Enfold

Research project · efficient embodied control

Enfold

Folding World-Generator Computation into Predictive Representations for Efficient Embodied Control

A world generator does not need to render a future at every control step. Enfold transfers the computation that constructs a future into a representation inferred from the present.

PaperPreprint coming soonCodeGitHub repositoryVideo12 real-robot demos
97.8%LIBERO
92.02%RoboTwin2.0
49 msEnfold-Flash
10.1×faster than Fast-WAM
training → inference
Enfold training and inference framework. A world generative model supervises a world understanding model during training; only the understanding model and task head are needed for direct control at inference.

Core insight

Internalize imagination. Unfold it only when needed.

Enfold changes the role of a world generator—from an action-time simulator into a training-time source of structured predictive supervision.

01 · observe the future

Expose generative computation

A teacher-forced generator processes the realized future and reveals multi-level hidden states across depth and corruption.

02 · fold into the present

Learn a current-only state

An encoder sees only the current visual context and instruction, and predicts the generator's future-conditioned computation.

03 · act or imagine

Share one predictive interface

Detached task heads read actions directly, while the same representation can condition an optional future rollout.

In one sentence

The generator's future is not repeatedly rendered at deployment; its predictable internal structure has already been enfolded into the present.

Why generator states?

A supervision space, not a single layer.

Pixels preserve every detail. Actions only indirectly describe scene evolution. Generator states expose how appearance, layout, and interaction are organized into a coherent trajectory.

Heat maps and scatter plot comparing predictive utility and noise sensitivity across generator blocks and corruption timesteps.
No block dominates. The most useful layer changes along the corruption trajectory, and semantic utility alone does not guarantee present-time predictability.
Plots showing illumination, layout, and noise sensitivity across generator depth.
Depth changes selectivity. Later states increasingly suppress illumination while retaining layout, but can become more coupled to stochastic generation.
Design consequence

Enfold combines complementary generator depths and uses a timestep-conditioned head to express their shared structure at each corruption level.

Control results

Strong control without generative rollout.

Enfold is evaluated on single-arm simulation, dual-arm simulation, and four real-world bimanual tasks—without embodied pre-training.

LIBERO

Single-arm manipulation
Enfold97.8%
Enfold-Flash97.5%

RoboTwin2.0

Dual-arm, clean + randomized
Enfold91.77%
Enfold-Flash92.02%

Real world

Four bimanual tasks
In distribution89.7%
OOD76.6%

Action latency · A100 40GB

49ms

Enfold-Flash is 10.1× faster than Fast-WAM while retaining 97.5% LIBERO success.

49 ms134 msFast-WAM · 10.1× slower
Four real-robot bimanual tasks and their out-of-distribution variants: store plate, fold towel, organize desktop, and spoon powder.
Real-world evaluation. Four bimanual tasks pair in-distribution rollouts with environment, material, object-set, and object-instance shifts.

Video demonstrations

One task. Three ways the world can change.

Each clip is a single autonomous rollout. The same policy is shown in its nominal setting, under an out-of-distribution shift, and after a human changes the scene during execution.

01

Store Plate

Bimanual object placement

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

02

Fold Towel

Bimanual deformable manipulation

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

03

Organize Desktop

Multi-object rearrangement

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

04

Spoon Powder

Granular-material transfer

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

Videos are muted and loop automatically; playback controls remain available for inspection.

Reimagining after intervention

The scene changes. Imagination and action change with it.

A human alters the scene while the instruction and preceding interaction remain fixed. Enfold re-encodes the new present and revises both its imagined continuation and executed action.

Store plate and fold towel rollouts showing human-induced perturbations followed by revised future imagination.
This coordinated departure from the pre-intervention trajectory is inconsistent with fixed trajectory replay and provides qualitative evidence of state-contingent recovery.

What Enfold encodes

A stable representation of change—not a copy of its teacher.

Predicting stochastic, future-conditioned states from the present acts as a conditional filter: nuisance variation is suppressed while interaction-relevant structure is retained.

7.9–10.4×lower nuisance sensitivity than generator states
31.8effective rank, arguing against feature collapse
18.2%lower long-horizon error on most-changing regions
Comparison of nuisance sensitivity, task retrieval, and effective rank for Cosmos generator states and the Enfold representation.
The student suppresses teacher nuisance while preserving task-selective, non-degenerate structure.
Future-change probe showing moving robot and object regions alongside prediction results across three future horizons.
Enfold's advantage grows with temporal distance and is largest on regions that actually change.
Query-point similarity overlays comparing frozen DINO and Enfold representations on real and simulated interactions.
Enfold connects the gripper to manipulated objects and contact-relevant regions, rather than responding only to visual similarity.

Resources

Code and videos are live. The paper will follow.

Explore the project repository or watch the real-robot rollouts above. Paper and citation details will be added with the preprint.

PaperPreprint coming soon
CodeGitHub repositoryVideo12 real-robot demos
Enfold — Predictive Representations for Efficient Embodied Control
Enfold

Research project · efficient embodied control

Enfold

Folding World-Generator Computation into Predictive Representations for Efficient Embodied Control

A world generator does not need to render a future at every control step. Enfold transfers the computation that constructs a future into a representation inferred from the present.

PaperPreprint coming soonCodeRelease coming soonVideo12 real-robot demos
97.8%LIBERO
92.02%RoboTwin2.0
49 msEnfold-Flash
10.1×faster than Fast-WAM
training → inference
Enfold training and inference framework. A world generative model supervises a world understanding model during training; only the understanding model and task head are needed for direct control at inference.

Core insight

Internalize imagination. Unfold it only when needed.

Enfold changes the role of a world generator—from an action-time simulator into a training-time source of structured predictive supervision.

01 · observe the future

Expose generative computation

A teacher-forced generator processes the realized future and reveals multi-level hidden states across depth and corruption.

02 · fold into the present

Learn a current-only state

An encoder sees only the current visual context and instruction, and predicts the generator's future-conditioned computation.

03 · act or imagine

Share one predictive interface

Detached task heads read actions directly, while the same representation can condition an optional future rollout.

In one sentence

The generator's future is not repeatedly rendered at deployment; its predictable internal structure has already been enfolded into the present.

Why generator states?

A supervision space, not a single layer.

Pixels preserve every detail. Actions only indirectly describe scene evolution. Generator states expose how appearance, layout, and interaction are organized into a coherent trajectory.

Heat maps and scatter plot comparing predictive utility and noise sensitivity across generator blocks and corruption timesteps.
No block dominates. The most useful layer changes along the corruption trajectory, and semantic utility alone does not guarantee present-time predictability.
Plots showing illumination, layout, and noise sensitivity across generator depth.
Depth changes selectivity. Later states increasingly suppress illumination while retaining layout, but can become more coupled to stochastic generation.
Design consequence

Enfold combines complementary generator depths and uses a timestep-conditioned head to express their shared structure at each corruption level.

Control results

Strong control without generative rollout.

Enfold is evaluated on single-arm simulation, dual-arm simulation, and four real-world bimanual tasks—without embodied pre-training.

LIBERO

Single-arm manipulation
Enfold97.8%
Enfold-Flash97.5%

RoboTwin2.0

Dual-arm, clean + randomized
Enfold91.77%
Enfold-Flash92.02%

Real world

Four bimanual tasks
In distribution89.7%
OOD76.6%

Action latency · A100 40GB

49ms

Enfold-Flash is 10.1× faster than Fast-WAM while retaining 97.5% LIBERO success.

49 ms134 msFast-WAM · 10.1× slower
Four real-robot bimanual tasks and their out-of-distribution variants: store plate, fold towel, organize desktop, and spoon powder.
Real-world evaluation. Four bimanual tasks pair in-distribution rollouts with environment, material, object-set, and object-instance shifts.

Video demonstrations

One task. Three ways the world can change.

Each clip is a single autonomous rollout. The same policy is shown in its nominal setting, under an out-of-distribution shift, and after a human changes the scene during execution.

01

Store Plate

Bimanual object placement

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

02

Fold Towel

Bimanual deformable manipulation

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

03

Organize Desktop

Multi-object rearrangement

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

04

Spoon Powder

Granular-material transfer

In distribution

Nominal setting

OOD

Appearance or object shift

Perturbation

Scene changed during execution

Videos are muted and loop automatically; playback controls remain available for inspection.

Reimagining after intervention

The scene changes. Imagination and action change with it.

A human alters the scene while the instruction and preceding interaction remain fixed. Enfold re-encodes the new present and revises both its imagined continuation and executed action.

Store plate and fold towel rollouts showing human-induced perturbations followed by revised future imagination.
This coordinated departure from the pre-intervention trajectory is inconsistent with fixed trajectory replay and provides qualitative evidence of state-contingent recovery.

What Enfold encodes

A stable representation of change—not a copy of its teacher.

Predicting stochastic, future-conditioned states from the present acts as a conditional filter: nuisance variation is suppressed while interaction-relevant structure is retained.

7.9–10.4×lower nuisance sensitivity than generator states
31.8effective rank, arguing against feature collapse
18.2%lower long-horizon error on most-changing regions
Comparison of nuisance sensitivity, task retrieval, and effective rank for Cosmos generator states and the Enfold representation.
The student suppresses teacher nuisance while preserving task-selective, non-degenerate structure.
Future-change probe showing moving robot and object regions alongside prediction results across three future horizons.
Enfold's advantage grows with temporal distance and is largest on regions that actually change.
Query-point similarity overlays comparing frozen DINO and Enfold representations on real and simulated interactions.
Enfold connects the gripper to manipulated objects and contact-relevant regions, rather than responding only to visual similarity.

Resources

Videos are live. Paper and code will follow.

Real-robot rollouts are available above. Paper, code, and citation details will be updated when the preprint is available.

PaperPreprint coming soon
CodeRelease coming soon
Video12 real-robot demos