Qantara Bridge-Flow Training for Multi-Paradigm JEPA Control

Jul 16, 2026

Upload by Noah

Joint-Embedding Predictive Architectures (JEPAs) for control from raw pixels typically commit during training to a single inference paradigm, either trajectory optimisation in a learned dynamics model or direct behaviour cloning. Qantara introduces an end-to-end JEPA that unifies these options by combining Brownian-bridge interpolation between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. A single checkpoint supports multiple inference modes without retraining, achieving strong control and new SOTA on OGBench-Cube and competitive performance on Push-T and Cube under matched compute.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

Introduction

Motivation: inference-time constraints

Prior work clusters

Qantara approach and objective

arXiv :2607 .04978v 1 [ cs .LG] 6 Jul 2026 Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control Ruslan Rakhimov∗ George Bredis Yuriy Maksyuta Daniil Gavrilov T-Tech 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. 1 Introduction A robot deployed for visuomotor control faces inference-time constraints (per-step latency, the size of the action search space, whether a goal observation can be supplied) that the training run cannot anticipate. The same learned predictor is useful in different ways under different constraints: rolled out as a forward dynamics simulator under a planner when the search budget allows, or read off asa behaviour-cloning policy when latency is tight. Every JEPA world model trained from pixels for control today commits to one of these inference paradigms at training time, and a checkpoint trained for one cannot serve the others without retraining. Two clusters of prior work tile the relevant design space. Sub-billion JEPA world models commit to a single inference paradigm: PLDM [36], DINO-WM [42], and LeWM [29, 4] train action-conditional latent dynamics and plan via trajectory optimisation, while V-JEPA-2-AC [3] post-trains a 300M action head on a 1B video JEPA for direct behaviour cloning. Multi-paradigm checkpoints exist in a second cluster that operates in pixel space at billion-class scale: PAD, UVA, UWM, DUST, and Cosmos Policy [16, 22, 43, 39, 21] denoise observations and actions jointly and serve forward dynamics, inverse dynamics, and behaviour policy from one set of weights. The (sub-billion JEPA, multi-paradigm) cell is empty. ∗ Correspondence: [ru.rakhimov@t-tech.dev](ru.rakhimov@t-tech.dev. Project page:)[. Project page:](ru.rakhimov@t-tech.dev. Project page:) [https://corl-team.github.io/qantara](https://corl-team.github.io/qantara) Accepted at ICML 2026 Workshop on Decision-Making from Offline Datasets to Online Adaptation: Black-Box Optimization to Reinforcement Learning. | Training | | | | | | | | --- | --- | --- | --- | --- | --- | --- | | | action latent state noised | v1 ẑ0 va0 ẑ1 va2 ẑ3 va3 ẑ4 vaT−1 ẑT action head latent head | | | | | | | | clean context noisy inputs | | | | | | | | | | | o3 | | | Edge-aligned (τa, τz) sampling | | | --- | --- | | τz | | | | τa | Figure 1: Qantara training and edge-aligned (τa ,τ z ) sampling. Left: block-causal predictor on (a, z) tokens;

FAQs

What problem does Qantara address in JEPA-based control models?

Existing JEPA world models are trained for one inference paradigm, so a checkpoint cannot serve planning and imitation-style control without retraining. Qantara targets this deployment-time mismatch by enabling multiple inference paradigms from one checkpoint.

How does Qantara’s training objective combine state and action modeling?

Qantara pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. It also samples per-token noise levels along edge-aligned modes to match inference queries.

Which inference paradigms can use the same trained Qantara checkpoint?

The same weights support latent planning, behaviour-cloning action sampling, and inverse dynamics via a video–inverse composition. This lets the model defer the planning-vs-imitation choice to inference under deployment constraints.

Jul 16, 2026

Upload by Noah