SourcesX-VLA authors
publisher · 2 sources
X-VLA authors
2 sources read from this publisher, newest first, with the methods each one describes.
X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
research paper · technical report · 2025-10-11 · 47 methods
Low-Rank AdaptationFlow-based vision-language-action modelEmbodiment-specific fine-tuningCross-attention-based resamplerCross-domain and cross-trajectory shufflingDomain-specific resamplers and action headsEmbodiment-aware language prompt concatenationEmbodiment-aware language promptsOpen-loop evaluationAction-modality isolationConcatenation of proprioceptive states, action tokens, and time embeddingsDAgger-style data collectionDomain-specific parameter learning with a shared backboneEmbodiment-specific soft promptsHeld-out validation set constructionHeterogeneity-guided MoE routingHeterogeneous Low-Rank AdaptationHPT-style domain-specific input projectionJoint fine-tuning across multiple embodimentsJoint policy adaptationLoad-balancing regularizationMultimodal Diffusion Transformer decoderNatural-language hardware-configuration promptsOptimal-transport path for flow matchingParallel MLP-Mixer-based action decoderPer-hardware-configuration learnable prompt embeddingsPeriodic policy training with failure-targeted demonstration collectionPretrained vision-language encoderPrivacy-sensitive content filtering and bias mitigationPrompt warm-up with the pretrained backbone frozenReduced learning rates for soft prompts and vision-language modulesRepeated stage-I demonstrations until cloth keypoints emergeRepresent the next four seconds of intended trajectory with 30 anchor pointsRotate6D encoding for absolute end-effector rotationSeparate embodiment-specific action projection headsShared vision backboneSoft prompt learningSoft prompting with learnable embodiment-specific data-source embeddingsStandard Diffusion Transformer decoder conditioned on multimodal featuresStandardize actions as end-effector pose and discretized gripper state
2toINF/X-VLA-Pt · Hugging Face
other · model card · 2025-10-11 · 4 methods