Noematrix releases RoboRSI, a multi-agent self-evolution framework for real-world robotics

The starting point of this discussion is a contrast. According to Machine Heart, Robocurve integrated GPT-6 Astra into a robot to execute tasks the day after its release, claiming a 95% success rate on one robot control task, higher than Fable 5.1's 40%; at the same time, output token usage was only about 1/6.2 of the latter, and cost about 1/2.3. These numbers come from a report's paraphrase, not first-party evaluation disclosure, and should be treated as source claims rather than confirmed facts.

But the same report provides evidence pointing in the opposite direction. The RoboDojo team noted in its official evaluation that Astra repeatedly issued physically unreasonable or unsafe actions in real-machine testing, with some incidents even causing hardware damage; for safety reasons, the team stopped testing early, and Astra failed to complete the originally planned RoboDojo-Real evaluation containing 18 tasks.

Putting the two together yields a cautious inference: a foundation model's performance on controlled single tasks and its ability to work safely and stably over time in a real physical environment are not the same thing. The former is a demonstration of capability ceiling; the latter depends on how to constrain actions, handle failures, and decide from which node to continue—these are classified in the report as system-level problems.

The report records two voices from the industry. A founder of a well-known embodied AI company said today's embodied AI has no real moat, general model companies have advantages in talent, compute, and capital, and the next one to two years will be a critical window. Robocurve mentioned that if this trend continues, large language models could achieve real-time control of robotic arms as early as the end of this year, and at the latest by 2029. Both are source-side judgments, not established facts.

The report itself also offers a view: embodied models obtained merely by fine-tuning on a semantic base are likely to suffer dimensional reduction from GPT-6. This is likewise a commentary judgment, but it explains why the discussion focus shifted from 'can the model control a robot' to 'what else is needed beyond the model.'

According to the report's narrative, the foundation model solves the understanding problem, but between understanding and sustained stable execution there is still a layer of system capability. This layer is summarized as three things: structured management and reuse of experience, behavioral boundaries and safety constraints, and continuous iteration capability after deployment.