DEEP Robotics releases five research results on embodied brain

Recently, Deep Robotics has released a dense series of papers centered on its self-developed embodied brain DEEPRobotics-Brain, announcing five research results in a concentrated manner: the world model DR-WM-1.0, the embodied foundation model DR-VLA-2.0, the embodied agent DR-Agent, the policy optimization framework DR-APO, and the data closed-loop platform DR-Benchmark. The five results cover the full-chain technology of the embodied brain and outline a clear architecture for Deep Robotics' embodied brain. These five studies are not isolated breakthroughs but coordinated efforts carried out around the embodied brain system. Deep Robotics has joined forces with seven top international universities—MIT, UC San Diego, the University of Illinois, Tsinghua University, Zhejiang University, the Hong Kong University of Science and Technology, and Beijing Institute of Technology—to conduct innovative exploration in five major areas: the data closed-loop platform, embodied agents, embodied foundation models, policy optimization algorithms, and spatial intelligence. The paper results are closely interlinked, together piecing together a map of embodied intelligence oriented toward real-world operational scenarios. As the pieces of the technological puzzle fall into place one by one, Deep Robotics' embodied brain DEEPRobotics-Brain is becoming increasingly clear—it is an embodied intelligence architecture oriented toward real-world operational scenarios. What DEEPRobotics-Brain seeks to answer is: how can embodied intelligence take root in the real physical world, understand people's intentions, adapt to people's needs, and autonomously iterate and evolve through continuous interaction, so that 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 move into the real world. Traditional robot training relies on repeated trial and error in real environments, which is time-consuming and costly; while emerging video-generation-based world models can simulate rich visual scenes, they generally lack accurate modeling of real physical laws and are hard-pressed to reliably predict the actual consequences of robot actions. A video-generation-based world model can generate an exquisite video of a robot crossing snow, yet overlook that extremely low friction will cause slipping. Visual plausibility does not equal dynamic feasibility—this is precisely the most dangerous gap at the current stage when generative world models are used for robot decision-making: the aesthetic imagination capability in pixel space is difficult to convert into the physical imagination capability used for embodied intelligence, and DR-WM-1.0, released by Deep Robotics, directly targets this gap. DR-WM-1.0, released by Deep Robotics in collaboration with Zhejiang University, Beijing Institute of Technology, Tsinghua University, MIT, and the University of Illinois, is the first world model in the quadruped robot field to introduce physical properties such as friction, roughness, and stiffness. On top of the traditional video-generation-based world model, it adds a physical-attribute prior module trained at large scale, so that the world model's predicted future frames are not only realistic but also conform to physical laws. The model is pretrained on 2.58 million images and over 100,000 real-machine clips; while ensuring generalization: the dual-ControlNet architecture introduces scene geometry and camera motion into video generation, and autoregressive inference continues the robot dog and controller states, keeping the simulation process consistent with the real control closed loop. This is equivalent to providing the robot with a "physical rehearsal device" before acting, helping it judge in advance: what will happen if it walks on this path, and how to walk more safely. DR-WM connects environmental perception, physical prediction, and action decision-making, providing an analyzable and verifiable technical path for narrowing the gap between visual realism and physical realism. Take substation inspection after rain or snow as an example: faced with a slippery shortcut and a detour route, the robot can first "rehearse" each separately, predict the possibility of slipping, posture instability, and blocked progress, and then make a decision by combining safety risks and passage efficiency. 02 Embodied Foundation Model DR-VLA-2.0: Understanding Human Social Rules and Truly "Understanding" Instructions With rehearsal capability in place, robots can move toward real roads and workstations—but the next test follows immediately: how can robots better integrate into human life and understand human social behavior? 01 DR-VLA For robots to truly integrate into human society and go to frontline scenarios such as streets, parks, and logistics delivery, navigation must not only solve path planning but also possess comprehensive capabilities for instruction understanding, obstacle avoidance, and real-time response to dynamic environments. However, robot teaching relies on teleoperation, which has high collection costs and makes it difficult to cover rich environments and interaction situations. Although robots can leverage massive amounts of human first-person videos to learn navigation experience such as obstacle avoidance, yielding, and path selection, humans and robots differ in camera viewpoint, imaging characteristics, and motion patterns, and the videos also lack action labels that robots can directly execute. How to extract transferable navigation knowledge from ordinary internet videos and convert it into robot-executable trajectories is the core issue focused on in the navigation direction. Researchers from MIT, the University of Illinois Urbana-Champaign, Tsinghua University, and Zhejiang University, in collaboration with Deep Robotics, explored a quadruped robot navigation learning path of "action-label-free human video pretraining—robot teaching adaptation." The EgoLynx dataset constructed by the team integrates 10,000 hours of human first-person internet videos and 2,000 robot teaching clips. With the innovative use of the DISR module, the robot can learn behavioral patterns from massive videos, understand traffic rules and navigation rules, understand human intentions, and learn social etiquette, improving the anthropomorphic degree of embodied intelligence. Technically, this is the first quadruped robot navigation paradigm based on domain-invariant scene representation (DISR). Human and robot perspectives belong to different domains. The key to cross-embodiment learning is finding a scene representation common to both domains—DISR fuses scene semantics, spatial localization, and spatiotemporal relations, enabling the model to focus on "what is there, where it is, and how things relate to one another," and establishing a cross-embodiment transfer interface through scene structure; the Mixture-of-Transformers architecture coordinates understanding, prediction, and action while also accounting for deployment efficiency. Backed by 10,000 hours of data, the gains are quantifiable: in evaluations on the EgoLynx-Robot test set, the comprehensive navigation score QNS improved by about 5%, and the final heading error decreased by 14.6%. With the support of this research, robots in scenarios such as "last-mile" delivery in communities and industrial parks are expected to combine route instructions to anticipate changes in pedestrians and obstacles, integrating more naturally into everyday spaces shared by people and vehicles. 02 Self-developed atomic motion coordinate large model Beyond walking, fine manipulation is the real challenge for robot deployment. In manipulation, an inherent flaw that current embodied models struggle to overcome is that language supervision is too weak—that is, they "cannot understand the relationship between language and action"; in reasoning for long-horizon complex manipulation, models are often dominated by large-scale visual supervision, so even if people change the language instruction, the robotic arm will only execute it as-is or hesitate, making it difficult to understand how language actually commands action; in contact-rich tasks such as plugging and wiping, language, vision, and force feedback in fact belong to different pathways, lacking a unified, interpretable, and continuously correctable motion expression—that is, manipulation steerability is insufficient. Deep Robotics proposes a self-developed atomic motion coordinate large model, the first to incorporate language control and force feedback into the same motion coordinate framework, decomposing robotic arm motion into 13 atomic coordinates and establishing a "unified coordinate language" for motion. Through first-stage training with vision masked, it builds a mapping that truly aligns language with atomic actions, and in subsequent stages introduces force feedback to correct actions, enabling the robot to "understand, see, feel, and correct," with the effect of language on motion direction directly testable and quantifiable. In experiments, the separation rates for opposite instructions on single-atomic and dual-atomic actions reached 92.5% and 83.1%, respectively; in real dual-arm tasks with unseen layouts, task progress increased to 87.8%; after incorporating force feedback, progress on contact tasks such as plug insertion and vase wiping further increased to 78.5% and 75.2%; on out-of-distribution (OOD) language generalization and force-response tasks, the success rate surpassed existing baselines such as PI05, ForceVLA, and LA4VLA, reaching the current industry-leading level. This capability can be extended to power scenarios such as electrical cabinet operation, connector insertion, and part alignment, and has important practical deployment value. 03 Policy optimization framework DR-APO: success rate raised to 89.55% With navigation and manipulation models successively in place, for model capabilities to keep improving, one question cannot be avoided: how to train them? Existing reinforcement learning methods are highly bound to specific policy representations—specialized models and targeted optimization mean one policy often covers only a single task. As foundation models scale up and action representations become increasingly diverse, the training paradigm itself becomes the bottleneck of the entire chain. Deep Robotics, together with Zhejiang University, proposes DR-APO—the first representation-decoupled general generative policy reinforcement learning framework, which thoroughly decouples policy improvement from the underlying policy representation and achieves unified reinforcement learning optimization from training continuous-control policies from scratch to post-training embodied navigation policies. The core mechanism, "advantage-guided target improvement," is proposed for the first time: it constructs improved action targets based solely on the local advantage gradient direction and advantage reweighting of a single executed action, transforming policy improvement into a target distribution optimization problem independent of policy parameterization, completely eliminating reliance on explicit policy likelihood computation and on Critic screening of multiple candidate actions during inference. "Representation-compatible direct action fitting" further maps the unified action-space optimization objective to Gaussian, diffusion, and flow policies, overcoming the difficulty in conventional native fitting that local improvement signals are weakened during multi-step generation. Notably, the general training results of Deep Robotics' DR-APO even surpass specialized model training. On five MuJoCo-v5 continuous-control benchmarks, DR-APO ranks first on nine of ten metrics with diffusion policies and first on all ten metrics with flow policies, achieving the best comprehensive performance for a single policy; in embodied visual navigation post-training, the simulation success rate of the pretrained model NavDP increased from 70.33% to 89.55%, a gain of 19.22 percentage points; in real-scenario tests on Deep Robotics' Lynx M20 wheeled-legged robot, the navigation success rate rose sharply from 25% to 65%. DR-APO provides an efficient and general technical path for policy learning across heterogeneous representations and for post-training of embodied large models—a single framework driving the unified evolution of models across the entire chain. 04 Embodied agent DR-Agent: the first continuously self-evolving embodied navigation agent With the training engine in place, robots can run, manipulate, and self-evolve—but one piece is still missing: the ability to think. Beyond execution, robots also need to judge and analyze, and that requires memory. Real environments are constantly and dynamically changing: household items are moved casually, hotel layouts are adjusted daily, and warehouse goods continuously come and go. Traditional robots rely on fixed environmental memory; once changes exceed the prior range, "memory expiration" leads to task failure—memory rigidity and insufficient adaptation are core difficulties in the long-term deployment of embodied agents. Deep Robotics, together with UC San Diego and Zhejiang University, has released DR-Agent, the first self-evolving 4D belief embodied navigation agent for a continuously evolving world. To address "persistent navigation after the displacement of unseen objects," the team proposes the EvolvingNav method: it builds a time-indexed predictive 4D belief model, paired with an event-driven "predict-observe-replan" complete filter, enabling the robot to perform temporal reasoning based on arrival time and to verify targets visually—it has memory, and can continuously optimize its behavioral policy based on that memory, achieving dynamic iteration and continuous updating of the memory system in human-robot interaction scenarios. The accompanying EvoWorldBench evaluation benchmark, built on real human activity trajectories, contains 54 simulated scenes, more than 800,000 task instances, and a total of approximately 239,000 dynamic object state changes. Experiments verify that DR-Agent significantly outperforms mainstream baseline methods in temporal prediction, evidence updating, and failure recovery, and achieves leading overall performance on benchmarks such as FindingDory and GOATBench. From long-term autonomous operation in home services and cross-time-period continuous operation in shared spaces to item search and inspection path adjustment assisted by predictive dynamic memory in warehousing, logistics, and industrial inspection—a memory that can update itself determines whether a robot can move from "demonstrated once" to "on duty long-term." 05 Data closed-loop platform DR-Benchmark: 833.6 h, 3,294.7 km of real-world operation data collected Whether it is rehearsal, interaction, training, or memory, the starting point and destination of all of this is 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 data that fully links supervision signals such as environmental perception and behavioral intent with real continuous trajectories is still lacking; 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 scene distribution. Without connected data, any comparison and iteration of model capabilities is impossible. Deep Robotics, together with the University of Illinois Urbana-Champaign, the Hong Kong University of Science and Technology, Zhejiang University, Tsinghua University, and MIT, has released the DR-Benchmark data closed-loop platform. Its core dataset, LynxScenes, is the first large-scale dataset for real outdoor scenes of wheeled-legged robots, filling the data gap for real embodied navigation scenarios and filling the gap in evaluation benchmarks. Based on the Deep Robotics Lynx M20 wheeled-legged robot, the team completed a cumulative 833.6 hours and 3,294.7 kilometers of real-world operation data collection, building approximately 100,000 navigation clips and 3 million annotated frames, and providing four types of supervision: 3D objects, road structures, behavioral instructions, and expert demonstration trajectories, covering 27 object categories, 4 road structure types, 3 levels, and 20 navigation instruction categories. LynxScenes is the first to unify the four tasks of instruction understanding, environmental perception, future prediction, and scene reconstruction into the same real-data benchmark, and on this basis completed a systematic "checkup" of 25 representative models, redefining the evaluation dimensions of embodied navigation. From robots' real-time decisions on passing, waiting, and detouring in campus, community, and park delivery, to path adjustment in narrow passages and around temporary obstacles during factory inspections, to R&D teams using real recordings to replay problems and compare models, DR-Benchmark provides benchmarks and a foundation for model development across industries, and also enables data from real scenarios to continuously flow back, becoming the wellspring for the continuous evolution of the embodied brain. Piece by piece, the technology puzzle pieces are falling into place and interlocking tightly, and the technical skeleton of the Deep Robotics embodied brain (DEEPRobotics-Brain) has taken shape. In this round of intensive research layout, Deep Robotics continues to tackle core challenges across the entire chain, breaking away from the industry's habit of fragmented, single-point R&D, directly confronting the challenges of embodied intelligence in the real physical world, and exploring a frontier technical path that is verifiable, iterable, and transferable. From physical prediction to dynamic memory, and from policy optimization to data benchmarks, each technological breakthrough is interlinked and mutually supporting. This systematic exploration lays a solid foundation for the development and deployment of multi-form robots; by building industry consensus around a unified evaluation benchmark, it pushes the field away from scattered trial and error toward systematic research, and solidifies the frontier technology foundation for the long-term vision of empowering industrial development with embodied intelligence and realizing human-machine collaboration.