Axiom 的三位人工智慧工程师 Aditya Ramabadran、Simon Mahns 与 Tobias Gessler 进行了一项实验，将 OpenAI 的 GPT-6 Astra 部署于一辆约 1.81 公吨（2 short tons）的 2024 年款 Toyota Corolla 上。该模型连结了挡风玻璃摄影机与动力转向系统，成功在加州湾区的一间 In-N-Out 得来速车道中引导车辆行驶以领取餐点。这项测试凸显了通用语言模型在零样本或未经专门微调的情况下，直接与真实物理世界互动的全新潜力。
学术界与产业界正加速评估人工智慧在物理世界中的理解与推理能力。由前 Google DeepMind 研究员 Andrew Dai 创立的 Elorian AI 与 Scale AI 合作建立了「Humanity's Sixth Sense」基准测试，以量化模型直觉理解物理场景的能力。研究人员认为，空间与物理推理能力的突破，将是实现餐厅服务、家庭机器人等复杂环境应用的关键门槛。
该工程团队进一步建立了名为 DrivingBench 的停车场封闭路线基准测试，以客观量化不同大型语言模型的驾驶表现。测试数据显示，仅有 GPT-6 Astra 能够以极低速完整通过测试路线；相较之下，Claude Fable 5.1 仅完成路线的 45%，而 Grok 则仅达成 11%。即便如此，工程师指出最新模型已展现即时错误修正与情境内学习（in-context learning）的能力，可动态适应车辆操控。
Three AI engineers at Axiom, Aditya Ramabadran, Simon Mahns, and Tobias Gessler, conducted an experiment deploying OpenAI's GPT-6 Astra into a 1.81-metric-ton (2-short-ton) 2024 Toyota Corolla. Interfacing the model with windshield cameras and the power steering system, Astra successfully navigated an In-N-Out drive-thru lane in the Bay Area to retrieve food without specialized prior driving fine-tuning, demonstrating emergent zero-shot physical navigation capabilities in multimodal language models.
The broader artificial intelligence industry is increasingly focusing on physical and spatial reasoning benchmarks. Elorian AI, founded by former Google DeepMind researcher Andrew Dai, partnered with Scale AI to launch Humanity's Sixth Sense, a benchmark evaluating intuitive physical scene comprehension. Dai emphasized that robust physical reasoning is an indispensable foundation for enabling robotic systems to operate effectively in complex environments like homes and restaurants.
To quantitatively evaluate vehicular performance, the trio developed DrivingBench, testing models on a simple parking lot obstacle course. Experimental results showed significant performance disparities: only Astra completed the entire course at low speeds, while Claude Fable 5.1 completed 45 percent and Grok achieved just 11 percent. Notably, engineers observed real-time error correction and in-context learning adaptations as models learned to manipulate controls dynamically.