	white alumina externally, titanium nitride internally
Tube internal volume	approx 12.5 cm3
Sealing temperature	approx 40 C
Proximity science standoff	a few cm from the natural or abraded surface	[13]

Witness tubes differ from sample tubes only in their interior contents and are used to characterize contamination [2]. The external alumina coating limits solar heating and the internal titanium nitride limits adsorption of organic contaminants. A tube is loaded into a coring bit before drilling so that core, tailings and headspace gas are captured directly in the tube; the handling arm then advances it through volume assessment with a probe and imaging by CacheCam before sealing [13].

The seal is a mechanically activated plug: a ram drives a ferrule into a gold-coated titanium seal cup, expanding a sealing tooth into the tube wall and holding it there by plastic deformation [2]. Leak-tightness is governed by tube wall thickness, sealing temperature and diametric tooth interference, the measure of contact between tooth and wall. The mission requirement was verified only as better than the 1e-10 scc/s detection limit of conventional helium leak detection [2]. Static-mode mass spectrometry on six flight-like tubes later measured room-temperature helium leak rates from 8.8e-17 to 4.6e-14 scc/s, below the 1e-13 scc/s at which gas ingress or egress becomes negligible over the intended storage and transit interval [2]. Filled tubes have been deposited at a surface depot, so the cache is recoverable independently of the rover’s continued health.

Sampling is executed against a prescribed sol path spanning eleven continuous sols [13]. The rover acquires survey images to select the abrasion and coring target, abrades, clears dust with the gDRT, takes proximity and remote science on the abraded patch to document the rock, and then takes two cores in succession from unabraded portions of the same rock.

Modes of operation

The sol is built as an explicit sequence of awake and asleep intervals, because the battery recharges only while the rover sleeps [4]. Each activity declares its unit resources, such as the arm, so that two activities claiming the same resource cannot overlap; its energy and data volume consumption; its peak energy rate; and non-depletable resources such as available sequence engines and peak power. An activity is generally preceded by preheat activities and followed by a return to sleep. The ground scheduler solves against this model, and the same core scheduling logic was intended for an eventual onboard scheduler operating under tighter CPU limits.

Sampling imposes its own mode structure: an eleven-sol prescribed path in which the observations documenting each pair of cores are made in a fixed order, so the plan for any one sol in the sequence is constrained by what the preceding sols established [13].

Drive modes follow the Curiosity pattern of directed and autonomous navigation with onboard fault protection, with ENav replacing the earlier hazard evaluation [3]. No open source enumerates Perseverance safe mode entry conditions in the detail available for Curiosity, so they are not restated here.

Ground operations

Perseverance is operated from JPL on a five-stage cycle implemented in COCPIT and Playbook [5].

Stage	Function
Parcel development	defines and validates reusable activity definitions
Strategic planning	look-ahead plan over approximately five sols including communication passes
Campaign implementation	plan fragment templates against timing constraints, heating and resource assessment
Tactical uplink	plans the next sol against latest downlink, verifies constraints, generates sequences
Tactical downlink	monitors the executing plan and analyzes returned data

Decisional data is the subset of returned data that the next tactical planning cycle depends on. A plan cannot be built until its decisional data has arrived, so the tools display when each activity’s data contribution becomes available on Earth and aggregate expected data volume by criticality, both at plan level and at a given time within the plan [5]. Relay pass timing constrains the planning cycle as directly as the rover’s own schedule does.

Scheduling is automated. Copilot places activities against the resource model above [4]. When the scheduler fails to place an activity the cause is a conflict among constraints rather than anything visible in the output, so Crosscheck was built to present the generated schedule visually and analyze which constraints prevented placement and how they would have to change. The team is distributed rather than co-located and the tooling was built for concurrent remote collaboration on a shared plan [5].

Drilling and coring approval is a separate gate, and the crater floor campaign record gives its scale: more than 5 km driven and seven abrasion patches produced before the first sampling campaign concluded [13].

Technologies developed

Terrain-relative navigation moved landing site selection from a pre-launch judgment about ellipse-scale terrain to an in-flight decision at hazard scale [3]. Hardware-accelerated onboard vision demonstrated that a dedicated co-processor, rather than a faster general-purpose radiation-hardened processor, is the practical route to real-time perception under flight computing constraints: the same actuators driven with the same maximum wheel speed produce roughly six times the autonomous traverse rate [6]. The sample tube hermetic seal established a verified leak rate floor for planetary sample containment, at 8.8e-17 to 4.6e-14 scc/s helium, three orders of magnitude below what conventional helium leak detection could confirm [2]. MOXIE established that solid oxide electrolysis of the Martian atmosphere works at flight scale across seasonal and diurnal variation in atmospheric density, and that its operating envelope is bounded by carbon deposition rather than by the electrochemistry [1].