# Robot control preprint finds demonstrations improve agent performance on task subset

_Published Thursday, November 5, 2026 at 11:53 PM EST · Robotics, AI, Science · Latest · Tier 2 — Notable_

A RoboARCade preprint reports that one successful demonstration raised GPT-6 Astra's robot control score from 53.5% to 73.0% on a five-task subset. The researchers introduced a simulation benchmark of long-horizon manipulation tasks across three robot embodiments to test adaptation, memory and precise control.

Across the benchmark, Astra scored 49.8% and Claude Fable 5.1 scored 43.6%. A richer visual interface added 1.5 to 6.8 points at 1.7 to 2.3 times the cost. Failures mainly involved poor manipulation strategies or imprecise execution; agents incorrectly claimed success on 24% to 39% of attempted subtasks.

## Sources

- [arXiv Query: search_query=cat:cs.RO&id_list=&start=0&max_results=30](https://arxiv.org/abs/2611.08209v1)

---
Canonical: https://techandbusiness.org/newswire/Becl18y83z9W6kaEUK8gXL
Published: 2026-11-06T04:53:17.002Z
Story chronology: 2026-11-05T16:00:16.000Z
Retrieved: 2026-11-06T23:13:22.684Z
Publisher: Tech & Business (techandbusiness.org)