How AutoNav Let Perseverance Out-Drive Every Vehicle Off Earth

On 11 October, a 91.31-metre drive carried Perseverance past a record that had stood since Opportunity set it. The rover's odometer reached 28.10 miles (45.22 km), above Opportunity's 28.06 miles (45.16 km). The margin is about 60 m, by my arithmetic from NASA's two figures. The timescale is the larger difference. Perseverance covered in five and a half years a distance that took Opportunity almost 15.

The wheels and suspension did not change much between generations. The difference is how much of the driving the rover now decides for itself.

01What the record does and doesn't show

The mission credits its self-driving software, AutoNav. According to JPL, the rover images the terrain while rolling, builds a 3D hazard map, and plots a course between waypoints chosen by planners on Earth. The numbers that matter:

• More than 90% of Perseverance's distance was driven with AutoNav evaluating the terrain. The figure for the first Martian AutoNav on Spirit and Opportunity was about 5%.
• Sojourner judged the path ahead with 20 measurement points. Perseverance processes 600,000 with every look.
• At or near top speed, the rover can cover as much as 1,351 feet (412 m) in a sol.

"AutoNav-evaluated" does not mean unsupervised. Humans still pick the destination. They plan the basic route from orbital images, such as those from Mars Reconnaissance Orbiter. The 412 m figure is also a ceiling, not a typical day. Dividing 45.22 km by 2,006 sols gives about 22 m per sol. That is my calculation, and the average is low because science stops, sampling and restricted drive windows dominate the schedule.

02Why earlier rovers crawled

The constraint is computing, not wheel speed. According to the Science Robotics paper by Vandi Verma, Mark Maimone and colleagues, Perseverance keeps Curiosity's motor controllers, a similar top drive speed of about 4.4 cm/s, the same six-wheel rocker-bogie suspension, and a 133 MHz RAD750-based main computer. That processor is radiation-hardened and slow, and almost 50% of its capacity goes to baseline rover functions other than navigation.

On the single-CPU design, prior rovers had to stop, collect images and wait for path selection to finish before turning the wheels. JPL's mobility page puts Curiosity's autonomous average at just 15 to 18 metres per hour. A rover that can move roughly 150 m/h loses most of its time standing still. That stop-and-think penalty is why earlier missions used AutoNav only to extend drives that humans had already planned.

03Thinking while driving: the architecture

Perseverance attacks the bottleneck in two ways: dedicated hardware and a cheaper planning algorithm.

Hardware. A second RAD750, the Vision Compute Element (VCE), was added to speed up stereo processing. The VCE uses a radiation-hardened Xilinx FPGA for the parallel operations: image rectification, prefiltering, stereo disparity and visual-odometry feature scoring. With that work offloaded, the rover can drive at nearly full speed with all AutoNav functions enabled. The Science Robotics paper reports average autonomous speeds of 11.55 m/h for Curiosity and 97.97 m/h for Perseverance. The newer IEEE Transactions on Field Robotics paper on ENav gives roughly 17.3 m/h for Curiosity against about 92 m/h for Perseverance. The two papers use different accounting, so treat the values as an order-of-magnitude gain and not a precise ratio.

Planning. The planner is called Enhanced Autonomous Navigation (ENav). It builds a height map from stereo imagery and ranks candidate arcs. Then, according to a JPL-affiliated machine-learning paper, it uses the Approximate Clearance Evaluation (ACE) algorithm to check whether the highest-ranked paths are safe. ACE is cheap by design. It estimates the lowest and highest heights each wheel might reach on the terrain, then evaluates the worst-case vehicle configuration. That replaces iterative suspension simulation, which is too slow for a RAD750. A benchmark on RAD750-class hardware measured about 12 ms per ACE check. The same paper's baseline ENav spends roughly 4.5 s per planning cycle on 377 such checks. Running the expensive check on only a handful of top candidates keeps that cost bounded.

The ENav paper explains why this matters for wheel-aware assessment. Earlier planners used a coarse disk model of the rover. ENav assesses hazards from each wheel's location using ACE. That allows it to drive through rockier terrain than a conservative footprint would permit.

04Latency, line of sight and drift

Communication delay shapes this design in a specific way. The binding limit is the once-per-sol uplink cycle, not only the light-time, so a rover that stops whenever its human-planned path runs out is idle most of the day. The ENav paper notes that AutoNav is not restricted to terrain visible from the starting position. That lets it drive far beyond the roughly 15–50 m line-of-sight in each directed drive. Planners supply goals and keep-out zones. The rover handles the metre-scale geometry that orbital imagery cannot resolve.

The cost of autonomy is position error. Perseverance tracks itself with visual odometry. JPL notes that small errors accumulate over a drive. On long drives the rover's estimate of its position can be off by more than 100 feet (up to 35 m). JPL's earlier fix was an onboard algorithm that matches camera imagery to orbital maps. Maimone now describes newer software that lets the rover check its position against orbital maps mid-drive. He also says it teaches Perseverance to steer less and roll more. JPL has not published performance numbers for that update, so its effect is unquantified.

05Limits and open problems

The first limit is sunlight. Bright light can disrupt the navigation cameras and the software's reading of terrain. When driving toward the Sun, planners currently stop driving by mid-afternoon. Maimone calls sunlight itself the next frontier, saying that teaching the rover to see past glare would give every future rover a longer working day.

The second is that ENav reasons about geometry, not material. The ENav paper flags terrain classification as future work. That would let rovers avoid non-geometric hazards such as slippery sand.

The third is how much ground-side planning can be automated. According to a secondary report citing JPL's January 2026 announcement, in December 2025 a vision-language model analysed orbital images and elevation data and generated waypoints for the rover. Engineers checked the commands in a digital replica against more than 500,000 telemetry variables before transmission. I could not verify the details against JPL's own release. If it holds, the pattern is the same as in AutoNav: a bounded model proposes and a conservative verifier disposes.

06Lessons for autonomous robotics

1. Architecture beats raw compute. The wheels and CPU class stayed close to Curiosity's. Adding a vision co-processor and removing the stop-and-wait cycle changed throughput by an order of magnitude.
2. Bound the expensive check. ACE runs only on a few ranked candidates, so worst-case safety cost stays predictable on slow hardware.
3. Autonomy needs a localisation budget. Faster driving raises drift risk, which is why map-matching against orbital data matters.
4. Humans stay in the loop. They set goals and keep-out zones. Autonomy shortens the loop but does not remove it.

The mission has more to run on. NASA has approved a first extended mission, running from October through roughly the end of 2028. Project manager Steven Lee expects the rover to keep adding to this record for many years to come. The next gains depend on software such as glare handling and mid-drive localisation, not on new hardware.