Imagination-free WAMs, scaled to 14B and DROID-scale data.
WAM-like training, VLA-like inference.
DreamFree uses a video model’s predictive features to predict actions without generating future videos.
DreamFree achieves 11 Hz on a single H200 and a 17.8× inference speedup over DreamZero, while improving real-world success.
Without future-video generation or classifier-free guidance (CFG) at inference, DreamFree can fit on a single A6000 with 48 GB of memory.
Scaling the video backbone from 5B to 14B improves RoboCasa success from 58.5% to 71.7%. Extra videos raise it to 74.4% without additional action labels.
Stand a bottle upright, pour into a pan, pick up bolts, and place a ring around a pole.
Explore the DROID evaluationTrained on DROID only · No task-specific adaptation.
Trained on a filtered 100-hour subset of ABC-130k (≈1/35 of its total duration)
Trained on ABC-130k only · No task-specific adaptation.
Explore the YAM rolloutsGenerate future videos to guide actions. Iterative denoising is slow, and video prediction errors can affect actions.
Trains video denoising and action prediction together, but skips video generation at inference. Denoising alone does not ensure the shared features capture future information.
Distills a frozen video diffusion model’s predictive knowledge into a visual encoder. Predicts actions from the current observation without generating future videos.
DreamFree adds future representation distillation to video generative training, making the video model’s predictive knowledge available in the features used for action prediction. This helps translate larger video backbones into better control; generative training alone shows little benefit from scaling.
An explicit alignment loss makes current-frame features predictive of the future.
At training
A frozen video diffusion teacher observes current and future frames. An alignment loss trains the current-frame features to predict the teacher’s future-aware representations, alongside the original video generative objective. This directly supervises the features used for control and needs no action labels.
Using action-labeled robot demonstrations, a readout module extracts action-relevant features from the visual encoder. An action decoder maps these features to executable action chunks.
Across 24 kitchen tasks, DreamFree reaches 71.7% success, rising to 74.4% with extra video data. Scaling the backbone from 5B to 14B adds 13.2 percentage points, compared with 0.3 for Fast-WAM.
| Method | Family | Success (%) |
|---|---|---|
| π₀ | VLA | 62.5 |
| GR00T-N1.5 | VLA | 64.1 |
| FLARE | VLA | 66.4 |
| Cosmos Policy | WAM | 67.1 |
| DreamFree (14B) | WAM | 71.7 |
| + extra video | WAM | 74.4 |
The base model uses 50 demonstrations per task, as in Cosmos Policy. “+ extra video” adds videos without extra action labels.
| Method | 5B | 14B | Gain |
|---|---|---|---|
| Fast-WAM | 44.4 | 44.7 | +0.3 pp |
| DreamFree | 58.5 | 71.7 | +13.2 pp |
| + extra video | 68.8 | 74.4 | +5.6 pp |
Gain: 5B → 14B within each row.
DreamFree effectively leverages video priors without imagination and benefits from additional video-only data.
We evaluate 10 tasks × 5 trials per policy, using new objects or scene arrangements within familiar task types. All policies run out of the box, without task-specific fine-tuning.
Success measures completed tasks. Task progress gives partial credit using each task’s scoring rubric.
| Method | Success (%) | Progress (%) | Latency (ms/chunk) |
|---|---|---|---|
| π₀.₅ | 16.0 | 33.5 | 81 |
| DreamZero | 56.0 | 71.8 | 1,632 |
| DreamFree | 64.0 | 83.0 | 91 |
Average time to predict an action chunk on H200 GPUs, using each policy’s own settings.
With its 14B video backbone, DreamFree averages 91 ms per action chunk on a single H200. Inference is 17.8× faster than DreamZero, without video step distillation or asynchronous execution.
Full DROID evaluations and a small-data pilot on YAM.
DreamFree learns from video-only demonstrations through its predictive encoder.
Atomic tasks use one skill; composite tasks combine several skills in sequence.
We add composite-task demonstrations to visual encoder pre-training, alongside atomic-task videos. Both models use the same atomic-task action labels to train the action module; no composite-task action labels are used.
Open the microwave → put food inside → close the door → turn it on.
For encoder pre-training
Composite-task demonstration video
Atomic videos + action labels only
Selected failure example
Inference instructionPick the cheese from the counter and place it in the microwave. Then turn on the microwave.
Added during pre-training
Selected success example
Inference instructionPick the cheese from the counter and place it in the microwave. Then turn on the microwave.
Note: The atomic training data typically includes the individual skills used in composite tasks, e.g., “Open the microwave” and “Close the door.”
Our results suggest that these demonstrations help it combine familiar skills into new task sequences, without additional action labels.
“put the orange block in the bowl”
The robot picks the red block instead of the requested orange one. Similar colors can still be confused.
Fold and stack the t-shirts
Rollout 11Erase the whiteboard
Rollout 01The policy completes some tasks with limited training data, but folding clothes and erasing the whiteboard are still unreliable. The t-shirt folding failure may reflect difficulty understanding the cloth’s state.
We have not yet investigated whether longer training would resolve these limitations.
@misc{jo2026dreamfree,
title = {Future Representation Distillation for Imagination-free World Action Models},
author = {Jaehyeong Jo and Sangwon Jang and Sihyun Yu and
Seonghyeon Ye and Yekyung Nah and Hojung Jung and
Kimin Lee and Sung Ju Hwang},
year = {2026},
note = {Manuscript}
}