Diffusion ReRoll: Revisable Denoising for Robotic Sequential Prediction

Diffusion ReRoll

Revisable Denoising for Robotic Sequential Prediction

A diffusion process that respects the temporal meaning of a<br>sequence

and preserves cross-horizon revisability .

Seonsoo Kim<br>Seongil Hong<br>Jun-Gill Kang

Agency for Defense Development (ADD)

arXiv

Paper

Code

Video

Project video

Diffusion ReRoll at a glance

Your browser does not support embedded video.

Full project video · 2 minutes 57 seconds

Overview · Three denoising processes

The schedule changes how information moves.

The same sequential prediction problem produces very different behavior depending on<br>which horizon regions remain uncertain, resolve first, or become revisable again.

01 · Equal noise level

Full-sequence diffusion

Every sequential timestep denoises together, without an explicit temporal order.

02 · Left-to-right order

Causal denoising

Earlier timesteps resolve first, giving the diffusion process temporal direction.

03 · Revisable denoising waves

Diffusion ReRoll

Periodic re-noising reopens resolved regions so updated context can revise them.

Visualization: predicted clean trajectories x̂0 are shown in<br>all three animations.

Robotics scope

One principle across robotic sequential prediction

ReRoll is evaluated beyond maze generation: it refines planned states, robot action<br>chunks, and coupled video–action sequences.

Offline RL benchmark

Long-horizon planning

Goal guidance and endpoint inpainting on OGBench PointMaze and AntMaze.

Guidance · Inpainting

Diffusion Policy based

Robot action generation

LIBERO-10 multi-task and RoboCasa single-task learning with different prediction<br>horizons and observation-history lengths.

LIBERO-10 · RoboCasa

UWM based

Unified video–action generation

Policy, inverse dynamics, future-video prediction, and action–video alignment.

Video · Action · Alignment

Core perspective

Sequential tokens have temporal meaning.

Earlier and later states, actions, and frames are not interchangeable. Their noise<br>levels determine what is visible, uncertain, and still modifiable during generation.

ReRoll turns this noise schedule into a cross-horizon revision mechanism: resolved<br>regions can become uncertain again, exchange information with the rest of the horizon,<br>and settle into a more compatible sequence.

Temporal meaning<br>Uncertainty control<br>Cross-horizon revision

Characteristic mistakes

What fails when the horizon cannot revise?

The problem is not simply whether denoising is global or causal. The crucial question is<br>whether early decisions can still change after new cross-horizon context becomes useful.

Full-sequence diffusion

Early global errors get committed.

Treating all timesteps equally provides global interaction, but does not express<br>which regions should stay uncertain for later correction.

Causal denoising

Temporal order can lock a wrong prefix.

Causal noise levels add temporal meaning, but low-noise prefixes become difficult to<br>revise when later predictions reveal a mismatch.

Diffusion ReRoll

An early mistake does not have to be final.

A ReRoll event re-noises the affected region. The model explores another compatible<br>candidate using updated context, naturally revising the earlier error.

Early error→Re-noise→Revise

Abstract

Iterative revision across the prediction horizon

Diffusion-based sequence predictors commonly follow a single, irreversible denoising<br>process. Diffusion ReRoll instead selectively re-noises sufficiently resolved regions<br>while the rest of the horizon continues denoising. These regions can then be refined<br>again using updated context from the sequence. The resulting structured re-noising<br>enables cross-horizon revision while preserving local consistency. Across long-horizon<br>planning, diffusion-policy-style action prediction, and unified video–action modeling,<br>ReRoll improves performance over full-sequence and causal denoising baselines.

Method

A programmable refinement schedule

ReRoll changes the token-wise noise schedule—not the underlying denoising<br>architecture—to control when information can be revised across a sequence.

Deployment schedule

Piecewise-linear denoising with periodic noise injection.

Linear noise-level chunks move across the prediction horizon. At the reset level, a<br>ReRoll event raises uncertainty in a resolved region, connects it to the next<br>denoising wave, and makes that region revisable again.

Training schedule

Train on the corresponding linear noise patterns.

Diffusion Forcing introduced the noise-as-masking view by assigning different noise<br>levels to different sequence positions. ReRoll builds on that idea with randomized<br>piecewise-linear assignments, including reversed chunks for bidirectional generation.

01

Noise as masking

This idea comes from Diffusion Forcing: a token's noise level acts<br>like a soft mask, controlling how much information is visible and how freely the<br>token can still change.

02

Linear-chunk...