Perseverance’s AI Waypoints Put Rover Driving Through a New Safety Pipeline
“This demonstration shows how far our capabilities have advanced and broadens how we will explore other worlds,” said NASA Administrator Jared Isaacman.
This was noteworthy about the two December crossings of Perseverance: the route geometry was generated by a generative AI system, which rover planners typically develop manually, and the verification machinery that the mission already had handled the rest. This led to two drives in Jezero Crater which made AI a new upstream operations contributor but retained the safety gates on the downstream.
During sols 1,707 and 1,709 of the mission the task of route planning switched to a vision-language model capable of interpreting both the imagery and the terrain products. This system combined high-resolution orbital images of HiRISE camera with slope data of digital elevation models and suggested a continuous route with waypoints. Practically it needed to be able to view the same operational hazards that rover drivers are worried about: bedrock, outcrops, boulder fields, and sand ripples, and also generate a waypoint chain that should be compatible with how Perseverance would like to be commanded. It is that compatibility requirement which frequently pits autonomy discourse as real engineering since a plausible path is not an uplinkable sequence.
The overruling stroke was yet subject to a wilfully cautious refring. The team tested the command products at the virtual JPL digital twin, a computer simulation, before any AI-generated drive could be cleared to fly to Mars, to reveal any minor conflicts in the flight-software. Over 500,000 telemetry variables were checked by the engineers, and the AI output was handled like any other artifact of high consequences: simulate, verify, and uplink. It was not until this step was cleared that Perseverance was given the drive plan via the same mission pipelines that human-built routes are issued.
By then, with such limitations imposed, the rover made an AI-planned 689-foot (210 meters) and 807-foot (246 meters) drive two days later. The milestone can be less in the raw meterage than the proven correspondence between the today’s generative planning and the safety training of the past: the source of the planning varied, yet the philosophical principle of validation remained.
According to Vandi Verma, a JPL space roboticist who is one of the members of the Perseverance engineering team, the foundational building blocks of generative AI are already proving to be promising when it comes to simplifying the pillars of autonomous navigation to drive off-planet: perception (seeing the rocks and ripples), localization (knowing where we are), and planning and control (deciding and executing the safest path). Generative AI and other intelligent systems will bring us to a point where our surface rovers can drive by kilometers with operator workloads reduced, and identify interesting surface features to our science team by scrubbing vast volumes of rover images.
The focus on workload is not symbolic. The nature of Mars operations is characterized by delay, limited contact windows, and a regular cycle of planning that should transform the unperfect knowledge into safe motion. The design of the routes takes the attention of experts since all decisions are combined with energy, slip risk, visibility and the possibility of the rover to get out of unexpected terrain at the next destination. An AI planner capable of writing orbital-context candidates of waypoints can reduce the human time spent on first-pass geometry, allowing teams to concentrate on the more valuable work: cross-checking with the rover-level imagery, reconciling science priorities, determining when a route is worth the risk.
Facilities are also important. The Rover Operations Center (ROC) at JPL is established to be the center of autonomy integration throughout the surface missions, and it transferring decades of rover practice into repeatable pattern of operations. The December drives in that framing are a test of systems: not whether an AI can draw a safe line, but whether the larger operations stack, tooling, simulation, review culture, and command interfaces, can receive AI outputs without compromising the guardrails.
“Imagine intelligent systems not only on the ground at Earth, but also in edge applications in our rovers, helicopters, drones, and other surface elements trained with the collective wisdom of our NASA engineers, scientists, and astronauts, etc”. said Matt Wallace, manager of the Exploration Systems Office at JPL.
