SenseNova Study: GPT-6-Astra Clears Robot Competence Threshold

"GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence," write the authors of a new SenseNova-led evaluation on arxiv that tests the frontier model across 12 robotics benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands.

Under a unified protocol that lets the model use in-context learning from training-set demonstrations and substantially increases the number of execution steps, the paper reports "unprecedented performance." The agentic harness behind those numbers pairs a high-level "System 2" reasoner, GPT-6-Astra itself, with a low-level "System 1" controller such as π0.5. That coupling, the authors write, is "costly but effective."

The real-world half of the study is less flattering. Experiments across 8 robot configurations surface two concrete gaps: the robot demonstrations required for in-context learning are "expensive to collect," and "it is implausible to trade longer execution for higher success rates with physical constraints." The paper lists five open problems for standalone GPT-6-Astra: generality of ICL, cost and execution time, physical-state verification, recovery from irreversible failures, and long-horizon preservation of progress. It also sketches a prototype next-generation benchmark that derives tasks from real-world needs, provides no in-domain demonstrations, and "judges agents on a world clock that keeps running while they deliberate."