CloudDeepTech Releases Five Key Achievements of Embodied Brain, Building a Complete Technical System

The development of embodied intelligence has entered a new stage of large-scale implementation in real-world scenarios. The real world is full of uncertainty, and robots must not only perceive, predict, and dynamically calibrate in real time, but also accumulate long-term memory to achieve continuous autonomous iteration, ultimately interacting and collaborating efficiently and naturally with humans. Systematic capabilities cannot be assembled from scattered demos; their core support is a truly refined embodied brain tailored for real-world operational scenarios.

Recently, Yunchudeshi Technology has released multiple papers around its self-developed embodied brain DEEPRobotics-Brain, concentrating on the publication of five research achievements: the world model DR-WM-1.0, the embodied foundational model DR-VLA-2.0, the embodied intelligent agent DR-Agent, the strategy optimization framework DR-APO, and the data closed-loop platform DR-Benchmark. These five achievements cover the full technical chain of the embodied brain, outlining a clear architecture for Yunchudeshi's embodied brain.

These five research achievements are not single-point breakthroughs, but rather collaborative efforts around the embodied brain system. Yunchudeshi Technology, in collaboration with seven internationally top-tier universities including MIT, UC San Diego, University of Illinois, Tsinghua University, Zhejiang University, Hong Kong University of Science and Technology, and Beijing Institute of Technology, has focused on innovation exploration in five areas: data closed-loop platforms, embodied intelligent agents, embodied foundational models, strategy optimization algorithms, and spatial intelligence. The paper results are interconnected, collectively forming a map of embodied intelligence for real-world operations.

As the technical pieces gradually fall into place, Yunchudeshi Technology's embodied brain DEEPRobotics-Brain is becoming increasingly clear—it is an architecture for embodied intelligence tailored for real-world operations. DEEPRobotics-Brain aims to answer: How can embodied intelligence take root in the real physical world, understand human intentions, adapt to human needs, and autonomously iterate and evolve through continuous interaction, making robots become trustworthy and reliable intelligent partners.

01.

World Model DR-WM-1.0: Introducing physical constraints to foresee the real future

The world model is an important 'training ground' for embodied intelligence to enter the real world. Traditional robot training relies on repeated trial and error in real environments, which is time-consuming and costly; while emerging video generative world models, although capable of simulating rich visual scenes, generally lack accurate modeling of real physical laws and are unable to reliably predict the actual consequences of robot actions. Based on video generative world models, it is possible to generate refined videos of a robot walking through snow, yet they overlook the low friction that leads to slipping.

Visual plausibility does not equate to dynamic feasibility, which is precisely the most dangerous gap in the current use of generative world models for robotic decision-making: the aesthetic imagination in the pixel space is difficult to translate into the physical imagination required for embodied intelligence, and the DR-WM-1.0 released by Cloud Depth Technology directly addresses this gap.

Cloud Depth Technology, in collaboration with Zhejiang University, Beijing Institute of Technology, Tsinghua University, MIT, and the University of Illinois, released DR-WM-1.0, which for the first time introduced physical properties such as friction, roughness, and stiffness in world models for quadruped robots. On the basis of traditional video-based generative world models, it added a prior module for physical attributes that enables large-scale training, allowing the world model to predict future scenes that are not only realistic but also conform to physical laws. The model was pre-trained on 2.58 million images and over 100,000 real robot segments, ensuring generalization while: the dual ControlNet architecture introduces scene geometry and camera motion into video generation, and autoregressive inference continues the machine dog and controller states, keeping the simulation process aligned with real control loops.

Fig. 1 DR-WM overall process and motion performance under different contact conditions.

This is equivalent to providing robots with a 'physical rehearsal' before taking action, helping it anticipate: what will happen if it takes this path, and how to proceed more safely. DR-WM connects environmental perception, physical prediction, and action decision-making, offering an analyzable and verifiable technical path to narrow the gap between visual realism and physical realism. For example, in the case of inspecting a substation after rain and snow: facing a slippery shortcut and a detour route, the robot can first separately 'rehearse' to predict the possibility of slipping, instability in posture, and obstruction in movement, then make a decision based on safety risks and travel efficiency.

02.

Embodied Foundation Model DR-VLA-2.0: Understanding Human Social Rules to Let Robots Truly 'Listen'

Preparation capabilities are in place, allowing robots to move toward real roads and workstations—

But the next challenge is imminent: How can robots better integrate into human life and understand human social behavior?

To truly integrate into human society, robots must go to frontline scenarios such as streets, parks, and logistics delivery. Navigation cannot only solve path planning; it must also possess the comprehensive ability to understand instructions, avoid obstacles, and respond in real-time to dynamic environments.

However, robot teaching relies on remote operation, which has high collection costs and is difficult to cover rich environmental and interactive contexts. Although robots can learn navigation experiences such as obstacle avoidance, yielding, and path selection by leveraging massive first-person perspective videos from humans, the camera perspectives, imaging characteristics, and movement modes of humans and robots differ, and the videos lack action labels that robots can directly execute. How to extract transferable navigation knowledge from ordinary online videos and convert it into trajectories that robots can execute is the core issue focused on in the navigation field.

Yunduchu Technology, in collaboration with researchers from MIT, University of Illinois at Urbana-Champaign, Tsinghua University, and Zhejiang University, is exploring a four-legged robot navigation learning path titled “Human Video Pretraining without Action Labels—Robot Teaching Adaptation.” The team has constructed the EgoLynx dataset, integrating 10,000 hours of first-person network videos and 2,000 segments of robot teaching footage. By innovatively utilizing the DISR module, the robot is able to learn behavioral patterns from massive video data, understand traffic rules and navigation rules, comprehend human intentions, and learn social etiquette, thereby enhancing the human-like level of embodied intelligence.

Figure 2 Overview of the DR-VLA method. On the left is a human-first-person video, in the middle are data from three experts and a robot, and on the right are future scene representations and predicted trajectories. The human video provides understanding and prediction supervision, while robot demonstrations further provide action supervision.

Technically, this is the first quadruped robot navigation paradigm based on domain-invariant scene representation (DISR) in a cross-embodiment learning scenario. The human and robot perspectives belong to different domains, and the key to cross-embodiment learning is to find a common scene representation between the two domains—DISR integrates scene semantics, spatial localization, and spatiotemporal relationships, allowing the model to focus on 'what is there, where it is, and how they are related,' establishing a cross-embodiment transfer interface through scene structure; the Mixture-of-Transformers architecture enables understanding, prediction, and action to work in synergy, while also considering deployment efficiency.

10,000 hours of data support, with quantifiable gains: In the EgoLynx-Robot test set evaluation, the comprehensive navigation score QNS improved by approximately 5%, and the final heading error decreased by 14.6%. With the help of this research, robots are expected to combine route instructions to predict pedestrian and obstacle changes in scenarios such as 'last-mile' delivery in communities and campuses, more naturally integrating into everyday spaces with mixed human and vehicle traffic.

Beyond mobility, precise manipulation is the real challenge for robot deployment. In terms of operation, the inherent limitations of current embodied models lie in weak language supervision, i.e., 'not understanding the relationship between language and actions'; in long-range complex operations, models are often dominated by large-scale visual supervision, and even when people change language instructions, the robotic arm will either execute the same action or hesitate, struggling to understand how language actually commands movement; in tasks with rich contact such as plugging and wiping, language, vision, and force feedback actually belong to different pathways, lacking a unified, interpretable, and continuously adjustable movement expression, i.e., insufficient manipulation steerability.

Yunduchu Technology proposes

self-developed atomic motion coordinate large model

For the first time, it integrates language control with force feedback within the same motion coordinate framework, decomposing the mechanical arm's movement into 13 atomic coordinates (atomic coordinate), establishing a 'unified coordinate language' for motion. Through the first phase of training that shields vision, it establishes a true alignment mapping between language and atomic actions. In subsequent phases, force feedback is introduced to refine the actions, enabling the robot to 'hear, see, feel, and act.' The control effect of language on motion direction can be directly tested and quantified.

Figure 3 When the image and state are fixed, and only the positive and negative atomic prompts are changed, the end-point translation and rotation trajectories generate responses in opposite directions.

In the experiment, the separation rate for opposite instructions of single-atom and double-atom actions reached 92.5% and 83.1%, respectively; in unseen layouts of real dual-arm tasks, the task progress was increased to 87.8%; after incorporating force feedback, the progress for contact tasks such as plug insertion and vase wiping was further increased to 78.5% and 75.2%, respectively; the success rate on language generalization and force response tasks outside the data distribution (OOD) surpassed existing baselines such as PI05, ForceVLA, and LA4VLA, reaching the current industry-leading level. This capability can be extended to power cabinet operations, connector insertion, and part alignment in power scenarios, possessing significant practical value.

03.

Strategy Optimization Framework DR-APO: Success rate increased to 89.55%

Navigation and operation models have been successively deployed, but to continuously enhance model capabilities, the same issue remains unavoidable: how to train? Existing reinforcement learning methods are highly bound to specific policy representations—specialized models, targeted optimization, and a single policy often only covers a single task. When the base model scale expands and action representations become increasingly diverse, the training paradigm itself becomes the bottleneck of the entire pipeline.

Cloud Depth Technology, in collaboration with Zhejiang University, proposed DR-APO—the first representation decoupled general generative policy reinforcement learning framework, completely decoupling policy improvement from the underlying policy representation, achieving unified reinforcement learning optimization from zero training of continuous control policies to post-training of embodied navigation policies.

Figure 4 DR-APO combines advantage-guided target improvement with action fitting of adapted policy representations, transitioning from continuous control validation to post-training of embodied navigation.

The core mechanism 'Advantage-Guided Target Improvement' is introduced for the first time: based solely on the local gradient direction of advantage from a single action execution and advantage reweighting, it constructs an improved action target, transforming the policy improvement into an optimization problem of an independent target distribution, completely eliminating the need for explicit policy likelihood calculations and the reliance on multiple candidate action Critics during the inference phase. 'Direct Action Fitting with Representation Compatibility' further maps the unified action space optimization objective to Gaussian, diffusion, and flow policies, overcoming the challenge of the weakening of local improvement signals in multi-step generation processes in traditional native fitting.

It is worth noting that the general training effectiveness of Yunchudeshi Technology's DR-APO even surpassed specialized training. In 5 MuJoCo-v5 continuous control benchmark tests, DR-APO achieved first place in nine out of ten metrics on diffusion strategies and first place in all ten metrics on flow strategies, achieving the best single-strategy comprehensive performance; in embodied visual navigation post-training, the pre-trained model NavDP's simulation success rate increased from 70.33% to 89.55%, an improvement of 19.22 percentage points; in Yunchudeshi Technology's Sanmao M20 wheel-legged robot real-world scenario testing, the navigation success rate increased significantly from 25% to 65%.

DR-APO provides an efficient and general technical path for cross-heterogeneous representation strategy learning and embodied large model post-training, directly supporting the post-training of navigation and operation models — a single framework, driving the unified evolution of the entire chain of models.

04.

Embodied Intelligent Agent DR-Agent: The first continuously self-evolving embodied navigation intelligent agent

Training engine in place, robots can run, operate, and self-evolve — Still missing one piece of the puzzle: thinking.

The real environment continuously changes dynamically: household items are moved around at will, hotel layouts are adjusted daily, and warehouse goods are constantly coming in and out. Traditional robots rely on fixed environmental memory; once changes exceed the prior range, 'memory expiration' leads to task failure—memory solidification and insufficient adaptation are the core challenges for embodied intelligent agents in long-term deployment.

This DR-Agent, released by Yunchudeshi Technology in collaboration with the University of California, San Diego, and Zhejiang University, is the first self-evolving 4D belief embodied navigation intelligent agent designed for continuous evolving worlds. The team proposed the EvolvingNav method to address 'persistent navigation after unseen object displacement': building a time-indexed predictive 4D belief model, along with an event-driven 'predict-observe-replan' complete filter, enabling robots to perform temporal reasoning based on arrival time and rely on vision for goal verification — it has memory and can continuously optimize its behavioral strategies based on memory, achieving dynamic iteration and continuous updating of the memory system in human-machine interaction scenarios.

Figure 5 DR-Agent Overview

The accompanying EvoWorldBench evaluation benchmark, built based on real human activity trajectories, includes 54 simulated scenarios, over 800,000 task instances, and a total of approximately 239,000 dynamic state changes of objects. Experimental results demonstrate that DR-Agent significantly outperforms mainstream baseline methods in temporal prediction, evidence updating, and fault recovery. It achieves leading comprehensive performance on evaluation benchmarks such as FindingDory and GOATBench. From long-term autonomous operation in home services, continuous operation across time periods in shared spaces, to predictive dynamic memory-assisted item search and inspection path adjustment in warehousing logistics and industrial inspections—the self-updating memory determines whether a robot can transition from 'demonstrating once' to 'long-term on duty'.

05.

Data Loop Platform DR-Benchmark: 833.6 hours, 3,294.7 kilometers of real-world operation data collection

Whether it is rehearsal, interaction, training, or memory, all of this starts and ends with the same thing: data. The industry has long faced two shortcomings: first, existing robot data and language navigation data each cover only part of the information, and there is still a lack of data that can fully associate environmental perception, behavioral intent, and other supervisory signals with real continuous trajectories; second, perception, navigation, future prediction, and scene reconstruction use different data and protocols, making it difficult to conduct consistent analysis on the same platform and in the same scenario. Without data flow, all model capability comparisons and iterations are impossible.

Cloud Depth Technology, in collaboration with the University of Illinois Urbana-Champaign, Hong Kong University of Science and Technology, Zhejiang University, Tsinghua University, and MIT, released the DR-Benchmark data loop platform. Its core dataset, LynxScenes, is the first large-scale dataset targeting real-world outdoor scenarios for leg-wheel robots, addressing the lack of data in real embodied navigation scenarios and filling the gap in evaluation benchmarks.

Figure 6 LynxScenes Overall Framework, covering real robot data collection, multimodal observations, spatial and behavioral annotation, as well as four downstream tasks.

Based on Cloud Depth Technology's Mountain Cat M20 wheeled-legged robot, the team has completed a total of 833.6 hours and 3,294.7 kilometers of real-world operation data collection, constructing approximately 100,000 navigation segments, 3,000,000 annotated frames, providing supervision for four categories: 3D objects, road structures, behavioral commands, and expert demonstration trajectories, covering 27 types of objects, 4 types of road structures, 3 levels, and 20 types of navigation commands. LynxScenes is the first to unify the tasks of instruction understanding, environmental perception, future prediction, and scene reconstruction under a single real-world data benchmark, and accordingly conducted a systematic 'check-up' on 25 representative models, redefining the evaluation dimensions of embodied navigation.

From campus, community, and park delivery, robots make real-time decisions on passage, waiting, and detours, to factory inspection scenarios where path adjustments are made for narrow passages and temporary obstacles, and then to R&D teams using real-recorded playback to identify issues and model comparisons, DR-Benchmark provides benchmarks and foundational models for various industries, allowing real-world data to continuously flow back, becoming the source of nourishment for the embodied brain's continuous evolution.

Piece by piece, the technical puzzle is gradually in place and tightly interlocked, forming the technical skeleton of the embodied brain (DEEPRobotics-Brain) in the cloud. In this round of intensive research layout, DeepRobotics Technology continues to tackle core challenges along the entire chain, breaking away from industry-specific, fragmented development habits, and directly facing the challenges of embodied intelligence in the real physical world. It explores a cutting-edge technical path that is verifiable, iterable, and transferable. From physical prediction to dynamic memory, from strategy optimization to data benchmarks, each technological breakthrough is interconnected and mutually supportive. This systematic exploration lays a solid foundation for the development and deployment of multi-form robots; by consolidating industry consensus through a unified evaluation benchmark, it drives the field away from scattered trial-and-error approaches and toward systematic research, laying a front-line technological foundation for empowering the development of embodied intelligence and achieving long-term visions of human-robot collaboration.

A series of preprint papers have now been fully released on the [DeepRobotics Technology official website] - [Developer Center] - [Frontier Exploration] section. Welcome to exchange and discussion https://developer.deeprobotics.cn/research/

END

Industrial robot companies

Eston Automation

Efte Robot

Fao Robot

Yuexiang Robot

Jieka Robot

Songling Robot

Luoshi Robotics

A Tongmu Robot

JiZhiJia

Hikvision Robot

Yifei Technology

Ailite Robot

Service and special robot companies

Yijiahe

Jingpin Tieshang

Qitem Robotics

Shihe Robot

Pudu Robot

Shiluo Robot

Kuma Technology MAMMOTION

Humanoid robot companies

UBTech Robotics

Unitree

Yunchedu

Xingdongjinyuan

Weijing Robot

Zhujidongli

Leju Robot

Daxiang Robot

Magic Atom

Zhongqing Robot

Pasi Ni Perception

Cyborg Robot

Digital Huaxia

Fourier Intelligent

Tianlian Robot

Kepler Humanoid Robot

Lingbao CASBOT

Qingbao Robot

Zhejiang Humanoid Robot Innovation Center

Dongyi Technology

Zhishi Technology

PNDbotics

Zhuoyide Robot

Qinglang Intelligent

Galileo GALILEO

Songyan Power

Tianji Intelligent

Woaan Robot

LiGong Huahui

Jiace Jinhua

Embodied intelligence companies

Cross Dimension Intelligent

Galaxy Universal

Qianxun Intelligent

Lingxin Qiaoshou

Ruierman Intelligent

Microintelligence

Tuisheng Technology

Zhongke Silicium Records

Shutu Technology

Lingqiao Intelligent

Xingchen Intelligent

Qiongche Intelligent

Fangzhou Infinite

| iFLYTEK |

Beijing Humanoid Robot Innovation Center

National and Local Co-built Humanoid Robot Innovation Center

Daimeng Robot

Shibitu Robot

Xinghaitu

Yuequan Bionics

Zeroth Power Robot

Zhongke ShenGu

Zhi Fangping

Daka Robot

Haocun Technology

Ju Shi Intelligent

Xynova Xiner Future

Feixi Technology

Weilu Power

Boden Intelligent

Qianjue Technology

Lingsheng Technology

Ji Cui Intelligent Manufacturing

Xinbeit Technology

Chen Hunxian Technology

Dexmal Original Power

Youliqi

Self-variable

Ruiyan Zhikong Lingqiao Hand

Qiwu Technology

RoboScience Machine Science

Zhongke Fifth Dynasty

Critical Point

Danghong Technology

Qiaojie Shu Wu

Vbot Veta Power

Tashan Technology

Ju Brain Pan Shi

Youai Zhijhe Robot

Zhixing Jian

Aimiao Robot

Lu Ming Robot

Quantum Intelligent

Ruoyu Technology

Tiehu Robot

Zhi Wuji

Medical robot companies

Yuanhua Intelligent

Tianzhihang

Sizhe Rui Intelligent Medical

Jingfeng Medical

Tuo Dao Medical

Zhenyida

Shuru® Robot

Luosenbot

Shuimu Dongfang

｜

Kangnuo Sistem

Disight Medical

Upstream industry chain companies

Lvde Harmonic

Yinshi Robotics

Kunwei Technology

Muta Intelligent

Qingpou Vision

Benmo Technology

Lan Dian Touch Control

| Xinjingcheng Sensor

BrainCo Strong Brain Technology

Yuli Instrument

Jiya Jingji

Sila Technology

Shenyuan Sheng

Feipu Navigation Technology

Inks

Juxi Intelligent Drive

Lingyun Guang Yuanke Vision

Xuanji Power

Yiyou Technology

| Ruiyuan Precision |

Lingzhu Era

HIT Huawei Ke

Xinghui Sensor

Lingdi Technology

Quan Zhi Bo

CubeMars Robot Power

Wanglong Robot Elevator

Hangkai Microelectronics

Wutong Sensing and Control