Diffusion ReRoll

Revisable Denoising for Robotic Sequential Prediction

A diffusion process that respects the temporal meaning of a sequence
and preserves cross-horizon revisability.

Seonsoo Kim Seongil Hong Jun-Gill Kang

Agency for Defense Development (ADD)

arXiv Paper Code Video

Project video

Diffusion ReRoll at a glance

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 which horizon regions remain uncertain, resolve first, or become revisable again.

Animated full-sequence diffusion trajectory denoising

01 · Equal noise level

Full-sequence diffusion

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

Animated Diffusion Forcing causal trajectory denoising

02 · Left-to-right order

Causal denoising

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

Animated Diffusion ReRoll trajectory denoising and revision

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 all three animations.

Robotics scope

One principle across robotic sequential prediction

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

Long-horizon maze planning trajectory

Offline RL benchmark

Long-horizon planning

Goal guidance and endpoint inpainting on OGBench PointMaze and AntMaze.

Guidance · Inpainting
Robot manipulation scene for diffusion-policy-style action prediction

Diffusion Policy based

Robot action generation

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

LIBERO-10 · RoboCasa
Generated robot video frames paired with a generated end-effector path

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 levels determine what is visible, uncertain, and still modifiable during generation.

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

Temporal meaning Uncertainty control 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 whether early decisions can still change after new cross-horizon context becomes useful.

Full-sequence diffusion committing errors, missing the goal, and passing through maze walls

Full-sequence diffusion

Early global errors get committed.

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

Causal denoising locking an erroneous trajectory prefix before later context is resolved

Causal denoising

Temporal order can lock a wrong prefix.

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

Diffusion ReRoll exploring revisable candidates and correcting an early denoising error

Diffusion ReRoll

An early mistake does not have to be final.

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

Early errorRe-noiseRevise

Abstract

Iterative revision across the prediction horizon

Diffusion-based sequence predictors commonly follow a single, irreversible denoising process. Diffusion ReRoll instead selectively re-noises sufficiently resolved regions while the rest of the horizon continues denoising. These regions can then be refined again using updated context from the sequence. The resulting structured re-noising enables cross-horizon revision while preserving local consistency. Across long-horizon planning, diffusion-policy-style action prediction, and unified video–action modeling, 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 architecture—to control when information can be revised across a sequence.

Diffusion ReRoll deployment schedule with piecewise-linear noise levels and periodic noise injection

Deployment schedule

Piecewise-linear denoising with periodic noise injection.

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

Training Diffusion ReRoll with corresponding linear noise-level assignments, including bidirectional schedules

Training schedule

Train on the corresponding linear noise patterns.

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

  1. 01

    Noise as masking

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

  2. 02

    Linear-chunk training

    Randomized piecewise-linear profiles expose the model to the partial-information patterns used at deployment.

  3. 03

    ReRoll and revise

    Periodic noise injection reopens resolved regions so updated context can correct them across the horizon.

Key idea

Noise is not only removed—it is actively managed as a mechanism for revising sequential predictions.

Bidirectional mode

Let both endpoints guide revision.

When both initial and terminal constraints are available, forward and backward denoising waves begin at opposite ends of the horizon. Their meeting point is adjustable, and each side can use its own slope, reset noise level, and number of ReRoll events. Unresolved regions remain noisy, so information from both endpoints can continue revising the sequence.

Initial condition Revisable meeting region Terminal condition

From the project video

Two denoising fronts, one revisable horizon.

This clip shows the bidirectional schedule in motion, followed by endpoint-conditioned examples in goal inpainting and inverse dynamics.

Goal inpainting Future-image conditioning Inverse dynamics
Comparison of taking-over and meeting bidirectional schedule matrices

Matrix behavior

Choose how long backward context remains active.

Taking over uses terminal-side context early and then returns to forward revision. Meeting preserves backward-driven context throughout ReRoll.

Taking-over matrix

Use terminal context early, finish causally.

Backward waves bring goal-side information into early revision. After the two sides overlap, forward waves take over, preserving a more causal final prediction structure.

Used for goal-inpainting maze planning

Meeting matrix

Keep both directions active.

Forward and backward waves meet and repeatedly refine the sequence from both sides, preserving terminal-to-initial influence throughout ReRoll.

Used for unified video–action modeling

Results

One revision principle, three robotic settings

Long-horizon planning OGBench
72.8 88.1%

Guidance-based replanning success

+21% relative over Diffusion Forcing
Policy learning LIBERO-10
40.1 62.7%

Average official-initialization success

+57% relative over Diffusion Policy
Video–action modeling UWM
57.2 71.0%

OOD joint-policy success

Video–action alignment error: 0.324 m → 0.301 m

Ablation · Noise injection

ReRoll events drive repeated improvement.

With zero ReRoll events, re-noising is removed and the schedule reduces to causal denoising. Adding selective noise-injection events improves long-horizon planning, policy learning, and unified video–action modeling before performance saturates, isolating repeated revision as a central source of the gains.

Plots showing performance as the number of ReRoll events increases for long-horizon planning, policy learning, and unified video-action modeling

Swipe horizontally to inspect all three settings →

0 eventsCausal denoising baseline
More eventsMore opportunities to revise
At saturationExtra resets add little benefit

Evaluation coverage

Experiments across planning, policy learning, and joint prediction

01

OGBench

Guidance-based planning and goal inpainting

Long-horizon PointMaze and AntMaze tasks test both guided generation and terminal conditioning.

02

LIBERO-10

Multi-task policy learning

A Diffusion Policy-based model predicts action chunks across ten manipulation tasks.

03

RoboCasa

Single-task learning across temporal contexts

Different prediction horizons and observation-history lengths test robustness to sequence design.

04

Video–action modeling

Joint generation and action–video alignment

UWM-based experiments evaluate future video, action prediction, inverse dynamics, and their alignment.

Conclusion

Keep uncertainty where revision still matters.

Diffusion ReRoll treats the noise schedule as a mechanism for information flow across time. By periodically restoring uncertainty, it preserves temporal structure without permanently locking early predictions.

The same principle improves long-horizon planning, manipulation-policy learning, and coupled video–action generation—providing a general route to revisable robotic sequential prediction.

BibTeX

Cite Diffusion ReRoll

@misc{kim2026diffusionreroll,
  title   = {Diffusion ReRoll: Revisable Denoising for
             Robotic Sequential Prediction},
  author  = {Kim, Seonsoo and Hong, Seongil and Kang, Jun-Gill},
  year    = {2026}
}