Can the EMG model be used directly on a different person without providing new data or performing individual training?
OctoSense: Octopus Dynamics achieves the world's first zero-shot EMG generalization across individuals. We're launching three data collection solutions to translate human motion and force into training data for robots.
On the eve of the World Robot Conference (WRC), Octopus Dynamics provided the answer:SynEMG EMG Foundation Model: A Global FirstEnable EMG for Embodied Data CollectionCross-Individual Zero-Shot Generalization, truly enabling myoelectric fromLook-aliketowardsOne model, thousands of users. For individuals not included in the training set,The model recovers hand pose and force state without requiring new individual data or incremental training.
围绕具身数据，章鱼动力正式推出OctoSense数采产品：以Fish-eye head ring, EMG wristband, and exoskeleton-integrated data acquisition gloveCore hardware, formingOctoS-Ego, OctoS-EgoBio, OctoS-DexUMIThree data acquisition solutions. All three have been productized and are ready for hands-on testing.
The Fish-Eye Ring captures first-person visuals via four fisheye cameras and uses the AINative Binocular Spatial Large Model to recover true scale and 3D depth.
The EMG wristband captures forearm electromyographic signals, which the SynEMG large model then uses to reconstruct hand posture and exertion state.
The exoskeleton-based data collection glove directly captures human hand motion data and aligns it with the dexterous robotic hand with near-zero loss.
Powered by the SYNData model algorithms and data engineering capabilities, visual, EMG, and other sensor data collected through three solution pathways undergo physics-informed recovery, multimodal alignment and slicing, quality cleaning, and semantic information generation. This process yields standardized, traceable datasets ready for downstream use.
• What's truly scarce isn't just raw data, but training-ready data.
Embodied AI data collection is scaling rapidly, but raw volume growth does not equate to proportional model capability gains. Large volumes of raw data often remain ineffective due toMultimodal timestamp misalignment, 3D scale drift, missing pose and force information under occlusion, and incomplete annotations.issues,Cannot start training。
Octopus Power officially introduces the value formula for embodied data: Embodied Data Value = Physical Completeness × Scenario Diversity × Collection Scale × Data Quality.
这四项Not addition, but multiplication.If any single component approaches zero, the overall value of the data is significantly diminished. Therefore, embodied data systems must answer:How to build data collection pipelines and transform raw data into trainable data assets.
• OctoSense: Three Acquisition Hardware Options, Three Data Collection Solutions
OctoSense is a system based onA data acquisition product system built around real human operationThe fish-eye headband, EMG wristband, and exoskeleton-based data acquisition glove each address a distinct critical gap. By combining these three hardware types in various configurations, they form three majorComplementaryData Acquisition Solution:
The three solutions are not mutually exclusive but rather complement each other in行为自然度、采集规模、物理模态完整度与真值精度Provide various combinations to maximize the embodied data value formula.
01 | SYNEMG EMG Large Model: First Global Achievement of Zero-Shot Cross-Individual Generalization for Electromyography
• Visual systems struggle to infer hand pose and actual force exertion when the hands are occluded.
In real-world scenarios, hands are frequently occluded; even when visible, it is difficult for vision alone to determine whether the operator is applying force and how much force is being exerted.
Electromyography signals originate from muscle activity.preceding visible actions. It can supplement the model with two types of information that are difficult to obtain visually:Hand pose and grip force under occlusion。
EMG Wristband: Only 80g, 1-channel forearm electrode array with bare hand for zero burden
• The real challenge in scaling EMG is: "Will it work for a different person?"
Muscle structure, skin condition, and movement habits vary across individuals. Traditional EMG models are typically trained for fixed users and tasks, keeping EMG in a "small-scale experimental" state: if each new user requires fresh data collection, training, and calibration, scaling embodied data acquisition becomes impractical.
The SYNEMG EMG large model has completely transformed this situation.
章鱼动力以大规模无标注多通道时序信号进行Self-supervised pre-training, let the model first learn across individuals and wearing statesGeneral EMG Representations, then extend to tasks such as gesture recognition, continuous hand motion recovery, and force state estimation. Learning from different signal distributions enables transferable features.Common Structurethereby enabling个体差异、佩戴位置和长时间漂移保持鲁棒。
SYNEMGWorld's first EMG cross-subject zero-shot generalization for embodied data collection：
• Zero-shot generalizationNew instances require no additional training data or fine-tuning calibration.Wear and use.
• Cross-person Pose Recovery: The same model can accurately reconstruct different individuals.Hand pose and exertion state.
• Unprecedented precision breakthrough: SynEMG achieves zero-shot cross-subject pose reconstruction误差降低50%，率先实现了Millimeter-level absolute positioning accuracy.
• End-to-end scaleEMG breakthroughs are no longer just about "better accuracy" — they're moving from "one model, one person" to "one model, thousands," enabling efficient, large-scale collection of embodied data.
When combined with the OctoS-EgoBio solution, the first-person visual input from an eyeglass headband and a myoelectric wristband can continue to reconstruct hand pose even when the hand is occluded, achievingMillimeter-level joint accuracy and 0.1N active force controlThis goes beyond just "what was seen" in a single action—it also captures "how the hand moved" and "how much force was actually applied." Additionally, because the hand remains naturally exposed, it offers advantages over traditional real-device capture methods.1000x efficiency的规模化潜力。
02｜外骨骼同构数采手套：让人手动作与发力，近乎无损直达灵巧手真机
OctoS-Ego and OctoS-EgoBio are designed for larger-scale, more diverse scenarios requiring higher behavioral naturalness, such as pre-training data collection. For post-training use cases...物理真值要求更高tasksThe exoskeleton-integrated isomorphic data acquisition glove further enhances data precision.。
It can sync records.Full-sole contact force distributionand亚毫米级手部姿态More critically, the exoskeleton-based data collection glove pairs with Octopus Power's self-developed dexterous hand.Uses the same joint degrees of freedom and kinematic structure.The acquisition module and the dexterous hand body feature an integrated, unified design.
Biomimetic Exoskeleton Data-Gathering Glove: Fully integrated with our proprietary dexterous hand, it maps human hand data to the dexterous hand with near-zero loss.
Why "Isomorphism" Matters More Than Just "High Precision"
Typically, in addition to real-device teleoperation data collection, any operation data without a physical body must undergoCross-App Retargeting: Map hand motions to dexterous hands with different numbers of joints, degrees of freedom, ranges of motion, and structures. This process introduces deformation and errors; while actions can be "forced" through,Accurately translating raw contact relationships, force distributions, and fine-grained joint states remains challenging.。
This isRetargeting GapThis isn't a problem that can be fully resolved through additional post-processing steps; it stems from structural differences between the acquisition and execution ends.Inherent Loss。
Exoskeleton-homomorphic data-gathering gloves eliminate this gap at the source:Human-powered data collection glove for manual operationWhen the data-gathering glove and the dexterous hand are fully isomorphic, the collected manipulation data can be directly mapped toThe灵巧 hand aligns with a ratio of approximately 1:1.。
The OctoS-DexUMI solution, which combines a fisheye head ring with an exoskeleton-integrated data-gathering glove, offers value beyond "higher accuracy." It isEnable dexterous manipulation by making data usable by robotic hands from the moment of collection.
03 | Fisheye Ring: The four fisheye lenses significantly reduce hand occlusion, while the dual-fisheye large model provides accurate 3D scale.
Embodied manipulation requires not only actions and force, but also awareness of what the operator sees, where the target is, and the spatial relationships in which the action occurs.
A single standard wide-angle camera has limited field of view, causing hands to easily move out of frame during large movements. Additionally, monocular images lack scale information, making it difficult to recover true spatial scale and distance.
The Octopus Power Headband integrates四颗鱼眼相机: Two in front and one on each side, forming approximately330° × 150° wide field of view, covering both hands and the interaction between the object and its surroundings. For movements such as walking, turning, or reaching out to grasp objects, the panoramic field of view canSignificantly reduce the likelihood of hands leaving the frame.。
Two forward-facing fisheye lenses form a stereo binocular system.基线为60mm。借助AINative Binocular Spatial Large Model, the system is capable ofRecovering Real-World Scale and 3D Depth from Fisheye Images, while maintaining a large field of viewMillimeter-level depth estimation accuracyEstablishes real-world 3D scale between hands, objects, and the environment.Four fisheye cameras solve "incomplete views," while front-facing binoculars solve "inaccurate measurements."。
Ego头环：环形布局集成四颗鱼眼相机，前视两颗兼作双目
04 | SYNData: Model Algorithms and Data Engineering to Transform Raw Data into Standardized Datasets
Hardware determines whether data can be collected; SYNData determines whether the collected data is usable.
SYNData is more than a data processing platform—it's a comprehensive data system powered by model algorithms and data engineering. It calculates the physical states and semantic information of operations from images, electromyography signals, and other sensor inputs, then performs multimodal alignment, slicing, cleaning, and organization to produce standardized, traceable datasets.
After SYNData processing, OctoSense's three solution data streams are no longer fragmented video and sensor signals. Instead, they become a reusable data asset with complete physical information, aligned multi-modal correlations, and traceable quality results for downstream retrieval and reuse.
OctoSense captures the complete record of real-world operations, while SYNData processes and organizes multimodal data into usable datasets.
• From one real-world action captured to a lifetime of robot expertise accumulated.
Returning to the original formula: Embodied Data Value = Physical Completeness × Scenario Diversity × Collection Scale × Data Quality.
SYNEMG EMG Large ModelEnable cross-subject generalization to eliminate per-user training and calibration for EMG.Exoskeleton-based Isomorphic Data Collection GlovesResolve cross-ontology alignment to enable near-lossless, direct deployment of ontology-free data on dexterous robotic hands.Four-Eye Head RingDual-camera completion for first-person vision and 3D scale; SYNData uses model algorithms and data engineering to convert multi-modal collected data into standardized, traceable datasets.
OctoS-Ego, OctoS-EgoBio, and OctoS-DexUMI cover diverse capture needs, while SYNData provides unified capabilities for data computation, processing, and delivery. Together, they transform a single real-world operation into robot-ready data.
Enable EMG signals to transcend individual differences, align human operational data with near-lossless fidelity to the real robot, and turn every physical action into learnable experience for the robot.