GPT-6 Astra Excels at Robot Control, but Real-World Safety Fails

A day after GPT-6 Astra's release, Robocurve connected it directly to a robot for task execution and found it performed strongly in robot control while consuming fewer tokens. In one robot-control task, GPT-6 Astra achieved a 95% success rate, compared with 40% for Fable 5.1, while using about one-sixth of the output tokens and costing about half as much.

The results fueled a wave of demos in which GPT-6-based models controlled robots for complex operations. A founder of a well-known embodied AI company said today's embodied intelligence has no real moat, and general-model companies have advantages in talent, compute, and capital; the next one to two years will be a critical window.

Strong single-task scores do not mean a foundation model can safely and stably control robots in real environments. The RoboDojo team noted in its official evaluation that Astra repeatedly issued physically unreasonable or unsafe actions during real-machine testing, with some incidents causing hardware damage.

For safety reasons, the team stopped testing early, so Astra did not complete the planned RoboDojo-Real evaluation of 18 tasks. The lesson is that breakthroughs in foundation-model capability are only a starting point; constraining actions, handling failures, and enabling robots to work safely and stably over time remain system-level problems.

Noematrix released RoboRSI, a multi-agent self-evolution framework for robots in real complex scenarios. It seeks to answer how, once a base model has strong understanding and reasoning, to build a complete loop of continuous execution, diagnosis, revision, and reuse at the system level.

RoboRSI splits planning, execution, diagnosis, and revision into different contexts and uses a collaborative architecture of four agent types: Manager, Planner, Engineer, and Reviewer. These four roles work in relay around the same task and the same execution evidence, forming a loop of execution, diagnosis, revision, and re-execution. Humans retain control over goals, value judgments, and safety boundaries.

Noematrix proposes the TSR mechanism, using a four-layer skill tree to organize all robot capabilities. Each modification is limited to clear skill boundaries; agents can focus on local implementation and revision without drifting from the global task goal during continuous debugging. Historical experience thus shifts from dialogue records and logs to capabilities that can actually be called in the next round.

Noematrix also designed a three-stage evolution mechanism: extract stable structures from successful call chains, parameterize object categories, target regions, and spatial relationships, and then solidify reusable processes into code skills. When similar tasks appear again, agents only need to select, monitor, and revise the entire skill if necessary. In LIBERO benchmark experiments, enabling code solidification raised episode pass rate from 21.5% to 29.0% and cut median token consumption by 29.4%, compared with the version without it.

Under RoboRSI, building the same level of capability completed a full loop within one day. In 2024, completing a similar household floor-cleaning demo required two engineers working continuously for about a month and writing around 2,000 to 3,000 lines of task code.

If a system can autonomously close the loop of task decomposition, skill implementation, execution diagnosis, and code revision, the core cost of adapting to a new scenario shifts from weeks of on-site engineers to setting goals plus autonomous iteration plus human review and safety checks.

This change opens new commercial space for highly personalized scenarios such as homes. Every household has different layouts, furniture arrangements, and object placements. If each deployment requires engineers on site for weeks, service costs cannot support large-scale rollout. Shortening the adaptation cycle means scattered scenarios that were previously too costly to cover can become serviceable markets.