Dexmal DM0.5 Explained: A 150,000-Hour Embodied Foundation Model for Real-World Robots

IntroductionEmbodied AI has no shortage of strong demos, new robot hardware, or research talent. The harder problem is still the same: how do you build a data flywheel that can keep improving robots in real-world tasks?Dexmal, known in Chinese as 原力灵机, introduced a new embodied foundation model called DM0.5 at its developer event. The model is positioned as a general-purpose foundation model for open-world embodied intelligence. It is meant to connect three things that often stay separate in robotics: large-scale robot data, model generalization, and deployment infrastructure.According to the original QbitAI report, DM0.5 is built on 150,000 hours of data, uses a 4B-parameter model, and is designed to support navigation, grasping, full-body control, long-horizon memory, and cross-embodiment deployment.## 150,000 Hours of Data: The Foundation Behind DM0.5Dexmal describes DM0.5 with three keywords: larger, stronger, and more practical.The model is positioned as a general embodied foundation model for open-world tasks. Compared with the previous DM0 model, DM0.5 doubles the parameter scale to 4B parameters and increases the training data volume by 400%.The 150,000-hour data foundation is mainly built from three types of high-quality data.###

1. Real Robot Operation DataDM0.5 uses 50,000 hours of high-precision robot operation data. This data covers more than 100 types of actions and supports fine-grained alignment between instructions and actions at the second level.This part of the dataset helps the model learn how physical actions are actually executed in the real world, not just how they are described in language.###
2. Egocentric First-Person DataThe model also uses 100,000 hours of egocentric data. This gives the model a more human-like view of the environment and supports high-precision 3D landmark generation at the millimeter level.First-person data matters because robots need to understand scenes from the perspective of an actor moving through the world, not only from static camera views.###
3. Scene Reconstruction DataDexmal also uses scene reconstruction data based on 1 million square meters of spatial data. The goal is to model complex indoor environments and reduce the Sim2Real gap between simulation and real deployment.This matters for embodied intelligence because a robot that performs well in a simulator may still fail when lighting, object placement, surface friction, or human behavior changes in a real scene.## Three Architecture Upgrades in DM0.5Beyond data scale, DM0.5 introduces three major architecture-level upgrades. These upgrades are designed to help robots move from “remembering actions” to “understanding tasks”.###
4. Context Abstraction LayerMany real-world robotic tasks are not short one-step commands. They can last several minutes and require the robot to remember what happened earlier.DM0.5 adds efficient context compression and supports up to 60 seconds of native memory, with average memory capability around 30 seconds. This allows the model to preserve useful historical context and better understand how the environment has changed during a longer task.For long-horizon tasks, this kind of memory can reduce fragmented behavior and help the robot keep a more consistent action sequence.###
5. Embodied Chain-of-Thought TasksDM0.5 still follows a VLA-style direction, but Dexmal adds more reasoning-oriented tasks, including task planning, target localization, and future-state prediction.In simple terms, the model is encouraged to plan before acting. When a command is complex, it can break the goal into smaller steps, organize the action order, and reduce trial-and-error behavior.###
6. Trajectory Alignment LayerNatural-language commands and robot joint movements live in very different spaces. A person may say one sentence, but the robot still needs to translate that instruction into millimeter-level arm, gripper, or body movement.The trajectory alignment layer helps the model learn movement as an aligned process rather than a set of isolated points. This makes the model better at understanding the pattern of motion between intention and execution.Together, these upgrades support Dexmal’s claim that DM0.5 is not only memorizing actions but learning a more task-aware representation of the physical world.## Benchmark Results and Efficiency GainsUnder this data and architecture setup, DM0.5 shows several reported performance improvements across real-robot and simulation evaluations.According to the report, DM0.5 outperforms current SOTA models across four public real-robot and simulation benchmarks. A single model can support multiple task types, including navigation, grasping, and full-body control.Compared with DM0, DM0.5 reportedly improves:- Zero-shot navigation success rate by 31%

• Few-shot success rate by 45%
• Fine-tuned task success rate by 20%
• **Inference efficiency by 25%**The reported latency is also low for robotics deployment:- 50 ms inference latency on NVIDIA H100
• 90 ms inference latency on RTX 4090These numbers matter because embodied AI cannot only be accurate. It also needs to respond fast enough for real-time physical control.## Generalization and Lower Costs: What “Useful” Means for DM0.5At the developer event, Dexmal co-founder and CEO Tang Wenbin argued that many embodied AI models are being released, but the real question is whether they are actually useful.Dexmal uses two practical standards to define a useful embodied foundation model:1. How strong is the model’s zero-shot generalization?
  This decides the upper bound of the model’s capability.

1. How efficient is post-training?
   This decides how quickly the model can move into real industrial or service scenarios.For post-training, DM0.5 reportedly lowers fine-tuning cost by 60%. A new downstream task can be fine-tuned and deployed to expert-level performance with just one RTX 4090 consumer GPU, in as fast as 18 hours.The larger change is in generalization. Dexmal emphasizes that DM0.5 is designed to understand the relationships among objects, scenes, tasks, robot embodiments, and natural-language instructions. The company says it has already observed signs of generalization emergence in DM0.5.##
2. Zero-Shot CapabilityZero-shot capability measures what a robot can do without seeing the exact task during training.Dexmal tested DM0.5 on Franka single-arm robots and Dexmal-Mint across eight core atomic actions:- Pick

• Put
• Cover
• Rotate
• Push
• Pull
• Wipe
• StackThe tests also included more complex instruction-following cases involving color, shape, size, state, absolute position, relative position, and long-sequence instructions.In the reported evaluation, DM0.5 matched PI 0.5 on one Franka single-arm Move capability, while outperforming DM0 and PI 0.5 across the other tested items.##

1. Fine-Tuning CapabilityFine-tuning capability shows how well the model adapts after entering a real task environment.On the RoboChallenge real-robot benchmark Table30 V2, DM0.5 was tested across 30 complex tasks and four heterogeneous robot embodiments. It reportedly scored 54.42 with a 43% success rate, ranking first in that evaluation.The model also refreshed results on several recognized grasping and navigation benchmarks. In LIBERO, DM0.5 reportedly reached 99.1% overall performance, including 94.1% in Clean scenarios and 94.4% in Random scenarios.The important practical detail is cost: Dexmal says a new downstream task can be fine-tuned within 18 hours on a single RTX 4090, while inference can run at 90 ms on the same class of GPU.##
2. Native Long-Horizon MemoryDM0.5 supports up to 60 seconds of native long-horizon memory. Dexmal notes that many current embodied models usually stay below 10 seconds of memory.Longer memory allows robots to handle multi-step tasks where the environment changes over time. For example, in a messy storage or household organization task, the robot needs to remember what it has already moved, what remains, and how the scene has changed.DM0.5 also supports video prompting. With long-memory capability, the model can understand a short human demonstration video and align its behavior to the demonstrated task, reducing reliance on language-only instructions.##
3. Robustness Against Interference

DM0.5 uses a dual-system design:- Sys2 handles high-level understanding, reasoning, and planning.

• Sys1 handles high-frequency action response.The model is strengthened against two common long-tail disturbances in embodied scenarios:- Camera disturbance: the robot should stay stable even when the camera viewpoint changes suddenly.
• Human interruption: the robot should understand that the task state has changed and adjust its next actions after a person interrupts the process.This is important because real-world deployment rarely looks like a clean lab demo. People move around, cameras shift, and objects are not always where the robot expects them to be.##

1. Multi-Embodiment SupportDM0.5 is trained with multi-robot and multi-task integration, so it is designed for cross-platform transfer.Dexmal says the model can adapt to several mainstream and heterogeneous robot types, including biped humanoids, wheeled robots, dual-arm and single-arm manipulators, and dexterous hands.The article lists examples such as Aloha, ARX, UR, W1, Unitree G1, Tiangong, Huaqin, and Realman. Dexmal also says the model can be deployed quickly on new unknown robot embodiments with lightweight adaptation.The broader point is that generalization should move robots away from fixed environments and fixed action scripts. A more general model can handle new objects, new scenes, new instructions, and new robot bodies with less custom work.## A Foundation Model Alone Is Not EnoughA strong embodied foundation model is only the first step. To make embodied AI work in industry, the model still needs tooling, deployment systems, data feedback, and real-world operation infrastructure.Dexmal uses a “three-stage rocket” idea to describe this stack:1. Stage one: DM0.5 as the general foundation model.
2. Stage two: DexDev developer platform, including DFOL2.0, MaaS, and DexOS.
3. Stage three: Ferrata, a multi-agent hybrid operation system for real scenarios.These layers are meant to make the model cheaper to call, easier to deploy across hardware, and more stable in real-world task environments.## I. DexDev Developer PlatformDexDev is Dexmal’s developer platform for reducing the complexity of applying embodied AI models.In the current embodied AI field, models, hardware, tasks, datasets, and deployment environments are often fragmented. Developers may need to understand algorithms, robot control, hardware adaptation, and scenario iteration at the same time.DexDev is built around three modules.###
4. DFOL2.0: A World-Model-Driven FrameworkDFOL2.0 is an embodied reinforcement learning and data-loop framework driven by Dexmal’s general world model, DW0.5.Its role is to help the model keep improving. Instead of relying only on expensive real-robot trial and error, DFOL2.0 uses high-fidelity virtual physical environments for lower-cost and lower-risk closed-loop policy training.It also feeds real-world task and failure data back to the cloud, helping the foundation model continue to evolve.According to Dexmal, DFOL2.0 can reduce real-robot training data demand by 60% and training cost by 40%.###
5. DexOS: A General Operating System for Embodied AIDexOS defines a standardized ECP, or Embodied Control Protocol, interface.The goal is to hide differences across heterogeneous robot hardware. Instead of solving a difficult “N × M” adaptation problem between many models and many robot bodies, DexOS tries to simplify deployment into an “N + 1” unified connection problem.This is meant to help DM-series models run across different hardware with lower cost, lower latency, and more stable control.###
6. Embodied MaaS ServiceDexmal also introduced an embodied MaaS service for the DM model family.The idea is to package foundation model capabilities as a service. Developers do not need to train models from scratch or handle every hardware adaptation detail themselves. They can call model capabilities more directly for robot deployment and upgrades.In Dexmal’s stack, DFOL2.0 helps the model improve, DexOS connects software and hardware, and MaaS makes the model easier to use at scale.## II. Ferrata: A Multi-Agent Hybrid Operation SystemOnce single-robot capabilities can be called through MaaS, the next challenge is multi-robot collaboration.Dexmal introduced Ferrata, a multi-agent hybrid operation system designed for real-world scenarios. It is meant to handle system-level scheduling across multiple goals, models, robot forms, and safety boundaries.Ferrata is built on the DM model family and Realtime-VLA. It is not limited to one specific robot. Instead, it coordinates tasks, models, hardware types, and safety mechanisms at the system level.Through task hierarchy, exception handling, human takeover, and data feedback, Ferrata aims to keep robots operating continuously in real environments.From DM0.5 to DexDev and Ferrata, Dexmal is trying to build a full infrastructure path from model capability to real productivity.The first layer is the general model. The second layer is platform infrastructure for training, MaaS, and operating systems. The third layer is a scenario-level operation system that helps embodied AI move from lab demos into production settings.## FAQ### What is Dexmal DM0.5?DM0.5 is an embodied foundation model introduced by Dexmal, also known as 原力灵机. It is positioned as a 4B-parameter general model for open-world robot tasks such as navigation, grasping, full-body control, and instruction-following.### How much data was used to train DM0.5?According to the original report, DM0.5 is built on 150,000 hours of data. This includes 50,000 hours of real robot operation data, 100,000 hours of egocentric first-person data, and large-scale scene reconstruction data covering 1 million square meters.### What makes DM0.5 different from DM0?DM0.5 doubles the model parameter scale to 4B and increases the data volume by 400% compared with DM0. It also adds long-horizon memory, embodied reasoning tasks, and trajectory alignment to improve task understanding and generalization.### Can DM0.5 run on consumer GPUs?The report says DM0.5 can complete expert-level fine-tuning for a new downstream task on one RTX 4090 in as fast as 18 hours. It also reports 90 ms inference latency on RTX 4090 and 50 ms on H100.### What is DexDev?DexDev is Dexmal’s developer platform for embodied AI applications. It includes DFOL2.0 for reinforcement learning and data loops, DexOS for cross-hardware control, and MaaS for easier model capability access.### What is Ferrata used for?Ferrata is a multi-agent hybrid operation system for real-world robot scenarios. It coordinates tasks, robot hardware, models, safety boundaries, exception handling, human takeover, and data feedback.### Is DM0.5 open source?The article focuses on DM0.5’s release and reported capabilities, but does not clearly state that DM0.5 itself is open source. Dexmal does maintain the open-source Dexbotic VLA toolbox, and DM0-related documentation is available through the Dexbotic GitHub repository.### Why does long-horizon memory matter in embodied AI?Robots often need to complete multi-step tasks where the scene changes over time. Longer memory helps a model track previous actions, current task state, and environmental changes, making the robot less likely to lose context during long tasks.## Related Tools- Dexmal: The official site for Dexmal, the company behind DM0, DM0.5, DexDev, and Ferrata.

• Dexbotic: An open-source VLA development toolbox from Dexmal for embodied intelligence research.
• RoboChallenge: A benchmark platform for evaluating robot task performance in embodied AI settings.
• Hugging Face: A model hosting platform referenced by Dexbotic documentation for DM0 model checkpoints.
• arXiv: A research paper platform where the DM0 technical paper is available.## Related Links- Dexmal Official Website: Official company and product information from Dexmal.
• Dexbotic GitHub Repository: Open-source VLA toolbox for pretraining, fine-tuning, inference, and evaluation.
• DM0 Tutorial in Dexbotic: Official Dexbotic documentation for using DM0-specific configurations.
• DM0 arXiv Paper: Research paper for “DM0: An Embodied-Native Vision-Language-Action Model towards Physical AI”.
• RoboChallenge DM0 Runs: Public RoboChallenge runs filtered for DM0 on Table30.
• Dexmal GitHub Organization: Dexmal’s GitHub organization page and open-source project list.## SummaryDM0.5 is Dexmal’s latest embodied foundation model, built around larger-scale data, a 4B-parameter architecture, longer memory, embodied reasoning, and trajectory alignment. Its main goal is to improve robot generalization across tasks, environments, and embodiments.The article also shows why the model alone is not enough. Dexmal’s DexDev platform, DFOL2.0, DexOS, MaaS, and Ferrata system are designed to turn model capability into deployable infrastructure.The key takeaway: DM0.5 is not just a model update; it is Dexmal’s attempt to connect embodied AI training, deployment, and real-world robot operations into one stack.