SourcesRobbyant
publisher · 5 sources
Robbyant
5 sources read from this publisher, newest first, with the methods each one describes.
From Foundation to Application: Improving VLA Models in Practice
research paper · technical report · 2026-07-07 · 26 methods
Mixture of expertsAuxiliary-loss-free load balancingBlockwise causal attention55-dimensional canonical state-and-action representationAutomated vision-language-model annotation of manipulation videosCausal video-representation distillation for temporal supervisionDepth-representation distillation for geometric supervisionDual-Query DistillationEgocentric SLAM for camera pose estimationFine-grained routed experts with a shared expertHand pose estimation for MANO trajectory recoveryHuman verification of robot state and video alignmentL2 action-regression lossMean-standard-deviation action normalizationMinimum valid-frame coverage filteringNear-static episode filteringPer-expert routing correction biasRelative joint-action targetsRobotic trajectory quality filtering pipelineSigmoid-based expert routingSign-based expert-load correction-bias updatesTrajectory smoothness filtering using jerk and derivative Z-scoresTrajectory-level quality controlVideo-adapted DINO and iBOT self-distillationVision-language-model filtering of egocentric videosVLM-based temporal subtask segmentation and instruction generation
A Pragmatic VLA Foundation Model
research paper · technical report · 2026-01-26 · 17 methods
Fully Sharded Data ParallelMixture-of-Transformers (MoT)Blockwise causal attentionAction-expert-specific sharding groupsAutomated anomaly screeningConditional flow matching for action generationDistillation-loss alignment of VLM queries with depth tokensFlexAttentionHuman review of filtered demonstration episodesLearnable image queriesMixed-precision policy with BF16 storage and communication and FP32 reductionsOperator fusion using torch.compileRandomize object poses during demonstration collectionRandomized-order sequential model evaluationRepeated trials for statistical robustnessShared self-attention across modalitiesVision distillation
Masked Depth Modeling for Spatial Perception
research paper · technical report · 2026-01-25 · 34 methods
Global-norm gradient clippingBF16 mixed-precision trainingDiffusion-policy action headAdditive Gaussian and shot-noise perturbationsBlock-wise depth masking for depth-completion evaluationBundle adjustment on tracked visual-odometry pointsCondition depth reconstruction on the complete RGB imageControl the monocular-depth and depth-completion tasks through the depth-token masking strategyConvStackDifferential learning ratesDINOv2 and Point Transformer grasp-policy conditioningGaussian-noise augmentation of depth mapsGenerate pseudo-depth labels from stereo infrared pairsHierarchical convolutional decoderImage augmentation with crops, flips, and synthetic degradationsInitialize framewise camera extrinsics to the SE(3) identityLeft-right consistency checking for stereo pseudo-depthLinear warmup followed by step decayMasked depth modeling for RGB-D depth completionMasking fully missing depth patchesModular multi-camera RGB-D capture systemNatural masks for missing depth regionsOmit global bundle adjustment for efficiencyOnline RGB-D tracking with SpatialTrackerV2Probabilistic masking of partially valid depth patchesRandom sampling of fully valid depth tokensRetarget human hand-object interactions using 3D keypoint correspondenceSemi-global matchingSeparate RGB and depth patch embeddingsShared spatial and modality-specific positional embeddingsSimulate real active RGB-D camera imaging to generate imperfect depthSynthetic RGB-D data generation using self-hosted 3D assetsUV positional encodingsValid-pixel-masked L1 depth loss
robbyant/lingbot-vla-v2-6b · Hugging Face
other · model card · 2025-08-04 · 8 methods
Mixture of expertsPredictive dynamics modelingDistilling current and future perceptual queries from pretrained perception modelsFine-grained expert segmentationLarge-scale multi-source pre-training data curationPerceptual-query token appendingShared expert isolationUnified representation for multi-part robot action signals
Robbyant/lingbot-vla
software documentation · code repo · 8 methods