Odyssey 已推出 Odyssey-3，这是一个世界模型。据这家位于帕洛阿尔托的人工智能实验室称，该模型能够控制机械臂和人形机器人、驾驶汽车、操控无人机以及玩电子游戏。该公司周四表示，目前已有研究预览版本可供使用。
Odyssey has launched Odyssey-3, a world model that the Palo Alto AI lab says can control robot arms and humanoids, drive a car, fly drones and play video games. A research preview is available now, the company said on Thursday.
世界模型通过大量视觉数据学习，以预测场景的后续变化，包括各种动作对事件发展的影响。Odyssey 表示，其新模型能让机器在掌握特定任务之前，先对世界形成一种普遍的理解。
World models learn from large amounts of visual data to predict how a scene will change next, including how actions affect what happens. Odyssey says its new model gives machines a general understanding of the world before they learn a specific task.
Odyssey 的联合创始人兼首席执行官 Oliver Cameron 表示：“Odyssey-3 是迈向单一智能体的重要一步，这种智能体能够理解并与我们周围的世界互动。”
“Odyssey-3 is a big step toward a single intelligence that can understand and operate in the world around us,” said Oliver Cameron, Odyssey’s co-founder and chief executive.
Odyssey 称，在 Physics-IQ Verified 的视频转视频测试中，Odyssey-3 Pro 取得了 66.1 分，为该排行榜上报告的最高分数。这一分数取自每项任务中八次尝试的最佳结果。该基准测试由 Anates Labs 和 DeepMind 推出，要求模型续接真实物理实验的视频。参与比较的模型还包括来自 Nvidia 和 Black Forest Labs 的产品。
The benchmark claims Odyssey-3 Pro scored 66.1 on the video-to-video test of Physics-IQ Verified, the highest score reported on that leaderboard, according to Odyssey. That score uses the best of eight attempts per task. The benchmark, from Anates Labs and DeepMind, asks models to continue videos of real physics experiments. The comparison includes models from Nvidia and Black Forest Labs.
Odyssey 还表示，该模型在 WorldMark 基准测试的四个类别中位列前三。这一结果来自该公司自身的评估。
Odyssey also says the model ranked first in three of the four categories of the WorldMark benchmark. That result comes from the company’s own evaluation.
Odyssey 展示了该模型在六个领域中的六种机器表现。在数十小时的机器人演示数据训练下，它能够完成机械臂操作任务。此外，它还能应对演示数据中未包含的错误，例如抓取失败。
Six kinds of machine Odyssey showed the model in six areas. With tens of hours of robot demonstrations, it completed tasks with robot arms. It also recovered from errors that were not in those demonstrations, such as a missed grasp.
人形机器人公司 Flexion 已在 Odyssey-3 上构建了自己的控制策略。Odyssey 称，这些策略能够处理光照变化，而光照变化曾导致其他基准模型测试失败。
The humanoid robot company Flexion has built control policies on Odyssey-3. Odyssey says they handled lighting changes that caused the tested baselines to fail.
该公司还利用在约 20 小时数据上训练的驾驶策略，在印度的真实道路上驾驶汽车。Odyssey-3 还能在室内引导无人机、生成用于训练人工智能代理的环境，以及运行《侠盗猎车手 V》。
The company also drove a car on real roads in India with a driving policy trained on about 20 hours of data. Odyssey-3 also guided drones indoors, generated environments for training AI agents and played Grand Theft Auto V.
从3.1亿美元到产品 自驾车老兵创立的Odyssey。6月，它以14.5亿美元的估值筹集了3.1亿美元，亚马逊和AMD Ventures等是投资者。
From $310m to a product Self-driving car veterans founded Odyssey. In June it raised $310 million at a $1.45 billion valuation, with Amazon and AMD Ventures among the investors.
Odyssey是几家构建世界模型的公司之一。上个月，AMD同意以82亿美元的股票收购Fei-Fei Li的World Labs初创公司。9月，前DeepMind世界模型研究员创立的初创公司Emulate据报正在洽谈筹集7亿美元。Nvidia提供其自己的开放Cosmos世界模型。
Odyssey is one of several companies building world models. Last month AMD agreed to buy Fei-Fei Li’s World Labs startup for $8.2bn in stock. In September, startup Emulate, founded by former DeepMind world-model researchers, was reportedly in talks to raise $700m. Nvidia offers its own open Cosmos world models.