Primary sourcearXiv cs.AI
View sourceEvent ·Multi-source
EBench 技术报告:评测 GPT-6-Astra 控制机器人的能力边界
报告在 EBench 上评测 GPT-6-Astra 作为机器人控制智能体的表现,与经过后训练的 VLA 和 WAM 策略对比。直接控制模式下,该智能体在短程任务上领先,且在泛化维度上性能波动最小,但在长程任务、桌面操作和高精度接触任务上不及后训练策略。引入策略调用的 harness 模式可将整体任务成功率提升到最强后训练策略的水平,同时明显降低 token 成本。成绩仍低于各任务单模型上限,原因是策略调用不当、缺乏子目标分解,以及智能体恢复控制时干扰抓取。
Current read
报告实测了 GPT-6-Astra 操控真实机器人:短程任务很强,长程和高精度操作还不行,调用 VLA 策略能补短板还省 token。
11 sources42° AI heatLast updated 2026/10/26 17:51:02
Evidence
11 sourcesRelated source 1AIGCLINK
View sourceRelated source 2Simon Willison’s Weblog
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4IT之家
View sourceRelated source 5marktechpost
View sourceRelated source 6Decoder
View sourceRelated source 7TestingCatalog
View sourceRelated source 8The Rundown AI
View sourceRelated source 9OpenAI Blog
View sourceRelated source 10Shubham Saboo
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.