DAWN

DAWN: Worlds Teach, Actions ExecuteDecoupled World-to-Action Distillation
for Robot Manipulation

Paper
Accelerated 12-second overviews, grouped by robot platform.

Motivation

A world model should teach during training, while an action model executes independently at inference.

Future visual predictions contain useful physical dynamics, but generating them at every control step adds computational cost. DAWN transfers the world model’s knowledge into the action policy during training, so deployment requires only the action model.

Comparison of World Action Models, efficient WAMs, and DAWN’s asymmetric training and inference paradigm
DAWN separates future-dynamics supervision from action execution.

Method

A frozen world model provides future structural features. Dynamics-appearance disentangled distillation encourages the action model to absorb manipulation-relevant dynamics while the auxiliary pathway handles appearance.

DAWN’s action model, frozen world-model teacher, appearance generator and feature decoder
The world model, appearance generator, and feature decoder are used during training. At inference, only the action model remains.

Dynamics-appearance disentanglement

World-model features mix object motion, gripper changes, and contact transitions with textures, lighting, and backgrounds. An appearance pathway models visual detail, encouraging the action representation to encode the complementary dynamics relevant to control.

The teacher uses LTX-Video; the standalone policy uses a Qwen3-VL backbone and an MLP action head. The appearance pathway is initialized from InstructPix2Pix.

Comparison of entangled supervision and dynamics-focused learning, with action features Q and an auxiliary appearance representation
Blue tokens denote action features Q; the yellow block denotes the appearance representation v. Solid arrows show decoder inputs and reconstruction. Dashed arrows indicate schematic learning pressure, not separate losses.

Real-world experiments

Dual-arm UR5e and dual-arm ARX ACONE platforms with head-mounted and wrist-mounted RealSense cameras
Dual-arm UR5e (left) and dual-arm ARX ACONE (right). Each platform uses one head-mounted RGB camera and two wrist-mounted RGB cameras.

Task demonstrations

Task 1 / 6

Insert the drill

Insert the drill into the holder.

Rollouts are shown at 2× speed.

1 / 6

Stack-blocks ablation

Representative trials without world-model distillation, without disentanglement, and with the full DAWN model.

The same additional 3× acceleration is applied to all clips. Shorter clips hold their final frame.

Failure case: fold the towel

Failure analysis

The policy grasps and moves the towel, but does not complete the final fold. This illustrates a remaining failure mode in deformable-object manipulation.

Experimental results

Evaluation on LIBERO, LIBERO-Plus, RoboTwin 2.0, and six real-world manipulation tasks.

Across simulation benchmarks

DAWN variants compared across three simulation benchmarks
MethodLIBEROLIBERO-PlusRoboTwin 2.0
DAWN ACTION ONLY97.278.771.3
DAWN w/ WM inference97.378.571.5
DAWN w/o WM distillation96.675.457.1

LIBERO: 40 tasks, 50 episodes per task. LIBERO-Plus: seven perturbation dimensions without fine-tuning. RoboTwin 2.0: ten jointly trained tasks in the clean setting.

Real-world comparison

Six tasks, evaluated over 20 trials each. All methods use 300 demonstrations per task.

Real-world success counts out of twenty trials
TaskGR00T N1.5VPPDAWN
Insert drill10/206/2011/20
Hang towel13/208/2015/20
Chain & close box16/2010/2016/20
Fold cloth14/207/2013/20
Fold towel11/205/2014/20
Stack blocks18/2015/2018/20
Average success68.3%42.5%72.5%

Real-world ablation

Stack the Blocks: 300 demonstrations and 20 trials with blocks placed at extreme positions.

8/20

Without world-model
distillation

15/20

Without
disentanglement

18/20

Full
DAWN