How CADRE passed its autonomy testing

Getting missions to land successfully on the moon has been difficult. Recent missions, such as IM-1 and IM-2, which the private company Intuitive Machines completed, have been qualified successes at best, with both landers settling at unintended angles and breaking parts of them off along the way. Such experiences offer excellent learning opportunities, though, and NASA is confident that a third time might be a charm for a flawless mission.

There will be a lot riding on IM-3, the third Commercial Lunar Payload Services (CLPS) mission, including a set of rovers and ground station for a NASA experiment called the Cooperative Autonomous Distribution Robotic Exploration (CADRE), which recently passed its Verification and Validation (V&V) test for one of its most essential parts. This software architecture handles tasks for each rover and binds them into a cohesive whole.

As reported, CADRE consists of three rovers and a ground station. The mission intends to link all four components into a cohesive, multi-agent system that works together to create a 3D map of the surface and sub-surface area around the moon’s Reiner Gamma region, where it should land somewhere before the end of 2026. It requires a software architecture that handles planning, scheduling, and execution (PS&E).

PS&E is a problematic concept in multi-agent autonomous robots, as each robot has its own environment in which they are a part. They can’t necessarily understand how/where another collective member is. Nor can they even guarantee that the members are still functional. That is especially true in harsh environments like the lunar surface, where one rover could end up on the side of a hill and get significantly less sunlight, forcing it to shut off earlier than its peers. With improper networking, such a fate could affect the entire group rather than just one individual rover.

CADRE’s PS&E architecture is designed to avoid that kind of problem. It consists of four modules, as described in a paper recently released by engineers at NASA’s Jet Propulsion Laboratory (JPL), which is responsible for building it. The four modules are leader election, a shared state database, a strategic planning module, and individual agent controllers. The work is available on the arXiv preprint server.

The leader election module assigns a single agent to plan and coordinate between the distributed assets of the system. Critically, the leader can be changed as part of this module if the agent currently elected to the position is incapacitated. Similar to the “designated survivor” famous for being the only cabinet member in the US to intentionally miss the President’s State of the Union address, another member can hold a backed-up copy of the status of the elected leader, and the system can switch to operating off of that in the event the leader cannot perform its role anymore.

Shared state databases (SSDs) are key to many underlying software architectures, and CADRE’s system is no exception. SSDs allow coordination in understanding current states and tasks. However, with CADRE, the system, which utilizes a custom design application layer, does not attempt to force synchronization between all agents. That is what the leader and designated survivor are for, and the specific instance of a database on an individual agent is not as crucial given this architectural choice.

Evaluating and scheduling tasks is the purview of the strategic planning module, as might be expected given the name. However, executing the tasks is assigned to the specific agent best placed to accomplish them, which is handled by the fourth module, the agent controller. The agents will then report to the leader about their progress and await further instructions as the leader updates its own models and executes the newly decided action plan.

The central point of the V&V testing described in the new paper was testing all of this interconnected communication and planning. The testing ranged from simulations to actual physical implementation on hardware rover analogs. The engineers also tested it outside on different terrains to make the simulated environment on Earth as similar to the moon as possible, where the overall system will eventually land.

With the successful completion of V&V testing of its planning and execution architecture, CADRE can move on to the next step in its development cycle. Even though its integration into its launcher has started, it still has a while to go before launch in 2025/2026, but soon, we might see a “wolf pack” of rovers driving over the lunar surface.