SourcesXiaomi
publisher · 3 sources
Xiaomi
3 sources read from this publisher, newest first, with the methods each one describes.
XiaomiRobotics/Xiaomi-Robotics-1-RoboCasa365 · Hugging Face
software documentation · model card · 2026-07-28 · 6 methods
Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories
research paper · technical report · 2026-07-16 · 24 methods
Flow matchingAdaptive Layer Normalization (AdaLN)Mixture-of-Transformers (MoT)Pre-training followed by post-training for robot policiesAutomated language annotation of manipulation trajectoriesBeta-distributed flow-matching timestep samplingWinner-takes-all action-candidate supervisionAsynchronous fine-tuningAuxiliary action-generation supervision for the vision-language modelChain-of-thought labelingChoice Policies for VLM action generationCross-embodiment end-effector frame orientation alignmentFiltering idle trajectory segmentsFive-step Euler integration for action samplingFixed-length segmentation of manipulation trajectoriesFour-timestep sampling to amortize VLM computationJoint flow-matching and VLM choice-policy training objectiveJoint vision-language and action trainingLanguage annotation of human-segmented post-training clipsLoss masking for missing embodiment-specific action dimensionsMasking action-related tokens from DiT attentionProducer-consumer parallel pipeline for trajectory captioningRelative-delta end-effector pose action representationVision-language token sequence packing
Xiaomi-Robotics-0: An Open-Sourced Vision-Language-Action Model with Real-Time Execution
technical report · 2026-02-13 · 21 methods
Flow matchingAsynchronous inference and executionAdaptive Layer Normalization (AdaLN)Mixture-of-Transformers (MoT)Beta-distributed flow-matching timestep samplingCausal attentionLearnable attention sink tokenWinner-takes-all action-candidate supervision1:6 vision-language-to-robot-trajectory samplingAction-chunk timestep alignmentChoice Policies for multimodal action predictionCross-validated consensus annotation for visual groundingJoint training on robot trajectories and vision-language dataOnline-error-weighted flow-matching lossPre-trained VLM relabeling to refine VQA and caption annotationsPrefixing the next inference with the current chunk's actions across the inference windowRoPE positional-index offsetShape attention mask for action-prefix conditioningTimestamp-based 30 Hz synchronization of input modalitiesTraining-time action-prefix conditioning for asynchronous executionVLM-generated embodied reasoning and planning data from robot trajectories