Recently, on the latest MolmoSpaces benchmark list for authoritative embodied intelligence zero-shot evaluations, initiated by AI2, the VLM Agent framework EchoPolicy-0 under the Magic Atom brand, which operates for general robot control, ranked first globally with a average success rate of 75.4% in robot operations, covering all 4 v1 tasks.
Under the same ranking list and group, the MagicBrain framework for self-embedded intelligent agents achieved rank fourth globally in the MolmoSpaces Combined list with a parameter scale of 3.4B and an average success rate of 72.4% in robot operations. Its抓取 success rate reached 94.4%, far exceeding the industry average.
The MolmoSpaces list is currently the most credible zero-sample evaluation benchmark in the field of embodied intelligence worldwide, targeting operations, navigation, and longer-distance embodied tasks. The MolmoSpaces ecosystem includes over 230,000 indoor environments and approximately 130,000 object assets, providing a rich source of variations for evaluation. Robots need to handle situations beyond fixed demonstration environments, find targets under different backgrounds, geometric structures, and object distributions, and complete tasks based on local conditions.
At the same time, MolmoSpaces includes movable rigid bodies and jointed interactive objects. The robot must not only identify the target in the image, but also change the object’s state through real simulation contact, making the robot’s task more challenging.
MolmoSpaces’ benchmarks clearly define the initial scenario, robot and camera configurations, task instructions, and success conditions. A complete list, unified scoring, and public results can transform scattered demonstrations into comparable data. The four tasks involved by Magic Atom EchoPolicy-0 and the MagicBrain framework include grabbing free objects, interacting with articulated objects, and multi-stage operations. These aspects examine the system capabilities of the embodied intelligence Agent framework from different perspectives, rather than focusing solely on a single metric.
Since 2026, embodied intelligence has moved from the demonstration phase into real production lines. The focus of competition has shifted to the “brain” of robots—the large embodied intelligence models. As robot motion control capabilities improve rapidly, the bodies of robots become more flexible and durable, and the practical applications of their brains evolve quickly as the boundaries of the body are opened up. Accordingly, market research is shifting towards: how to perform more complex tasks and how to quickly get started with new tasks. The evaluation by MolmoSpaces also focuses on the generalization ability of large embodied intelligence models.
The person in charge of the MagicLab Studio, a magic atom research and development laboratory, said: “Regarding the choice of the embodied brain approach, we recognize the hierarchical architecture approach, and we have found that instruction following ability is a challenge for current models in general tasks. EchoPolicy ensures instruction following ability through the Agent framework and trajectory scoring; MagicBrain optimizes instruction following performance through end-to-end visual guidance. Both have achieved good results.”
EchoPolicy-0 : Let the robot multitask share a set of thinking and action methods
The robot needs to continuously link the language goal, visual observation, and physical action to complete a task. For example, after grasping an object, it must determine whether it has actually held it securely; before opening the cabinet door, it must find a position where it can interact; if a single action does not meet expectations, the robot needs to adjust its plan based on the new situation.
These actions appear different, but they are common challenges that embodied intelligent robots face across tasks: where is the goal, what is the current state, which capability should be called, whether the action achieves its purpose, and how to continue after failure.
MagicLab Studio revolves around these common issues across tasks. It organizes visual language models, geometric operation tools, and general action strategies into a unified VLM Agent framework called EchoPolicy-0, enabling the observation of the environment, understanding objectives, decision-making capabilities, action execution, and verification of results to form a closed loop. Different tasks share this mechanism, and specific interactions and movements are carried out by the corresponding executors.
The person in charge of the EchoPolicy-0 framework development said: “The versatility of EchoPolicy-0 lies in enabling capabilities to be uniformly called, combined, and verified. The current version of EchoPolicy-0 has completed verification for four types of operations on the Franka one-arm platform.”
Regarding the usage of VLA, MagicLab Studio introduces multi-candidate trajectory selection based on VLM target guidance: the motion model first proposes various possible movements, the candidates are projected onto the image plane, then the more suitable trajectory is selected according to the target specified by VLM, and continuous verification occurs after execution. High-level visual understanding thus participates in low-level motion selection, rather than merely staying at task decomposition.
EchoPolicy-0 framework completed 3915 full-scale tests on four operation tasks of MolmoSpaces, achieving a 75.4% equal-weight average success rate for all four tasks. As of October 13, 2026, it ranks first on the MolmoSpaces Combined list.
MagicLab Studio regards four evaluations of the EchoPolicy-0 framework on MolmoSpaces as a systematic verification of a unified framework, and uses the individual results to guide subsequent iterations. EchoPolicy-0 presents a system roadmap for general robot operations. The unified Agent framework can organize different types of operational capabilities: the VLM Agent manages semantic goals and task states, geometric tools provide inspectable spatial interactions, general action strategies generate candidates, trajectory selection constraints guided by VLM goals direct the actions, and continuous verification connects them together.
EchoPolicy-0 addresses the implementation scenario of embodied intelligence models by answering a question: how can we ensure that the action proposed by the model is indeed the correct one? It does not modify the model’s input; instead, among multiple candidate actions provided by the model, it filters out the most suitable one based on the visual target, and continuously verifies it after execution. For downstream customers, the value of this approach lies not in the single-point success rate—when a new type of task emerges, there is no need to build a new system from scratch; simply integrate the capabilities into the same framework is sufficient.
Self-developed brain: MagicBrain enables the model to “see” from “demo available” to “production line available”
The embodied AI Agent framework MagicBrain comes from the MagicLab Studio. In terms of four operation tasks—grabbing, placing, opening, and closing—MagicBrain completed 3915 full evaluations. The average success rate for the four v1 tasks ranked fourth globally. Among these, both the grabbing (Pick-v1) and grabbing and placing (Pick & Place-v1) tasks topped the global rankings, with a grabbing task success rate of up to 94.4%.
Visual-Language-Action (VLA) is a model architecture that has received significant attention in the field of embodied intelligence. The ability of models to follow language is a long-standing common issue in the industry. Having a model execute commands like “put an apple in the basket” involves two difficult tasks: first, understanding the term “apple,” and second, identifying at the pixel level which one is the “apple.”
In the research and testing of MagicLab Studio, it was found that in most cases of failed抓取, it wasn’t due to unstable action execution, but rather because the correct object was grabbed incorrectly or placed in the wrong position. In other words, VLA is strong enough in action learning, but its real weakness lies in binding noun phrases to image regions—this is also the part where training signals are the least dense: trying to grab “apple” but getting “pear” instead, the mistake lies not in the hand, but in the eyes.
In fact, VLA’s ability to follow visual signals is stronger than its ability to follow language instructions: “understanding speech” is VLA’s weakness, while “understanding images” is its strength. For this reason, MagicBrain uses a technology called In-Context Visual Guidance, and VLA is only responsible for performing actions; finding targets is left to the more skilled VLM Agent that is better at interpreting images.
In the future, this visual guidance technology can also be applied to more scenarios such as interaction. After all, creating what the model should see is not only useful for抓取. MagicBrain provides a more efficient division of labor paradigm. Language specifies “what to do”, while pixels specify “which one and where it is”. The two types of information are assigned to the appropriate models. It does not introduce new model structures or pile up larger parameters; it simply reconfigures the information transmission path.
“Based on our understanding of language in VLA, we enable it to ‘see’ further and know what to look at,” said the head of MagicLab Studio. This technical approach does not rely on increasing parameter scale to obtain large model performance. The parameter scale of the models involved in this evaluation is only 3.4B.
This means that any team with basic data capabilities can replicate it at low cost, and gain benefits without worrying about the model size – on the path toward the scale-up of embodied intelligence, a division of labor that focuses on combinations rather than full functionality may be closer to reality than “one model serving everything”.
At the industry level, Magic Atom answered a highly relevant question: whether VLA is actually affordable and applicable. Grabbing and placing are the high-frequency basic actions that enable embodied intelligence to move from the lab into real-world scenarios—factory sorting, warehouse handling, and home organization all rely on them. The increase in grabbing success rate from 70% to 94.4% means that VLA is approaching the threshold of being “available in production lines.” For the industry, every slight increase in success rate expands the range of scenarios that can be commercialized.
The design philosophy of the “embodied brain” of magical atoms
“We came in first in the MolmoSpaces ranking, but what Magic Atom is concerned about is the real issue behind that ranking: how the brain of embodied AI robots should be structured, and where it should be located.” said the person in charge of MagicLab Studio.
Magic Atom believes that embodied intelligence should progress along a layered collaboration technology path: the upper-level model is responsible for understanding, planning, and error correction, while the lower-level model converts task intentions into reliable physical actions.
The upper-level model is responsible for visual understanding, task decomposition, planning and decision-making, error correction, and memory management. Currently, cutting-edge basic models already have a good foundation in these areas, but they still need to be adapted to spatial relationships, physical constraints, and continuous interactions in embodied scenarios. The relevant person in charge believes that in the short term, the industry should focus on building Harness around the model operation, organizing context management, tool invocation, execution feedback, and evaluation mechanisms, and exploring feedback-driven iterative optimization, including the RSI direction. In the long term, autonomous training and post-training processes based on real-world scenario data are needed, gradually enhancing the model’s embodied understanding and decision-making capabilities.
The responsible person stated that the underlying model for action generation and control execution can be implemented using VLA, WAM, or other hybrid architectures. The choice of architecture should serve the actual tasks, with emphasis on instruction compliance, action accuracy, execution efficiency, and generalization capabilities.
Currently, the layered architecture itself has been widely discussed in the industry, but one key issue remains unresolved: how the task intentions at the upper level can be accurately understood and reliably executed by the lower level.
In experiments with MagicLab Studio, it was found that existing action models still tend to make errors in target selection and have misunderstandings regarding spatial constraints in OOD (Out-of-Distribution) scenarios. As a result, the team tried two complementary approaches to solve this issue.
The first one is EchoPolicy-0, which ranked first on the MolmoSpaces list. Through Agent scheduling, candidate selection, and closed-loop verification, it improves the consistency between execution and instructions. The upper-level model breaks down tasks into clear sub-targets, and the system evaluates multiple candidate action trajectories generated by the lower-level model to select the one that best matches the current instruction. After execution, based on new observations of progress, it decides whether to continue, retry, re-plan, or backtrack, and calls geometric tools as needed to perform specific operations.
The advantage of this technical approach is that it allows for rapid iteration by improving scheduling, selection, and validation mechanisms, without having to retrain the underlying model, thereby enhancing the task completion capabilities of existing models.
The concept of MagicBrain is to improve the target alignment ability during the action generation phase through explicit visual guidance and accompanying training.
What is worth learning from is the visual guidance mechanism between the upper and lower layers. MagicBrain injects the interactive targets, placement areas, and short-term motion history identified by the upper layer into the observation of the lower model in the form of visual markers, so that the task intent corresponds directly to the specific positions in the image. This method does not require changing the VLA network architecture; through subsequent training, the model learns to connect the visual markers with operational semantics and behavioral actions, thereby reducing ambiguity when relying solely on language instructions.
Two approaches operate on two different stages: “action generation” and “execution, selection, verification”. They complement each other well: visual guidance helps generate actions that align with goals, while Agent scheduling is responsible for selection, execution, inspection, and error correction.
This time, EchoPolicy-0 took first place, and MagicBrain took fourth place. The effectiveness of both paths has been verified.
When talking about the future, Magic Atom said: “In the next phase, we will further explore the combination of both, creating a complete feedback loop from task understanding to physical execution. We will also enter real industrial scenarios, continuously testing the task success rate, operational stability, and deployment costs, and improve the data and model platform during this process.” We solidify verified capabilities into reusable skill modules, standard interfaces, and solutions. At the same time, the system accumulates data from successes, failures, and manual corrections, establishing a continuous iteration mechanism of “deployment—collection—training—evaluation—re-deployment” to transform engineering experience into foundational model capabilities.