mlsanityData to 7 Oct 2026

Juno: Taming Predictive Latents for Vision-Language-Action Models

arXiv 2610.09940 ↗First posted 7 Oct 2026cs.RO cs.CVcited by 0

On the map: Vision-language-action models → World model-based VLA also fits: VLA architectures and generalization

Abstract (arXiv)

Joint-embedding predictive architectures (JEPAs) predict masked or future observations in representation space, offering a natural source of predictive latents for vision-language-action (VLA) models. Yet making these latents useful across pretraining, policy learning, and deployment requires addressing three failures: mismatch with embodiment-specific control, interference with action learning, and teacher miscalibration under distribution shifts. We introduce Juno, a unified framework built around one action-conditioned JEPA that serves as a control-aligned representation backbone, a predictive teacher, and an adaptable dynamics model. During pretraining, we train it on embodiment-matched trajectories and use a dynamic CLS loss to transfer motion-weighted patch dynamics to a compact global state. During policy learning, we fuse current-frame JEPA patches into VLA perception and use a decoupled reasoning branch with separate transformation parameters to distill future latent states for action generation. During deployment, we adapt the world model on all observed transitions, including failed rollouts, freeze the adapted teacher, and re-align the policy on verified executions using LoRA adapters and a trainable action head, without expert corrections or task rewards. On SimplerEnv, Juno raises average success from 60.9% to 68.5% over Qwen3GR00T, the strongest baseline, and test-time adaptation further reaches 72.7%; on a real robot, it retains 70%--75% success under background, height, and object shifts where the base policy collapses to 0%.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in World model-based VLA, one of 13 directions in Vision-language-action models. Builds on 42 of its 64 references that are in the map. First posted on arXiv 7 Oct 2026.

Builds on

42 of 64 references.

Closest ancestors · 7

Same direction, newest first

  1. 2026LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies14 Jun
  2. 2026Test-Time Training for Visual Foresight Vision-Language-Action Models6 May
  3. 2026FutureVLA: Joint Visuomotor Prediction for Vision-Language-Action Model11 Mar
  4. 2026World Guidance: World Modeling in Condition Space for Action Generation25 Feb
  5. 2026FRAPPE: Infusing World Modeling into Generalist Policies via Multiple Future Representation Alignment19 Feb
  6. 2026JEPA-VLA: Video Predictive Embedding is Needed for VLA Models12 Feb
  7. 2026VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model10 Feb

Other references · 35

  1. 2024OpenVLA: An Open-Source Vision-Language-Action Modelcited by 2,777
  2. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,297
  3. 2025π0.5: a Vision-Language-Action Model with Open-World Generalizationcited by 1,653
  4. 2025V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planningcited by 738
  5. 2025SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Modelcited by 502
  6. 2024Revisiting Feature Prediction for Learning Visual Representations from Videocited by 490
  7. 2024Evaluating Real-World Robot Manipulation Policies in Simulationcited by 484
  8. 2020Learning Invariant Representations for Reinforcement Learning without Reconstructioncited by 424
  9. 2024CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulationcited by 411
  10. 2025UniVLA: Learning to Act Anywhere with Task-centric Latent Actionscited by 386
  11. 2024DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planningcited by 357
  12. 2025WorldVLA: Towards Autoregressive Action World Modelcited by 351
23 more
  1. 2024Learning to (Learn at Test Time): RNNs with Expressive Hidden Statescited by 329
  2. 2024Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representationscited by 279
  3. 2026LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixelscited by 237
  4. 2025LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristicscited by 227
  5. 2020Self-Supervised Policy Adaptation during Deploymentcited by 121
  6. 2025FLARE: Robot Learning with Implicit World Modelingcited by 105
  7. 2026V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learningcited by 104
  8. 2026Being-H0.7: A Latent World-Action Model from Egocentric Videoscited by 68
  9. 2020Tent: Fully Test-time Adaptation by Entropy Minimizationcited by 59
  10. 2025What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?cited by 37
  11. 2025Latent Action Pretraining Through World Modelingcited by 21
  12. 2025EVOLVE-VLA: Test-Time Training from Environment Feedback for Vision-Language-Action Modelscited by 20
  13. 2026InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalizationcited by 18
  14. 2026AdaJEPA: An Adaptive Latent World Modelcited by 16
  15. 2026RoboTTT: Context Scaling for Robot Policiescited by 13
  16. 2026Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decodingcited by 11
  17. 2026AdaWorldPolicy: World-Model-Driven Diffusion Policy with Online Adaptive Learning for Robotic Manipulationcited by 9
  18. 2026RepWAM: World Action Modeling with Representation Visual-Action Tokenizerscited by 9
  19. 2026TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action Modelscited by 7
  20. 2026H-WM: Robotic Task and Motion Planning Guided by Hierarchical World Modelcited by 6
  21. 2026WAM-TTT: Steering World-Action Models by Watching Human Play at Test Timecited by 6
  22. 2026VLAFlow: A Unified Training Framework for Vision-Language-Action Models via Co-training and Future Latent Alignmentcited by 2
  23. 2026VANE: Reliable Test-Time Training for Vision-Language-Action Models via Future Visual Representation Predictioncited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026ATI-VLA: Action-Centric Predictive Vision-Language-Action Models via Actionable Alignment Then Adaptive Injection
  2. 2026V-JEPA Policy: Building Effective World-Action Models on Predictive Visual Latents
  3. 2026Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models
  4. 2026JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling
  5. 2026Devol-ONE: One Autoregressive Mixture of Transformers to Unify Vision-Language-Action and Latent World Modeling
  6. 2026LARA: Latent Action Representation Alignment for Vision-Language-Action Models
  7. 2026SLIP-VLA: Single-Step Latent Imagination for Policy Learning in Vision-Language-Action Models
  8. 2025Seeing Space and Motion: Enhancing Latent Actions with Geometric and Dynamic Awareness for Vision-Language-Action Models