Shield AI, Sedaro and Novi Space announced the completion of a demonstration of credible autonomy on-orbit on 24 August. For the first time in a multi-wheel experiment, Shield AI ' s Livemind Autonomous Software operated on low-Earth orbit satellites; a NOVI real satellite was set up as a team leader node of six satellites and five virtual satellites. Livemind works with Sedaro ' s SAFE margin autonomy framework for mission planning, validated and validated by NOVI operators.
In the 24-hour experiment window, the real spacecraft implemented 189 commands generated by Hivemind and approved by SAFE. Systems need to balance imaging missions, spacecraft health, battery charge levels and formation direction at the same time. In another experiment, the SAFE and Sedaro edges could deploy simulators to implement similar multi-target optimizations, which increased the percentage of Iridium communications initiated by satellites by 8 percentage points. Three figures indicate that the experiment did indeed enter the real orbital implementation phase, but do not mean that the satellite has been de-supervised or that the constellations have achieved full autonomy.
The key to credible autonomy is not to generate a plan, but to validate it before implementation.
With the expansion of the constellation, the ground team is not facing a single vehicle, but a constantly changing communications window, energy constraints, heat, attitude and mission priorities. Traditional manual processes can carefully handle a small number of spacecraft, but it is difficult to expand linearly with the number of nodes. The value of an autonomous system lies in the quick-calculating combination programme, which assigns tasks to the appropriate nodes; the risk is that the wrong command may consume limited power, miss communications windows and even affect spacecraft safety.
The presentation separated Hivemind's mission decision from Sedaro's certification floor. Livemind proposes optimal action at the formation level, where SAFE uses predictive models of spacecraft and their environment to check the results before the instructions are actually implemented. The 189 orders are therefore not evidence that “AI can control the satellite at will”, but rather a closed-ring sample consisting of generation, simulation, approval and execution. Press releases do not disclose the type of instruction, the number of programmes rejected, the number of manual interventions or the complete efficiency compared to the manual baseline, so that the outside world cannot rely solely on successful announcements to infer that the system is reliable under all unusual conditions.
The mixing of real and five virtual satellites also requires accurate understanding. It allows teams to test distributed task coordination without launching a full six-star formation and to place real telemetry, in-orbit computing and simulation members in the same environment. Its value is to reduce the cost of testing and to expand the coverage of the scene; the limitation is that virtual nodes do not reproduce all hardware drift, communications packages, space weather and multi-star interactions. More real stars, more orbital cycles and fail-injecting tests are still needed to prove the autonomy of the entire real constellation.
From vehicle to spacecraft is capability migration, not full product deployment
Livemind used to be primarily oriented towards aviation platforms, and this is the first time that he has entered the space field and the highest-level and longest-duration deployment to date. Shield AI stated that the demonstration showed that the satellite, which was not originally designed for autonomous operation, could also be superseded to organize intelligence, surveillance and reconnaissance missions. This is a positive signal of compatibility, but it is still a technical demonstration that does not mean that existing satellites can be upgraded without modification, much less that the relevant capabilities are already operational on a large scale in operational or military constellations.
The increase of 8 percentage points in the launch rate of communication links in the second experiment should also be interpreted in the context of specific indicators. It describes a change in the performance of the Iridium connection launch under one of the experimental conditions, which does not amount to an 8 per cent increase in the bandwidth of communications, the successful downloading of data or the overall increase in mission benefits. If more communication windows are accompanied by additional energy consumption, the system still needs to weigh mission value against spacecraft lifetime. The significance of multi-target optimization lies precisely in the fact that no single indicator can represent all results.
The three companies announced strategic collaboration at the end of 2025: Hivemind became the preferred autonomous software for Sedaro's on-orbit demonstration, while the Sedaro platform was used for space scenario development, testing and validation. The result of this exercise was to advance cooperation from ground simulation to on-orbit implementation. Part of the study was funded by the United States Government, with the official indication that publication should not be interpreted as representing an explicit or implicit policy of the United States Government. This qualification indicates that there is still a gap between technical milestones and government procurement and deployment decisions.
For satellite operators, the next phase of attention should include how to retreat when the command fails, who has the final authority to predict a model and a real spacecraft deviation, how software updates are certified, and to what extent the system is allowed autonomy during a communication interruption. Credible autonomy is not the removal of operators, but the upgrading of human control from an article-by-article order to the setting of targets, competencies and stoppage boundaries. The presentation provided real evidence in orbit and left a sufficiently clear checklist for validation: longer duration, more real nodes, public failure samples and quantifiable changes in manual workload to determine whether it could move from experiment to normal operation.
Source: Shield AI, Shield AI and Sedaro demonstrate trusted autoony Capabilities on NOVI Satellite, 24 August 2026, https://shield.ai/shield-ai-and-sedaro-demontate-trusted-autonomy-capabilities-on-novi-satellite/