Robot learning / World action modelsResearch project

DIDO: Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models

Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models

Jing Lyu^*,1,2,3Shuanghao Bai^*,4Runze Xiao³Zhenyu Liao⁶Wenxing Tan³

Zihan Tang⁵Ruochuan Shi^1,2,3Cheng Peng^1,2,3Yuheng Ji^1,2,3Yihao Wang³

Badong Chen⁴Pengwei Wang³Zhongyuan Wang^†,3Xiaoguang Zhao^†,1,2

¹ Institute of Automation, Chinese Academy of Sciences

² School of Artificial Intelligence, University of Chinese Academy of Sciences

³ Beijing Academy of Artificial Intelligence (BAAI)

⁴ Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University

⁵ Tsinghua University⁶ Amazon

^* Equal contribution.^† Corresponding authors.

GitHub arXiv Watch real-world demos

99.0%LIBERO · success rate

76.6%LIBERO-Plus · success rate

92.0%RoboTwin · success rate

384msAction chunk · H100, batch 1

01The observation

Visual structure settles first.
Interaction dynamics follow.

Static backgrounds form early during video denoising. The gripper, object, and their interaction become clear later. Truncating the teacher after one step loses precisely the future dynamics that robot control needs.

One-, two-, and four-step teacher predictions and latent dynamics, compared with DIDO's one-step prediction. — Observation
The four-step teacher progressively resolves interaction regions. DIDO learns to recover these dynamics within a single video denoising step.
Original PDF ↗

Read the full abstract

World Action Models (WAMs) use video generation models to predict future visual dynamics for robotic manipulation, but iterative denoising introduces additional latency for closed-loop control. We empirically find that visual content converges at different rates during denoising. Static background structure forms early, whereas the gripper and manipulated object remain blurry after the first step, with their interaction dynamics emerging only through subsequent denoising. Consequently, naively truncating a multi-step video model to one step preserves scene structure but loses the interaction-centric dynamics most critical for manipulation. To address this issue, we propose DIDO, which distills the converged dynamics of a multi-step video model into a single denoising step. DIDO combines distribution matching distillation with interaction-centric representation guidance. Beyond compressing multi-step generation into one forward pass, DIDO explicitly models the gripper, manipulated object, and their interaction using supervised bounding-box visual reasoning tokens. Additionally, DIDO aligns the target object’s representations across multiple model layers with features from a pretrained DINOv3 encoder. This interaction-centric guidance helps the distilled model preserve both the relevant entities and their future dynamics in a single step, while substantially reducing inference latency. DIDO achieves an average success rate of 99.0% on LIBERO, 76.6% on LIBERO-Plus, and 92.0% on RoboTwin, while also demonstrating effective transfer to long-horizon and generalization tasks in real-world robotic manipulation.

02The method

Preserve interaction dynamics
in one step.

DIDO combines one-step distribution matching with explicit supervision for the object, gripper, and their joint motion. An action expert reads the resulting latent representations directly.

DIDO architecture: object, interaction, gripper and alignment tokens in a one-step Wan video world model, connected to an action expert. — Architecture
Four token groups guide the one-step world model. Shared interaction tokens support both bounding-box prediction branches; DINOv3 supervises object-centric alignment across multiple layers.
Original PDF ↗

I / Distillation

From four steps to one

A robot-domain four-step teacher guides a one-step video generator through distribution matching. Trajectory supervision and multi-layer DINOv3 alignment preserve manipulation-relevant entities and dynamics.

II / Policy learning

Read the latent future

The world model and action expert are optimized jointly, retaining video, interaction, and alignment objectives. The action expert attends to the full latent stream, without decoding and re-encoding future frames.

“One step” refers to the video world model. The flow-matching action expert still uses iterative action generation. Distribution matching is used in Stage I only.

03The evidence

The one-step model
holds up under shift.

DIDO reaches 99.0% on LIBERO and 76.6% on LIBERO-Plus. On RoboTwin, its 92.0% average is the strongest reported result among methods without embodied pretraining.

**LIBERO and RoboTwin**Success rate (%) across suites and conditions
Method	Emb. PT	LIBERO					RoboTwin
Method	Emb. PT	Spatial	Object	Goal	Long	Average	Clean	Rand.	Average
π₀	Yes	96.8	98.8	95.8	85.2	94.1	65.9	58.4	62.2
π₀.₅	Yes	98.8	98.2	98.0	92.4	96.9	82.7	76.8	79.8
LingBot-VA	Yes	98.5	99.6	97.2	98.5	98.5	92.9	91.5	92.2
Motus	Yes	96.8	99.8	96.6	97.6	97.7	88.7	87.0	87.8
Faster-WAM	No	98.4	100.0	97.0	97.8	98.5	89.7	88.6	89.2
Flash-WAM	No	97.0	92.8	96.4	98.0	96.1	88.4	82.7	85.5
Fast-WAM (No co-video)	No	89.2	99.2	95.4	90.0	93.5	82.8	84.8	83.8
Fast-WAM	No	98.2	100.0	97.0	95.2	97.6	91.9	91.8	91.8
DIDO (No co-video)	No	97.8	99.4	97.2	96.4	97.7	89.6	89.0	89.3
DIDO	No	99.8	99.6	98.6	98.0	99.0	92.0	92.0	92.0

**LIBERO-Plus**Success rate (%) across seven perturbations
Method	Emb. PT	Camera	Robot	Language	Lighting	Background	Noise	Layout	Average
OpenVLA-OFT	Yes	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
π₀	Yes	81.4	13.8	58.8	85.0	68.9	6.0	79.0	53.6
X-VLA	Yes	96.0	23.4	75.7	88.2	71.8	89.7	62.7	71.4
Fast-WAM	No	53.7	16.4	68.9	78.2	60.7	44.5	37.7	51.5
ST-WAM	No	55.4	60.1	79.3	93.0	74.2	79.5	74.3	72.8
Faster-WAM	No	57.0	67.9	92.1	94.3	82.7	49.0	82.3	75.0
DIDO	No	45.3	74.3	94.4	96.6	72.7	75.0	83.3	76.6

Simulation results are success rates. “No co-video” removes the video-generation objective during policy learning. Averages follow the manuscript's reported values; on RoboTwin, DIDO is the strongest method without embodied pretraining.

What each component adds

DIDO is built cumulatively from the one-step truncation baseline. Naively truncating the fine-tuned four-step video model to one denoising step achieves 97.7% on LIBERO and 71.4% on LIBERO-Plus.

One-step distillation preserves more of the converged multi-step dynamics within a single step, improving performance to 98.3% and 72.3%.

Adding interaction-centric visual reasoning explicitly models the target object, gripper, and their interaction, raising performance to 98.9% and 75.9%.

Finally, dynamics-based token refinement achieves 99.0% and 76.6% while reducing the conditioning context from 323 to 191 tokens and end-to-end latency by approximately 30 ms. The larger improvement on LIBERO-Plus suggests that selectively compressing near-static regions is particularly beneficial under visual distribution shifts.

Cumulative ablation · Success rate (%)
Configuration	LIBERO	LIBERO-Plus
One-step truncation	97.7	71.4
+ One-step distillation	98.3	72.3
+ Interaction-centric reasoning	98.9	75.9
+ Dynamics-based refinement	99.0	76.6

Why later denoising steps matter

With the same trained four-step video model and action expert, later denoising representations improve LIBERO from 97.7% to 98.4% and LIBERO-Plus from 71.4% to 72.6%.

This experiment studies the teacher before DIDO distillation. Performance generally improves as denoising proceeds, confirming that later steps expose additional action-relevant dynamics.

Success rates improve with later denoising representations from the four-step teacher

Inference cost

On an NVIDIA H100 with batch size 1, DIDO predicts an action chunk in 384 ms end to end, compared with 562 ms for the four-step teacher, yielding a 32% latency reduction while improving LIBERO performance from 98.4% to 99.0%.

Compared with Fast-WAM, which removes future imagination at inference, DIDO introduces only 28 ms of additional latency while improving performance by 1.4 points on LIBERO and 25.1 points on LIBERO-Plus.

End-to-end action chunk prediction

Four-step teacher562 ms

DIDO384 ms

Fast-WAM356 ms

04Beyond simulation

New positions.
More clutter. The same policy.

Four manipulation tasks on Galbot G1, with 150 demonstrations per task. The basic protocol evaluates position generalization; the advanced protocol evaluates both position generalization and environment generalization. Policies are evaluated without additional adaptation.

Real-world demonstrations

4 tasks · 2 settings

Basic

Position generalization

Corn on plate

Drawer and duck

Stain wiping and cloth storage

Bowl and pepper placement

Advanced

Position generalization + Environment generalization

Corn on plate

Drawer and duck

Stain wiping and cloth storage

Bowl and pepper placement

Measure progress through the task

Each rollout receives a staged score on a 10-point scale, normalized to a percentage of the maximum. Evaluation uses 10 rollouts per task under each protocol. These are task-progress scores, rather than binary success rates.

Average normalized task-progress score
Method	Basic	Advanced
Fast-WAM	53.8	33.0
π₀.₅	74.5	53.3
DIDO	80.8	66.0

Reference

DIDO

Read the paper on arXiv or visit DIDO-WAM on GitHub.

@misc{lyu2026didodistillinginteractioncentricdynamics,
      title={DIDO: Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models},
      author={Jing Lyu and Shuanghao Bai and Runze Xiao and Zhenyu Liao and Wenxing Tan and Zihan Tang and Ruochuan Shi and Cheng Peng and Yuheng Ji and Yihao Wang and Badong Chen and Pengwei Wang and Zhongyuan Wang and Xiaoguang Zhao},
      year={2026},
      eprint={2609.15570},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.15570},
}

Visual structure settles first.Interaction dynamics follow.

Preserve interaction dynamicsin one step.