Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control
2026-07-06 • Machine Learning
Machine LearningComputer Vision and Pattern RecognitionRobotics
AI summaryⓘ
The authors introduce Qantara, a model that learns a flexible world representation from images to control actions, allowing multiple ways to decide what to do without retraining. Unlike previous models that stick to one method, Qantara can plan ahead, imitate expert behavior, or deduce actions from changes in its internal prediction. They achieve this by carefully training with a special technique that focuses on the edges of a noise pattern used during learning. Tested on control tasks, Qantara matches or beats previous best models while supporting different decision methods from the same training.
Joint-Embedding Predictive Architectureslatent world modelstrajectory optimisationbehaviour cloningBrownian-bridge interpolantnoise-to-data flow matchinglatent planninginverse dynamicsreinforcement learningcontrol from pixels
Authors
Ruslan Rakhimov, George Bredis, Yuriy Maksyuta, Daniil Gavrilov
Abstract
Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning. A single checkpoint that serves both would defer this choice to inference, when deployment constraints (rollout cost, observation accessibility) determine which path wins. We present Qantara, an end-to-end JEPA whose joint training objective pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. The same checkpoint serves three inference paradigms without retraining: latent planning, behaviour-cloning action sampling, and inverse dynamics, which we query through a video-inverse composition that first predicts the next latent without action conditioning, then extracts the action. Training concentrates mass on the edges of the (action-time, state-time) noise square, where inference queries the predictor: replacing it with uniform interior sampling drops Push-T planning from 90.1 to 53.3 SR at matched compute. On the LeWM control suite, Qantara reaches a 91.2 SR three-train-seed average and sets new SOTA on OGBench-Cube (+7.7 SR over DINO-WM, +19.7 over LeWM). From the same weights, the behaviour-cloning and video-inverse paths reach 82-83 SR on Push-T and 71-73 SR on Cube. These results move JEPA world models from single-paradigm planners to multi-paradigm controllers.