Driving · 2026-10-08

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

Yuchen Zhu, Chenyi Xu, Yulin Zhang, Gang Xu, Wentao Zhu

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

Juno is a unified action-conditioned JEPA framework for vision-language-action models. The authors train a control-aligned JEPA backbone with a Dynamic CLS loss, fuse its patch features into a VLA with a decoupled predictive reasoning branch, and run dynamics-first test-time adaptation that adapts the world model on all transitions then re-aligns the policy on verified executions. They report improved success in simulation and robust real-robot performance under distribution shifts.

Why it matters

The authors address three failures of predictive latents for VLA: representation mismatch with embodied control, interference with action learning, and teacher miscalibration under deployment shifts. Their dynamics-first test-time training uses failed rollouts as dynamics supervision while avoiding failed actions as policy targets, and they re-align the policy using the adapted teacher held fixed.

Method

Control-Aligned Pretraining: train an action-conditioned JEPA on embodiment-matched trajectories and use Dynamic CLS loss to transfer motion-weighted patch dynamics into a compact CLS token.
Decoupled Predictive Reasoning: fuse frozen JEPA patch features into VLA tokens and use a decoupled reasoning branch (with separate transformation parameters) distilled to future JEPA latent states for action generation.
Dynamics-First Test-Time Adaptation (Juno-TTT): adapt the world model on all deployment transitions (including failed rollouts), freeze the adapted teacher, and re-align the policy on verified successful executions using LoRA adapters and a trainable action head without expert corrections or task rewards.

Abstract (from 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 $0\%$.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-08 UTC

preprint