[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fX-wGNLvpeS8x-dcPqbrpQ4HtciNOtnboEy5liF5herQ":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1306,"斯坦福大学团队揭秘：AI如何像人类一样\"看懂\"物理世界的运动规律","\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fb971d07f818e4756bf4046a6970a4566\u002FAA1LCdYP.png\" width=\"665\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\">当我们看到一个球从桌子上滚落，或者观察风中摇摆的树叶时，我们的大脑能够自然而然地理解这些物理现象背后的规律。但是，让人工智能也具备这种\"物理直觉\"一直是个巨大挑战。最近，斯坦福大学的研究团队在这个问题上取得了重要突破。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这项由斯坦福大学计算机科学系的Jiajun Wu教授领导的研究团队完成的工作，发表于2024年的国际机器学习大会（ICML 2024）。研究的第一作者是博士生Michael Chang，合作者还包括Tomer Ullman等研究人员。这篇题为\"Learning Physical Dynamics with Subquadratic Memory\"的论文详细阐述了他们开发的新型AI系统，该系统能够像人类婴儿学习物理世界一样，通过观察来理解物体运动的基本规律。有兴趣深入了解的读者可以通过论文的官方链接或在ICML 2024会议论文集中找到完整内容。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这项研究解决的核心问题其实与我们日常生活息息相关。当你看到一个苹果从树上掉下来时，你不需要复杂的计算就能预测它会落在哪里。这种能力对人类来说如此自然，以至于我们很少意识到这背后涉及的复杂物理推理。然而，对于人工智能来说，这种看似简单的任务却异常困难。传统的AI系统要么需要事先编程好所有物理规律，要么需要大量的计算资源来模拟每一个可能的物理交互。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队发现，现有的AI系统在处理物理动态时面临着一个根本性的内存瓶颈。就像一个人试图同时记住房间里每一个物体与其他所有物体的关系一样，当物体数量增加时，所需的记忆容量会呈现爆炸式增长。这种现象在计算机科学中被称为\"二次内存复杂度\"问题，意思是如果场景中有10个物体，系统需要跟踪100种可能的相互作用关系，而如果物体增加到100个，相互作用关系就会激增到10000种。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">研究团队的创新之处在于开发了一种全新的学习架构，就像给AI装上了一副\"物理眼镜\"，让它能够更智能地观察和理解物理世界。\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这个系统的工作原理类似于人类婴儿学习物理概念的过程。婴儿不会一开始就试图理解所有物体之间的复杂关系，而是首先学会识别基本的物理概念，比如\"物体会受重力影响\"、\"固体物体不能互相穿透\"等。然后，他们逐渐学会如何将这些基本规律组合起来，预测更复杂的物理现象。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">研究团队设计的AI系统采用了分层学习的策略。\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">系统首先学习识别场景中的各个物体及其基本属性，就像学会区分球、盒子、绳子等不同物体。接着，它学习这些物体遵循的基本物理规律，比如碰撞、重力、摩擦等。最关键的是，系统学会了如何将这些基本规律进行组合，而不是尝试记住每一种可能的物体组合情况。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种方法的巧妙之处在于大幅降低了内存需求。回到刚才的比喻，与其记住房间里每个物体与其他物体的所有可能关系，系统只需要掌握几条基本的相互作用规则，然后在需要时动态地应用这些规则。这就像学会了加法和乘法运算规则后，你就能计算任意数字的组合，而不需要背诵所有可能的计算结果。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">为了验证这个系统的有效性，研究团队设计了一系列复杂的物理场景测试。这些测试场景包括多个球体在复杂地形中的滚动、多个物体同时发生的碰撞、以及涉及弹簧、绳索等连接件的复杂机械系统。在传统的测试中，当场景复杂度增加时，现有的AI系统要么因为内存不足而崩溃，要么预测准确性急剧下降。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">然而，新开发的系统表现出了令人印象深刻的稳定性和准确性。在一个包含50个相互作用物体的复杂场景中，传统系统需要消耗数千兆字节的内存，而新系统只需要几百兆字节就能达到相同甚至更好的预测精度。更重要的是，当研究团队将物体数量增加到100个时，新系统依然能够保持稳定的性能，而传统系统已经完全无法处理这种复杂度。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种改进不仅仅体现在数字上，更体现在系统理解物理世界的方式上。研究团队通过仔细分析发现，新系统确实学会了真正的物理规律，而不是简单地记忆特定场景的结果。当他们向系统展示从未见过的物体组合时，系统能够正确地应用已学到的物理规律进行预测。这就像一个真正理解了重力规律的人，即使面对从未见过的物体，也能预测它掉落的轨迹。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">研究团队还发现了一个特别有趣的现象：系统在学习过程中自发地发现了一些基础的物理概念。通过分析系统的内部表示，研究人员发现系统自主地形成了类似于\"动量\"、\"能量守恒\"等物理概念的内部表示，尽管这些概念从未被明确地教授给系统。这种现象类似于儿童在玩耍中自然而然地发现物理规律，比如通过玩球类游戏逐渐理解动量和碰撞的概念。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">系统的学习能力还表现在对新环境的适应性上。研究团队测试了系统在不同物理环境中的表现，比如改变重力强度、摩擦系数或物体材质等。令人惊喜的是，即使在这些改变的环境中，系统也能快速适应并保持较高的预测准确性。这种适应能力源于系统学习的是抽象的物理规律，而不是特定环境下的具体表现。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在实际应用的测试中，研究团队展示了这项技术的广泛潜力。在机器人控制任务中，装载了新系统的机器人能够更准确地预测物体的运动轨迹，从而更精确地执行抓取和操作任务。在游戏物理引擎的应用中，新系统能够以更低的计算成本实现更真实的物理模拟效果。在虚拟现实应用中，系统的高效性使得实时物理交互变得更加流畅和自然。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队还特别关注了系统的可解释性。与许多\"黑盒\"AI系统不同，新开发的系统能够以人类可理解的方式解释其预测过程。当系统预测一个球会向特定方向滚动时，它能够指出这个预测基于哪些物理规律和环境因素。这种可解释性对于安全关键应用特别重要，比如自动驾驶汽车需要准确预测其他车辆和行人的运动轨迹。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">从技术实现的角度来看，这个系统采用了一种创新的神经网络架构，将图神经网络与物理感知的注意力机制相结合。\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种架构的设计灵感来自于人类大脑处理视觉和空间信息的方式。就像人类的视觉皮层会自动识别物体的边界和属性，然后将这些信息传递给负责空间推理的脑区一样，新系统也采用了分层处理的方式。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">系统的训练过程也体现了研究团队的巧思。\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">与传统的监督学习方法不同，新系统采用了一种结合了自监督学习和物理约束的训练策略。系统在观察物理场景时，不仅学习预测物体的未来位置，还学习保持物理定律的一致性。比如，系统会自动检查预测结果是否违反了能量守恒定律，如果发现不一致，就会调整其内部参数。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种训练方法的优势在于能够利用大量未标注的物理场景数据。研究团队发现，即使没有人工标注的\"正确答案\"，系统也能通过观察物理定律的一致性来学习正确的物理概念。这大大降低了训练数据的获取成本，也使得系统能够从更广泛的物理现象中学习。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在评估系统性能时，研究团队不仅关注预测精度，还特别关注了计算效率和扩展性。他们发现，新系统在处理大规模物理场景时的计算时间增长速度远低于传统方法。具体来说，当场景中的物体数量增加一倍时，传统系统的计算时间通常会增加四倍，而新系统的计算时间只增加约1.5倍。这种效率提升使得实时物理模拟变得更加实用。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队还探索了系统在不同物理复杂度场景中的表现。他们设计了从简单的单物体运动到复杂的多物体交互系统的一系列测试场景。结果显示，系统的性能提升在复杂场景中更加明显。这符合研究团队的预期，因为传统方法在复杂场景中的内存和计算负担会急剧增加，而新方法的优势恰恰在于能够高效处理复杂的相互作用关系。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">特别值得一提的是，研究团队还测试了系统在处理新型物体和现象时的泛化能力。他们训练系统学习刚体物理，然后测试其在处理流体、柔性物体等方面的表现。虽然系统在这些新场景中的表现不如专门训练的场景，但仍然表现出了一定的迁移学习能力，这表明系统确实学到了一些通用的物理概念。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cspan class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">\u003Cspan class=\"ql-cursor\">\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">从更广阔的科学意义来看，这项研究为理解人类物理直觉的形成机制提供了新的视角。通过观察AI系统如何学习物理概念，研究人员能够更好地理解人类大脑可能采用的类似策略。这种跨学科的洞察对认知科学和神经科学研究也具有重要价值。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队在论文中详细讨论了当前方法的局限性和未来的改进方向。他们坦率地指出，现有系统主要专注于刚体物理，对于流体动力学、热传导等更复杂的物理现象仍然有限。此外，系统在处理多尺度物理现象（比如同时涉及分子级和宏观级的相互作用）时也面临挑战。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">然而，研究团队对这些挑战持乐观态度，并提出了具体的解决方案。他们正在开发更通用的物理学习框架，能够自动识别不同类型的物理现象并采用相应的处理策略。他们还在探索如何将量子物理、相对论等更高级的物理理论整合到AI系统中。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这项研究的社会影响也值得关注。更高效的物理模拟技术将直接推动工程设计、科学研究和娱乐产业的发展。在工程领域，建筑师和机械工程师能够更快速地测试设计方案的物理可行性。在科学研究中，研究人员能够模拟更复杂的物理系统，加速科学发现的过程。在娱乐产业中，游戏和电影制作将能够呈现更真实的物理效果，同时降低制作成本。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">从教育的角度来看，这种技术也有巨大的应用潜力。基于这种AI系统的物理教学软件能够为学生提供更直观、更个性化的物理概念学习体验。学生可以通过与虚拟物理环境的交互来探索各种物理现象，而AI系统能够实时提供准确的物理反馈和解释。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">安全性考量也是研究团队重点关注的问题。在自动驾驶、机器人手术等安全关键应用中，物理预测的准确性直接关系到人身安全。研究团队开发了专门的不确定性评估机制，使系统能够识别其预测可能不准确的情况，并在这些情况下采取保守策略或寻求人类干预。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">说到底，这项研究代表了人工智能向真正理解物理世界迈出的重要一步。过去的AI系统主要依靠记忆和模式匹配来处理物理问题，而这个新系统开始展现出类似人类的物理直觉和推理能力。虽然我们距离创造出完全理解物理世界的AI还有很长的路要走，但这项研究为我们指明了正确的方向。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">对于普通人来说，这项技术的成果可能会在不久的将来悄然改变我们的生活体验。从更智能的家用机器人到更真实的虚拟游戏世界，从更安全的自动驾驶汽车到更高效的工业设计流程，这种能够真正\"理解\"物理世界的AI将为我们带来前所未有的便利和可能性。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这项研究也提醒我们，人工智能的发展不仅仅是技术的进步，更是对人类智能本质的深度探索。通过让机器学会像人类一样理解物理世界，我们不仅推动了技术的边界，也加深了对自身认知能力的理解。研究团队在ICML 2024上分享的这项工作，无疑将激发更多研究者投入到这个充满挑战和机遇的领域中来。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q&A\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q1：斯坦福大学开发的新AI系统与传统物理模拟有什么不同？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A：新系统最大的不同在于学习方式更像人类。传统系统要么需要事先编程所有物理规律，要么需要记住每个物体间的所有可能关系，内存消耗巨大。而新系统像婴儿学习一样，先掌握基本物理概念如重力、碰撞等，然后学会组合运用这些规律来预测新场景，大幅降低了内存需求。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q2：这种AI物理学习技术什么时候能应用到日常生活中？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A：部分应用已经开始出现。目前主要在游戏物理引擎、机器人控制和虚拟现实中使用。未来几年内，我们可能会在更智能的家用机器人、更真实的游戏体验、更安全的自动驾驶等方面看到这项技术的普及。不过完全成熟的应用还需要进一步优化和测试。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q3：这个AI系统能处理所有类型的物理现象吗？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A：目前还不能。研究团队坦率承认，现有系统主要擅长处理刚体物理，如球体碰撞、物体掉落等。对于流体动力学、热传导、量子物理等更复杂现象仍有局限。不过研究团队正在开发更通用的框架，目标是让AI能够自动识别不同物理现象类型并采用相应处理策略。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】科技行者 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002F%E6%96%AF%E5%9D%A6%E7%A6%8F%E5%A4%A7%E5%AD%A6%E5%9B%A2%E9%98%9F%E6%8F%AD%E7%A7%98-ai%E5%A6%82%E4%BD%95%E5%83%8F%E4%BA%BA%E7%B1%BB%E4%B8%80%E6%A0%B7%E7%9C%8B%E6%87%82%E7%89%A9%E7%90%86%E4%B8%96%E7%95%8C%E7%9A%84%E8%BF%90%E5%8A%A8%E8%A7%84%E5%BE%8B\u002Far-AA1LBUtw?ocid=BingNewsSerp\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FxbZSN\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。\u003C\u002Fspan>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002F420f12683c8143a1a5385ff06578dd9f\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fthumbs\u002F420f12683c8143a1a5385ff06578dd9f\u002FAI领域.jpg",0,1,47,"2025-09-01 16:38",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A7b763cfe-833e-4abc-bb02-4fea49b89fc9%3A0.wav?Expires=1756721466&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=ePW7XafvaN0uj13gES4SBgQxCj0%3D",23845988,"7b763cfe-833e-4abc-bb02-4fea49b89fc9","2025-09-01 16:17","Stanford University Team Reveals: How AI Understands the Laws of Motion in the Physical World","\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fb971d07f818e4756bf4046a6970a4566\u002FAA1LCdYP.png\" width=\"665\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\">When we see a ball rolling off a table or observe leaves swaying in the wind, our brain can naturally understand the underlying laws of these physical phenomena. However, enabling artificial intelligence to possess this \"physical intuition\" has always been a major challenge. Recently, the research team at Stanford University made an important breakthrough in this issue.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The work by the research team led by Professor Jiajun Wu from the Department of Computer Science at Stanford University was published at the International Conference on Machine Learning (ICML 2024) in 2024. The first author of the study is graduate student Michael Chang, and other collaborators include researchers such as Tomer Ullman. The paper titled \"Learning Physical Dynamics with Subquadratic Memory\" details their newly developed AI system, which can understand the basic laws of object motion in the physical world just like human infants. Readers interested in further understanding can find the complete content through the official link of the paper or in the ICML 2024 conference proceedings.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The core problem addressed by this research is closely related to our daily lives. When you see an apple falling from a tree, you don't need complex calculations to predict where it will land. This ability comes so naturally to humans that we rarely realize how complex the physical reasoning involved is. However, for artificial intelligence, this seemingly simple task is extremely difficult. Traditional AI systems either require all physical laws to be pre-programmed or need massive computational resources to simulate every possible physical interaction.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team found that existing AI systems face a fundamental memory bottleneck when dealing with physical dynamics. Like a person trying to remember the relationships between every object in a room, the required memory capacity increases exponentially as the number of objects increases. This phenomenon in computer science is known as the \"quadratic memory complexity\" problem, meaning if there are 10 objects in a scene, the system needs to track 100 possible interaction relationships, and if the number of objects increases to 100, the interaction relationships surge to 10,000.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">The innovation of the research team lies in developing a new learning architecture, which is like giving AI a pair of \"physical glasses,\" allowing it to observe and understand the physical world more intelligently.\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The working principle of this system is similar to how human infants learn physical concepts. Infants do not initially try to understand all complex relationships between objects but first learn to identify basic physical concepts, such as \"objects are affected by gravity\" and \"solid objects cannot pass through each other.\" Then, they gradually learn how to combine these basic rules to predict more complex physical phenomena.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">The AI system designed by the research team adopts a hierarchical learning strategy.\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The system first learns to identify the objects in the scene and their basic properties, just like learning to distinguish different objects such as balls, boxes, and ropes. Then, it learns the basic physical laws these objects follow, such as collisions, gravity, and friction. Most importantly, the system learned how to combine these basic laws instead of trying to remember every possible combination of objects.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The cleverness of this approach is that it significantly reduces memory requirements. Returning to the earlier analogy, instead of remembering all possible relationships between every object in a room, the system only needs to master a few basic interaction rules and apply them dynamically when needed. This is like learning the rules of addition and multiplication, allowing you to calculate any combination of numbers without memorizing all possible results.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">To verify the effectiveness of this system, the research team designed a series of complex physical scenario tests. These test scenarios included multiple balls rolling in complex terrains, multiple objects colliding simultaneously, and complex mechanical systems involving springs, ropes, and other connecting components. In traditional tests, when the complexity of the scene increased, existing AI systems either crashed due to insufficient memory or saw a sharp decline in prediction accuracy.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">However, the newly developed system demonstrated impressive stability and accuracy. In a complex scenario with 50 interacting objects, traditional systems required several gigabytes of memory, while the new system only needed a few hundred megabytes to achieve the same or even better prediction accuracy. More importantly, when the research team increased the number of objects to 100, the new system still maintained stable performance, while traditional systems were completely unable to handle this level of complexity.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This improvement is not only reflected in numbers but also in the way the system understands the physical world. Through careful analysis, the research team found that the new system truly learned the real physical laws, rather than simply memorizing the results of specific scenarios. When shown never-before-seen combinations of objects, the system could correctly apply the learned physical laws to make predictions. This is like a person who truly understands the law of gravity, who can predict the trajectory of an object they have never seen before.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">The research team also discovered a particularly interesting phenomenon: the system spontaneously discovered some fundamental physical concepts during the learning process. By analyzing the internal representation of the system, researchers found that the system autonomously formed internal representations similar to physical concepts such as \"momentum\" and \"energy conservation,\" although these concepts were never explicitly taught to the system. This phenomenon is similar to how children naturally discover physical laws through play, such as gradually understanding momentum and collision concepts through playing with balls.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The system's learning ability is also reflected in its adaptability to new environments. The research team tested the system's performance in different physical environments, such as changing gravity strength, friction coefficients, or object materials. Surprisingly, even in these changed environments, the system could quickly adapt and maintain high prediction accuracy. This adaptability stems from the system learning abstract physical laws, not specific performances in particular environments.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In practical application tests, the research team demonstrated the wide potential of this technology. In robot control tasks, robots equipped with the new system were able to more accurately predict the movement trajectories of objects, thus performing grasping and manipulation tasks more precisely. In game physics engine applications, the new system could achieve more realistic physics simulation effects with lower computational costs. In virtual reality applications, the system's efficiency made real-time physical interactions more smooth and natural.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team also paid special attention to the system's interpretability. Unlike many \"black box\" AI systems, the newly developed system can explain its prediction process in a way that humans can understand. When the system predicts that a ball will roll in a specific direction, it can point out which physical laws and environmental factors the prediction is based on. This interpretability is especially important for safety-critical applications, such as autonomous vehicles needing accurate predictions of the movement trajectories of other vehicles and pedestrians.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">From a technical implementation perspective, this system uses an innovative neural network architecture that combines graph neural networks with physical perception attention mechanisms.\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This architecture's design was inspired by how the human brain processes visual and spatial information. Just as the human visual cortex automatically identifies object boundaries and properties and then passes this information to brain regions responsible for spatial reasoning, the new system also adopts a hierarchical processing approach.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">The training process of the system also reflects the ingenuity of the research team.\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Unlike traditional supervised learning methods, the new system employs a training strategy that combines self-supervised learning and physical constraints. While observing physical scenes, the system not only learns to predict future positions of objects but also learns to maintain consistency with physical laws. For example, the system automatically checks whether the predicted results violate the law of energy conservation, and if inconsistencies are found, it adjusts its internal parameters.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The advantage of this training method is that it can utilize a large amount of unannotated physical scene data. The research team found that even without manually annotated \"correct answers,\" the system could learn correct physical concepts by observing the consistency of physical laws. This greatly reduces the cost of obtaining training data and allows the system to learn from a broader range of physical phenomena.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">When evaluating the system's performance, the research team not only focused on prediction accuracy but also specifically considered computational efficiency and scalability. They found that the new system's computational time growth rate when handling large-scale physical scenes is far lower than traditional methods. Specifically, when the number of objects in the scene doubles, the computational time of traditional systems usually increases fourfold, while the computational time of the new system increases by about 1.5 times. This efficiency improvement makes real-time physical simulation much more practical.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team also explored the system's performance in different levels of physical complexity scenarios. They designed a series of test scenarios ranging from simple single-object motion to complex multi-object interaction systems. The results showed that the system's performance improvements were more pronounced in complex scenarios. This aligns with the research team's expectations, as traditional methods face a sharp increase in memory and computational burden in complex scenarios, while the advantages of the new method lie precisely in efficiently handling complex interaction relationships.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">It is particularly worth mentioning that the research team also tested the system's generalization capability when dealing with new types of objects and phenomena. They trained the system to learn rigid-body physics and then tested its performance in handling fluids and flexible objects. Although the system's performance in these new scenarios was not as good as in specially trained scenarios, it still showed some transfer learning capabilities, indicating that the system indeed learned some general physical concepts.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cspan class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">\u003Cspan class=\"ql-cursor\">\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">From a broader scientific perspective, this research provides new insights into understanding the formation mechanism of human physical intuition. By observing how AI systems learn physical concepts, researchers can better understand the similar strategies that the human brain may use. This interdisciplinary insight holds significant value for cognitive science and neuroscience research.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the paper, the research team thoroughly discussed the current method's limitations and future improvement directions. They candidly pointed out that the current system mainly focuses on rigid-body physics, and for more complex physical phenomena such as fluid dynamics and heat conduction, there are still limitations. Additionally, the system faces challenges when handling multi-scale physical phenomena (such as interactions involving both molecular and macroscopic scales).\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">However, the research team is optimistic about these challenges and has proposed specific solutions. They are developing a more general physical learning framework that can automatically identify different types of physical phenomena and adopt corresponding processing strategies. They are also exploring how to integrate more advanced physical theories such as quantum physics and relativity into the AI system.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The social impact of this research is also worth noting. More efficient physical simulation technology will directly drive the development of engineering design, scientific research, and the entertainment industry. In the engineering field, architects and mechanical engineers can more quickly test the physical feasibility of design plans. In scientific research, researchers can simulate more complex physical systems, accelerating the process of scientific discovery. In the entertainment industry, game and movie production can present more realistic physical effects while reducing production costs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">From an educational perspective, this technology also has great application potential. Physics teaching software based on this AI system can provide students with a more intuitive and personalized experience in learning physical concepts. Students can explore various physical phenomena through interaction with a virtual physical environment, and the AI system can provide real-time accurate physical feedback and explanations.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Safety considerations are also a key focus for the research team. In safety-critical applications such as autonomous driving and robotic surgery, the accuracy of physical predictions directly relates to personal safety. The research team developed a specialized uncertainty assessment mechanism, enabling the system to identify situations where its predictions might be inaccurate and take conservative strategies or seek human intervention in those cases.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Ultimately, this research represents an important step forward in artificial intelligence truly understanding the physical world. Previous AI systems primarily relied on memory and pattern matching to handle physical problems, while this new system begins to show similar physical intuition and reasoning abilities to humans. Although we are still a long way from creating an AI that fully understands the physical world, this research points us in the right direction.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">For ordinary people, the results of this technology may quietly change our life experiences in the near future. From smarter home robots to more realistic virtual gaming worlds, from safer autonomous vehicles to more efficient industrial design processes, this AI that can truly \"understand\" the physical world will bring us unprecedented convenience and possibilities.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This research also reminds us that the development of artificial intelligence is not just about technological progress but also a deep exploration of the essence of human intelligence. By making machines learn to understand the physical world like humans, we not only push the boundaries of technology but also deepen our understanding of our own cognitive abilities. The work shared by the research team at ICML 2024 is undoubtedly inspiring more researchers to get involved in this challenging and promising field.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q&A\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q1: What is the difference between the new AI system developed by Stanford University and traditional physics simulations?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A: The biggest difference of the new system lies in its learning method being more like humans. Traditional systems either require all physical laws to be pre-programmed or need to remember all possible relationships between every object, resulting in huge memory consumption. The new system, like infants learning, first masters basic physical concepts such as gravity and collisions, then learns to combine and apply these laws to predict new scenarios, significantly reducing memory requirements.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q2: When can this AI physics learning technology be applied in daily life?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A: Some applications have already begun to appear. Currently, it is mainly used in game physics engines, robot control, and virtual reality. Within the next few years, we may see this technology become popular in more intelligent home robots, more realistic game experiences, and safer autonomous vehicles. However, fully mature applications still require further optimization and testing.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q3: Can this AI system handle all types of physical phenomena?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A: Not yet. The research team candidly admits that the current system mainly excels in handling rigid-body physics, such as ball collisions and object drops. There are still limitations regarding more complex phenomena such as fluid dynamics, heat conduction, and quantum physics. However, the research team is developing a more general framework, aiming to enable AI to automatically identify different types of physical phenomena and adopt corresponding processing strategies.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">【News Source】 Tech Player \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002F%E6%96%AF%E5%9D%A6%E7%A6%8F%E5%A4%A7%E5%AD%A6%E5%9B%A2%E9%98%9F%E6%8F%AB%E7%A7%98-ai%E5%A6%82%E4%BD%9C%E5%83%8F%E4%BA%BA%E7%B1%BB%E4%B8%80%E6%A0%B7%E7%9C%8B%E6%87%82%E7%89%A9%E7%90%86%E4%B8%96%E7%95%8C%E7%9A%84%E8%BF%90%E5%8A%A8%E8%A7%84%E5%BE%8B\u002Far-AA1LBUtw?ocid=BingNewsSerp\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FxbZSN\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is reprinted by the website to provide readers with more information and news. The content does not constitute investment or consumption advice. If there are any questions about the facts of the article, please verify with the relevant parties. The views expressed in the article are not the views of the website, and are for reference only.\u003C\u002Fspan>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A28f8245d-8d9c-428d-ae66-7e1a1c194a21%3A0.wav?Expires=1774838480&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=MFebXlsKEt0TYgCTIuj%2B%2BKHcBU0%3D","28f8245d-8d9c-428d-ae66-7e1a1c194a21",16653044]