導(dǎo)讀：2025世界機(jī)器人大會于8月8日開幕，開幕式上發(fā)布了《2025具身智能機(jī)器人十大發(fā)展趨勢》，以下為全文。

趨勢一

第一，物理實(shí)踐、物理模擬器與世界模型協(xié)同驅(qū)動的具身感認(rèn)知。物理實(shí)踐是具身智能的本質(zhì)，物理模擬器可以構(gòu)建高保真的訓(xùn)練環(huán)境，世界模型可以提供環(huán)境當(dāng)中比較本質(zhì)的內(nèi)部特征。三者融合既可以保證豐富、有效、真實(shí)的環(huán)境，也可以用于訓(xùn)練具身智能機(jī)器人與環(huán)境的接觸和非接觸交互的感認(rèn)知能力，為其決策和控制奠定基礎(chǔ)。

趨勢二

第二，多層次端到端的具身決策。由多模態(tài)大模型啟發(fā)的，具有數(shù)理基礎(chǔ)的認(rèn)知與規(guī)劃研究，與生命科學(xué)家的成果融合，并與實(shí)時的控制模塊融合，可以顯著增強(qiáng)具身智能機(jī)器人在非結(jié)構(gòu)化環(huán)境下的泛化性和實(shí)用性。

趨勢三

第三，從控制角度來看，可以融合模型預(yù)測、強(qiáng)化學(xué)習(xí)和生命科學(xué)的具身智能控制。一方面可以把模型預(yù)測控制的動態(tài)優(yōu)化能力，把強(qiáng)化學(xué)習(xí)自適應(yīng)決策融合起來，更進(jìn)一步的與生命科學(xué)的冗余多環(huán)路控制機(jī)制相融合。這樣的話，可以更加讓具身智能機(jī)器人向人發(fā)展，實(shí)現(xiàn)具身智能的新控制，提升其在新環(huán)境當(dāng)中的適應(yīng)性和高性能。

趨勢四

第四，生成式人工智能驅(qū)動的具身智能機(jī)器人設(shè)計。通過對于電機(jī)、減速器、驅(qū)動器、結(jié)構(gòu)、連接件和材料的統(tǒng)一優(yōu)化，同時與工材領(lǐng)域的科學(xué)成果相結(jié)合，在物理模擬器當(dāng)中實(shí)現(xiàn)硬件與控制策略的協(xié)同優(yōu)化，可自動探索任務(wù)中實(shí)現(xiàn)最優(yōu)的具身智能的機(jī)器人設(shè)計。

趨勢五

第五，高度協(xié)同與動態(tài)適配的具身智能軟硬件一致性。具身智能機(jī)器人需要軟硬件的一致性，在硬件開發(fā)的階段需預(yù)置適配算法的接口規(guī)范，在算法的設(shè)計當(dāng)中又會內(nèi)嵌物理約束，就是軟中有硬，硬中有軟，并且通過聯(lián)合仿真驗證，就是有軟有硬的情況下，讓系統(tǒng)更加保持一致，讓軟件模塊更加接近硬件，讓整體系統(tǒng)更加符合我們的軟硬件一致性的期望。

趨勢六

第六，具身智能機(jī)器人大工廠，在仿真環(huán)境當(dāng)中實(shí)現(xiàn)自然語言交互、環(huán)境生成、機(jī)器人本體設(shè)計、決策-控制算法以及軟硬件一致性算法等研發(fā)，讓他們有機(jī)的結(jié)合在一起，并且反復(fù)進(jìn)化。這樣的系統(tǒng)可以根據(jù)性能和需求實(shí)現(xiàn)快速設(shè)計和實(shí)現(xiàn)高質(zhì)量具身智能機(jī)器人系統(tǒng)，為社會服務(wù)。

趨勢七

第七，具身智能大規(guī)模高質(zhì)量數(shù)據(jù)集，基于物理實(shí)體采集與仿真合成構(gòu)建大規(guī)模高質(zhì)量具身智能數(shù)據(jù)集。這里高質(zhì)量是一個關(guān)鍵，關(guān)于大規(guī)模，科研的期望是讓它規(guī)模要變小。同時可以顯著提升具身智能機(jī)器人的本體構(gòu)型優(yōu)化、多模態(tài)訓(xùn)練效率及跨場景策略遷移能力。

趨勢八

第八，具身智能機(jī)器人集群及與人協(xié)同的發(fā)展，融合多智能體的協(xié)同機(jī)制，構(gòu)建具身智能機(jī)器人集群。同時不斷提升具身智能機(jī)器人的安全性，及其與人的共情能力，讓具身智能機(jī)器人真正走向我們，真正走向人類，成為人類的朋友。

趨勢九

第九，跨學(xué)科的具身智能機(jī)器人開源社區(qū)。首先具身智能機(jī)器人的發(fā)展需要信息科學(xué)、工程與材料科學(xué)、數(shù)學(xué)物理科學(xué)、生命科學(xué)等多學(xué)科協(xié)作，將在全球范圍內(nèi)聚集各領(lǐng)域的頂級科學(xué)家和工程人員，促進(jìn)具身智能領(lǐng)域的技術(shù)探討，助力產(chǎn)業(yè)鏈的上下游深度融合和協(xié)作發(fā)展。

趨勢十

10 Trends ofEmbodied Intelligent Robots

TREND 1 Embodied Cognition Driven by the Synergy of Physical Practice, Physical Simulators, and World Models

Physical practice is the essence of embodied intelligence. Physical simulators construct high-fidelity training environments, while world models provide the internal characteristics of the environment. The integration of these three elements enables the creation of rich, effective and realistic environments for training embodied intelligent robots in perceptual and cognitive abilities regarding both contact and non-contact interactions with the environment, laying a solid foundation for decision-making and control.

Multi-level end-to-end embodied decision-making: Inspired by multimodal large models, the research on cognition and planning with mathematical and physical foundations, integrated with life sciences and combined with real-time control modules, significantly enhances the task generalization ability of embodied intelligent robots in unstructured environments.

By integrating the dynamic optimization capabilities of model predictive control, adaptive strategies of reinforcement learning, and the redundant multi-loop control mechanisms from life sciences, an embodied intelligent robot control system is constructed to improve the generalization and adaptability of embodied intelligent robot control in dynamic environments.

Generative AI-driven intelligent robot design realizes unified optimization of motors, reducers, drivers, structures, connectors, and materials. In combination with research and development technologies of robotic structure, it achieves co-optimization of hardware and control strategies in physical simulators, enabling automatic exploration of task-optimal embodied intelligent robot designs.

Embodied intelligent robots require consistent software/hardware co-design. By predefining interface specifications of algorithms during hardware development and embedding physical constraints of hardware into algorithm design, system-level consistency and optimization are achieved through joint simulation and validation.

In the simulation environment, research and development efforts such as natural language interaction, environment generation, robot body design, decision-control algorithms, and software-hardware consistency algorithms are integrated into an organic whole, enabling rapid design and manufacturing of high-quality embodied intelligent robot systems according to specific performance and requirements.

TREND 7 Large-Scale and High-Quality Dataset for Embodied Intelligence

The construction of large-scale and high-quality datasets for embodied intelligence through physical entity collection and simulation synthesis can enhance embodied intelligent robots’ capabilities in morphological optimization, multimodal training efficiency, and cross-scenario policy transfer.

Integrating multi-agent coordination mechanisms to construct embodied intelligent robot swarms, while continuously improving the safety of embodied intelligent agents and their ability to empathize with humans, will facilitate such robots’ entry into human society and realize their integration with humans.

TREND 9 Interdisciplinary Open Community for Embodied Intelligent Robots

The development of embodied intelligent robots requires collaboration across multiple disciplines, such as information sciences, engineering and materials sciences, mathematical and physical sciences, and life sciences. This initiative will bring together experts and scholars worldwide to promote technical discussions in the field of embodied intelligence, and facilitate in-depth integration and collaborative development throughout the industrial chain.

Through behavior norm verification, decision interpretability analysis, and data security research, a safety assessment framework and ethical norms for embodied intelligent robots will be established to ensure decision reliability and behavioral safety in complex and open environments.