OpenAI 重新分配机器人研究人员至其他团队,因通用模型可能使专用机器人 AI 变得不必要

独立观察 The Decoder 原文语言:英文

OpenAI 重组机器人团队,将研究人员并入主研究组和应用组,认为 GPT-6 Astra 等通用模型可直接控制机器人;Nvidia 研究显示 Astra 在模拟任务中成功率 74.8% 对比专用模型 51%,StationeryBench 基准测试中 Astra 完成 7/100 任务而 MolmoAct2 为 0。OpenAI 曾于 2020 年关闭机器人部门,2024 年重建,目标 2027 年公开机器人演示。

全文

Update AI in practice

Copy the url to clipboardShare this articleGo to comment section

OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary

Matthias BastianView the LinkedIn Profile of Matthias Bastian

Nov 14, 2026

GPT-Image-2 prompted by THE DECODER

Update – Nov 14, 2026

The opening sentence of this article originally stated that OpenAI was "apparently disbanding its robotics team again." OpenAI has since clarified that it reorganized the team to work more closely with research and product and that robotics work continues. The opening sentence has been updated to reflect this.

OpenAI has reorganized its robotics team, though for different reasons than when it first disbanded the unit in 2020.

Rather than abandoning robotics, the company wants to accelerate it by folding researchers into other teams. Robotics researchers are moving to the main research team under Chief Research Officer Mark Chen, while the applied robotics and hardware teams join Greg Brockman's Applied organization.

OpenAI made the decision on such short notice that it had already booked a booth at the Conference on Robot Learning (CoRL) in Austin, which then sat empty, Fortune reports. OpenAI apparently believes that general-purpose models like GPT-6 Astra can control robots directly, reducing the need for specialized robotics models.

In an Nvidia study, an Astra-powered agent achieved a 74.8 percent success rate on simulated robotics tasks, compared to 51 percent for the best specialized models. The tests used a simulated DROID robot, not a physical one.

A new robotics benchmark called StationeryBench compared GPT-6 Astra with Ai2's MolmoAct2 on five tasks involving desk objects on dual-arm YAM robotic arms. Across 200 trials, Astra fully completed 7 out of 100 tasks, while MolmoAct2 didn't complete a single one. Astra's average progress score was 46, compared to 12 for MolmoAct2. In another test, Astra running on a humanoid robot navigated and cleaned up a kitchen.

It remains unclear how much robotics data OpenAI and others incorporated during training. Other large general-purpose models from Anthropic and Google have also made recent progress on robotics tasks. Specialized robotics models may still have advantages in cost and reliability as well as speed and energy consumption, even if they don't generalize as well.

Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, also sees Astra as a major leap in spatial understanding. On the still-unpublished REMAP benchmark, GPT-6 Astra reaches near-human accuracy, according to Artzi. He suspects that OpenAI may have trained the model on large amounts of 3D training data such as Blender scenes.

OpenAI and robots, it's complicated

OpenAI shut down its robotics division in 2020 because, according to co-founder Wojciech Zaremba, the company lacked the training data it needed and large language models were a faster path to AGI. The company rebuilt the team in 2024. CEO Sam Altman is reportedly targeting a first public robot demo for 2027, and OpenAI wants to build its own everyday robots in the long term. But the company also faces heavy financial pressure to focus on its core AI business, where it has lost significant ground to Anthropic.

正文以来源语言显示。

参考来源

← 返回每日简报