Target-Consistent Action Guidance via Semi-Unbalanced Optimal Transport for Drifting World Models

Md Abtahi Majeed Chowdhury, Md Rifat Ur Rahman, Md Sakib Hossain Shovon, Jaesik Choi, Minhyuk SungarXiv 2026 · arXiv

TL;DR

This paper introduces Mass-Share Accentuation (MSA), a novel training method for one-step "drifting" world models that fixes a fundamental "target bias" in action guidance. By framing action following as a semi-relaxed optimal transport problem, MSA ensures that guidance vanishes exactly at the ground-truth target, leading to more accurate robot "imagination" and better downstream planning without any extra cost at inference time.

What Problem Does This Paper Address?

For a robot to plan its next move, it needs to "imagine" the future. Action-conditioned world models do exactly this, predicting how an environment changes based on a robot's actions. While diffusion models are the current gold standard for quality, they are computationally expensive because they require multiple steps to generate a single frame.

Drifting models offer a faster, one-step alternative by moving the iterative process from inference to training. However, to make these models actually follow actions (rather than just copying the current frame), researchers often use "accentuation"—a technique that pushes predictions away from the current frame.

The authors identify a critical flaw: target bias. Standard accentuation (borrowed from classifier-free guidance) keeps pushing predictions even after they've reached the correct future state. This leads to "overshooting," where the model imagines exaggerated or physically impossible motions.

Figure 1: In a controlled toy, standard accentuation (top) overshoots the target, while MSA (bottom) lands exactly on the recorded future.

Methods

The core of the paper is Mass-Share Accentuation (MSA). Instead of a simple "repulsive push," MSA treats the training process as a competition between two destinations: the recorded future (where we want to go) and the current frame (the "no-motion" counterfactual).

Key technical components include:

  • Semi-Relaxed Optimal Transport: MSA calculates how much "mass" a prediction should send to the current frame versus the target. If a prediction is far from the target, it sends more mass to the "no-motion" state, signaling that a lot of motion is still missing.
  • Bias Correction: The authors mathematically derive a "floor" share—the amount of mass the target itself would send to the current frame. By subtracting this floor, they ensure the guidance vector becomes exactly zero when the prediction matches the target.
  • Shared Translation: Unlike previous methods that apply different pushes to different samples (which can deform the generated distribution), MSA applies a single, shared translation vector to the entire batch of predictions at each location.
  • One-Step Inference: Crucially, all this complexity happens during training. At test time, the model remains a standard, lightning-fast one-step generator.

Datasets

The authors evaluated MSA across several challenging robotics benchmarks:

  • Push-T: A simulated task involving pushing a T-shaped block to a goal.
  • Robomimic (Lift & Can): Simulated manipulation tasks based on human demonstrations.
  • BridgeV2: A large-scale dataset of real-world robot videos.
  • T-block Toy: A simplified synthetic environment used to theoretically verify the removal of target bias.

Key Insights and Ideas

  • Guidance Needs an "Off" Switch: The most important insight is that guidance shouldn't be a constant force. It should fade as the model gets closer to the truth. MSA provides a principled way to achieve this "vanishing guidance."
  • Counterfactuals as Destinations: Treating the "no-action" state as a competing destination is a powerful shift from treating it as a "negative" to be repelled. It allows the model to measure "missing motion" more precisely.
  • The "Under-reaction" Paradox: Interestingly, the authors found that while theory predicts overshooting, many trained models actually under-react to actions. MSA helps here too: by providing a cleaner, bias-free signal, it allows the model to respond more strongly to actions without losing accuracy.

Key Results

MSA consistently outperformed baselines (like DriftWorld and W-Flow) across multiple metrics:

  • Better Action Ranking: On Push-T, MSA improved the Kendall $\tau$ (a measure of how well the model ranks different actions) by about 0.02, a significant margin in this domain.
  • Improved Response Slope: MSA increased the "response slope"—how much the predicted object moves relative to the actual action—bringing it closer to the ideal value of 1.0.
  • Lower Motion Error: It reduced moving-region Mean Squared Error (MSE) by 1% to 8% across different benchmarks.
  • Qualitative Fidelity: As seen in Figure 3, MSA predictions (far right) more closely track the ground-truth green contours compared to other drifting baselines.
Figure 3. Qualitative rollouts. Last context frame, ground truth six steps later, and each model’s prediction under the recorded actions. BridgeV2: top-left 191 × 191 crop. The green contour is the ground-truth object (arm on BridgeV2) at +6.
Figure 2: Quantitative sweep showing MSA (red star) achieving a better trade-off between response slope and outcome bias compared to accentuated baselines.

Limitations and Future Work

  • Stochasticity: The current formulation assumes a single "correct" future for every action. In the real world, the same action might lead to different outcomes (e.g., a ball rolling left or right). Extending MSA to handle multi-modal distributions is a natural next step.
  • Latent Space Constraints: The experiments were conducted in the latent space of a pre-trained VAE. While efficient, this might miss fine-grained pixel-level details that a direct pixel-space model could capture.
  • Benchmark Isolation: The authors honestly note that while MSA improves performance, current benchmarks don't perfectly isolate "bias removal" as the sole cause, as models often struggle with under-reaction for other reasons (like architectural limits).

Summary updated

Abstract

Action-conditioned world models let robots imagine the outcomes of their actions for planning and offline policy evaluation, but these uses need many rollouts, which multi-step diffusion makes expensive. Drifting models move the iteration from inference to training and generate samples in one network pass. DriftWorld brought this paradigm to action-conditioned world models; to strengthen action following, it repels predictions from the current frame, a construction borrowed from classifier-free guidance. We show that this accentuation has a target bias: the push persists even when a prediction matches the recorded future, in kernel, balanced, and unbalanced Wasserstein drifting fields alike. We introduce Mass-Share Accentuation (MSA) to remove this bias. MSA treats the current frame as a counterfactual destination competing with the recorded future in a semi-relaxed optimal transport problem; subtracting the closed-form share of the correct future makes the guidance vanish at the target, and the remainder is shared by all predictions. We prove that the recorded future is the unique fixed point along the motion direction for collapsed predictions, and that the guidance fades with prediction error. On Push-T, Robomimic, and BridgeV2, where trained models under-react rather than overshoot, MSA improves action ranking, response slope, and moving-region error over latent-space baselines at no inference cost, though these benchmarks do not isolate bias removal as the cause.
BibTeX
@misc{chowdhury2026target,
  author = {Md Abtahi Majeed Chowdhury and Md Rifat Ur Rahman and Md Sakib Hossain Shovon and Jaesik Choi and Minhyuk Sung},
  title = {{Target-Consistent Action Guidance via Semi-Unbalanced Optimal Transport for Drifting World Models}},
  year = {2026},
  eprint = {2611.24222},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24222},
}
Figure