A robot completes an action.
But does it know the task is actually finished?
SenseNova-RoboRSI documents where embodied AI still fails: misjudging success, forgetting completed subtasks, undoing earlier progress, and struggling with irreversible mistakes.
The research is open here:
arxiv.org/abs/2610.22471
But does it know the task is actually finished?
SenseNova-RoboRSI documents where embodied AI still fails: misjudging success, forgetting completed subtasks, undoing earlier progress, and struggling with irreversible mistakes.
The research is open here:
arxiv.org/abs/2610.22471
We release the most comprehensive ๐ฒ๐บ๐ฝ๐ถ๐ฟ๐ถ๐ฐ๐ฎ๐น ๐๐๐๐ฑ๐ ๐ผ๐ณ ๐๐ฃ๐ง-๐ฒ-๐๐๐๐ฟ๐ฎโ๐ ๐ฒ๐บ๐ฏ๐ผ๐ฑ๐ถ๐ฒ๐ฑ ๐ถ๐ป๐๐ฒ๐น๐น๐ถ๐ด๐ฒ๐ป๐ฐ๐ฒ across 12 simulation benchmarks and 8 real-robot platforms, together with ๐ฆ๐ฒ๐ป๐๐ฒ๐ก๐ผ๐๐ฎ-๐ฅ๐ผ๐ฏ๐ผ๐ฅ๐ฆ๐: an open-source harness framework and t ~1/50 the cost of GPT-6-Astra.
Here's what we found:
๐น ๐ง๐ต๐ฒ ๐ฐ๐ฎ๐ฝ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ ๐ฐ๐ฒ๐ถ๐น๐ถ๐ป๐ด ๐ถ๐ ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐บ๐๐ฐ๐ต ๐ต๐ถ๐ด๐ต๐ฒ๐ฟ. Under a unified protocol (ICL + up to 10ร the interaction budget), GPT-6-Astra achieves clear SOTA across all 12 benchmarks: RoboDojo: 27.6% at 1ร โ 81.9% at 10ร, DexJoCo: 87.3% vs. 40.3% prior SOTA, while the LIBERO benchmarks are pushed to saturation.
๐น ๐ง๐ผ๐ผ ๐บ๐๐ฐ๐ต ๐ฏ๐๐ฐ๐ธ ๐ณ๐ผ๐ฟ ๐๐ต๐ฒ ๐ฏ๐ฎ๐ป๐ด. Even a simple task like placing pens in a holder costs ~$55 in inference.
๐น ๐ง๐ต๐ฒ ๐๐ผ๐ฟ๐น๐ฑ ๐ฑ๐ผ๐ฒ๐๐ปโ๐ ๐๐ฎ๐ถ๐. Time matters as real robots overheat, and an egg burned while the model was still thinking.
Our takeaway: GPT-6-Astra has largely crossed the threshold of ๐ฒ๐บ๐ฏ๐ผ๐ฑ๐ถ๐ฒ๐ฑ ๐ฐ๐ผ๐บ๐ฝ๐ฒ๐๐ฒ๐ป๐ฐ๐ฒ, but ๐ฝ๐ฟ๐ฎ๐ฐ๐๐ถ๐ฐ๐ฎ๐น ๐ฒ๐บ๐ฏ๐ผ๐ฑ๐ถ๐ฒ๐ฑ ๐ถ๐ป๐๐ฒ๐น๐น๐ถ๐ด๐ฒ๐ป๐ฐ๐ฒ is not here yet.
๐ Project page: sensenova.ai/has-gpt-6-astrโฆ
๐ Report: sensenova.ai/gpt-6/assets/dโฆ
๐ป Code & model: github.com/OpenSenseNova/โฆ
Oct 22, 2026 ยท 10:15 AM UTC
8
40
167
26,514
@Damn_coder Forgetting completed tasks sounds like a serious issue for long-running robotics workflows.
0
0
1
49
@Damn_coder SenseNova-RoboRSI is highlighting problems that really matter outside benchmarks.
0
0
1
95
@Damn_coder Robots still have a lot to learn about fixing their own mistakes.
0
0
1
26
@Damn_coder Misjudging success and undoing progress is such key embodied failure
0
0
0
15
@Damn_coder Task completion sounds simple until a robot has to judge it.
0
0
0
18