The one-sentence version
Signals between Earth and Mars take 4 to 24 minutes one way. It is impossible for a Mars rover to wait for step-by-step remote control from the ground; it must see the path itself, decide how to move, and even select rocks worth studying. NASA's "Perseverance" rover completed most of its driving distance using an autonomous navigation system, the "Ingenuity" helicopter completed dozens of autonomous flights, and the AEGIS system can automatically select scientific targets without human intervention. Space exploration is one of the earliest fields where AI autonomy has been implemented; its experience (limited computing power, zero tolerance for error, and mandatory verifiability) provides valuable references for today's AI Agents.
About Timeliness
This article is organized based on public information available at the time of writing and has not been cross-checked online item by item.
Layer 1: The Intuition
1.1 Remote controlling a car with a twenty-minute delay
Imagine you are remote controlling a car, but every command takes 20 minutes to arrive, and the video feed from the car takes 20 minutes to return. You say "move forward one meter," but by the time you see the result, 40 minutes have passed.
Early Mars rovers worked exactly like this: ground teams planned a small segment of the path every day, and the rover followed it the next day. The efficiency was very low.
1.2 Three levels of autonomy
| Layer | Example |
|---|---|
| Obstacle Avoidance | Seeing a rock and driving around it |
| Path Planning | Given a target point, finding a safe route by itself |
| Scientific Decision Making | Judging for itself which rock is worth photographing and analyzing |
Layer 2: Representative Achievements
| System | Key points |
|---|---|
| AutoNav (Perseverance, since 2021) | Uses stereo cameras to build a 3D map of the surrounding terrain, evaluates the danger level of every patch of ground, and plans routes while moving. Compared to its predecessor "Curiosity," it can "think" while driving instead of stopping to calculate, significantly increasing travel speed; most of Perseverance's driving distance is completed via autonomous navigation. |
| Terrain-Relative Navigation (during landing) | During the descent of Perseverance, it takes real-time images of the ground and compares them with pre-stored maps to determine its position and automatically avoid dangerous landing zones. This allowed it to land in Jezero Crater, which was previously considered too dangerous. |
| AEGIS | First deployed on "Curiosity" and later used on "Perseverance": automatically identifies rocks that meet specific criteria in images and analyzes them with a laser spectrometer. It does not need to wait for ground instructions, significantly improving the efficiency of scientific data acquisition. |
| Ingenuity Helicopter (2021–2024) | The first aircraft to achieve powered flight on another planet. Originally planned for only 5 technical demonstration flights, it ultimately completed 72; every flight was completed autonomously by the onboard system using visual navigation. |
| Deep Space 1 Remote Agent (1999) | An early NASA experiment: allowing the AI system on a spacecraft to autonomously plan tasks and detect/handle faults; a milestone in space autonomy research. |
Layer 3: Lessons and Insights
3.1 Characteristics of Space Autonomous Systems
| Constraint | Response |
|---|---|
| Extremely limited computing power (radiation-hardened chips are much slower than smartphones) | Simple, efficient, and predictable algorithms, rather than Large Language Models |
| Zero tolerance for error (no one can go there to fix it) | Conservative safety margins; stop and wait for ground instructions when uncertain |
| Must be verifiable | Extensive ground simulation and testing; interpretable behavior |
| Limited communication windows | Autonomously handle daily tasks; leave major decisions to humans |
3.2 Insights for AI Agents
- Hierarchical Autonomy: Autonomous for daily operations, while high-risk decisions are handed over to humans (similar to the logic of Agent permission design);
- Stop when uncertain: Autonomous navigation stops and waits when encountering terrain that cannot be judged, rather than taking risks. For an Agent, this means "ask when uncertain" (see The Free Energy Principle as a Training Objective);
- Thorough testing in simulation: Mars rover software is repeatedly tested in ground replicas and simulated environments.
3.3 Future Directions
- Stronger onboard computing power to allow more complex vision models to go into space;
- Collaboration between multiple probes (e.g., a helicopter scouting the path for a rover);
- More distant missions (e.g., moons of Jupiter or Saturn) have longer communication delays, requiring even higher levels of autonomy.
Summary
| Layer | Key points |
|---|---|
| Motivation | Communication latency makes remote control unfeasible, necessitating autonomy |
| Achievements | Perseverance AutoNav, Terrain-Relative Navigation during landing, AEGIS, Ingenuity's 72 flights |
| Characteristics | Limited computing power, zero tolerance for error, must be verifiable |
| Insights | Hierarchical autonomy, stop when uncertain, thorough testing in simulation |