LoopedWAM: Looped World Action Models for Elastic Inference

Store 12 blocks, run 30 layers, and choose the depth at test time.

Anh Dao1 Xinxin Liu1 Hang Yu1 Jeffrey A. Chan Santiago1 Yuzhang Shang1

Paper coming soon Code coming soon Models coming soon Robot videos
"pick up the can and place it on the plate"

LoopedWAM controlling a real AgileX PiPER arm from a front and a wrist camera. Front view shown; uncut episodes from one evaluation session.

0.58Bpolicy parameters, 2.4× fewer than an equally deep dense model
91.9%RoboTwin 2.0, on par with the ~6B Fast-WAM (91.8%)
24.7–86.2 msnine latency budgets from one LoopedWAM-E checkpoint
7.4×less peak GPU memory than Fast-WAM (2.05 vs 15.25 GB)

Abstract

Compact world action models with depth you can choose at test time

World action models (WAMs) couple generative video modeling with robotic action prediction but typically require large video and action transformers, limiting deployment efficiency. We present LoopedWAM, a compact WAM that recurrently applies a shared six-block core in each expert, executing 30 transformer layers while storing only 12 blocks per expert. To account for their asymmetric inference costs, the video and action experts can operate at different depths: visual features are computed once per query, whereas action computation repeats at every sampling step. We further propose elastic self-distillation, which trains a single elastic checkpoint, LoopedWAM-E, across action-depth and sampling-step budgets. With only 0.58B policy parameters and no robot pretraining, LoopedWAM achieves 96.1% success on LIBERO, 91.9% on RoboTwin 2.0, and 52.2% on RoboCasa-GR1, while using 7.4× less peak GPU memory than Fast-WAM under matched inference measurements. On LIBERO, LoopedWAM-E achieves 96.0% success in 27.6 ms or 95.2% in 24.7 ms on an H100, providing multiple accuracy–latency operating points from a single checkpoint; with two sampling steps and four action loops, it stays within 0.5 points of LoopedWAM on all three benchmarks. Ablations demonstrate that weight-shared depth preserves competitive performance, while aggressive action-depth reduction requires explicit budget-aware training. On a real AgileX PiPER arm, LoopedWAM completes four manipulation tasks from 100 demonstrations each.

Three panels: (a) stored versus executed layers, (b) RoboTwin 2.0 success versus parameters, (c) LIBERO success versus latency
LoopedWAM is small, fast, and elastic. (a) Each expert stores 12 blocks and loops its shared 6-block core K times, reaching the 30-layer depth of its dense donor with 2.4× fewer parameters. (b) RoboTwin 2.0 success against parameters, as reported by each paper. (c) LIBERO success against latency on one H100 for every setting of LoopedWAM-E; at (Ka, N) = (2, 2) it is 3.9× faster than Fast-WAM.

Method

Separate what is stored from what is executed

A query runs the video expert once and the action expert once per sampling step. Latency follows the number of blocks executed in sequence; memory follows the number of blocks stored. LoopedWAM controls the two independently.

Method overview: looped video and action experts with late alignment, and elastic self-distillation from a frozen teacher
(a) LoopedWAM. Each expert stores a prelude, a shared core, and a coda, and loops the core Kv times in the video expert and Ka times in the action expert. The video expert encodes the current frame once per query and caches its keys and values; late alignment lets every action loop read the deepest visual features. (b) LoopedWAM-E. One student learns every setting (Ka, N); each macro step is matched to the frozen teacher's fine Euler map started from the student's own state.
1

Looped experts

A 3-block prelude, a 6-block shared core, and a 3-block coda, copied from a Wan2.1 video model. Looping the core four times recovers the donor's 30 layers with 12 stored blocks: 0.58B parameters in total.

2

Decoupled depth, late alignment

Video depth is paid once per query, action depth at every sampling step, so the two experts loop a different number of times. Action loop r attends to video loop σ(r) = max(1, Kv−Ka+r), so even one action loop sees the deepest visual features.

3

Elastic self-distillation

One student, LoopedWAM-E, is trained for Ka ∈ {1, 2, 4} action loops and N ∈ {2, 5, 10} sampling steps. Targets come from the teacher started at the student's own intermediate states, so the student learns to correct its own errors.

Cost of one policy query

T(Kv, Ka, N) ≈ T0 + L(Kv)·tv + N·L(Ka)·ta
L(K) = 3 + 6K + 3
T0 fixed cost: encoding and I/O
Video depth paid once per query
Action depth paid at each of the N sampling steps

So LoopedWAM-E keeps the video expert at full depth (Kv = 4) and makes only the action loops Ka and the steps N adjustable.

Elastic inference

One checkpoint, nine operating points

Pick the number of action loops and sampling steps at deployment. Every setting below is the same LoopedWAM-E weights, evaluated on LIBERO (6,000 episodes each) with latency on one H100 at batch size 1.

LIBERO success versus latency Nine LoopedWAM-E settings span 24.7 to 86.2 milliseconds at 95.2 to 96.2 percent success; Fast-WAM is at 107.5 milliseconds and 96.6 percent.
96.0%LIBERO success
27.6 mslatency per query
36action-block calls per query
3.9×faster than Fast-WAM*

Show all nine settings as a table
KaNBlock callsLatency (ms)Mem. (GB)SpatialObjectGoalLongAvg.

*Fast-WAM re-evaluated by us with the released checkpoint under the same protocol.

Three plots: success versus action loops, one checkpoint versus fixed-depth models, latency versus executed blocks
Depth must be trained to be removable (LIBERO, shorter ablation schedule). (a) Running the fixed 4/4 model with fewer action loops collapses its success (12.3% at one loop), while the elastic student holds every loop count, including ones it never trained on. (b) The elastic checkpoint dominates every separately trained fixed-depth model and Dense-12. (c) Latency is linear in executed action blocks (R2 = 0.999), with a 19.8 ms per-query floor.
+81.4 pts

Elastic training rescues one-loop inference

Truncating the trained model to a single action loop gives 12.3% on LIBERO; the elastic student keeps 93.7% at the same setting.

60 vs 60

At equal compute, depth beats steps

With 60 action-block calls each, (Ka, N) = (4, 2) beats (1, 5) by 0.4, 2.1, and 5.0 points on LIBERO, RoboTwin 2.0, and RoboCasa-GR1.

≤0.5 pts

One recipe, three benchmarks

At two steps and four action loops, a fifth of the teacher's action-block calls, the student stays within 0.5 points of its teacher on all three benchmarks.

87%

Looping recovers most of the benefit of depth

Looping 12 stored blocks to 30 layers closes 87% of the LIBERO gap between Dense-12 (90.0%) and Dense-30 (95.6%), without adding parameters.

Benchmarks

Competitive with world action models ten times larger

No robot pretraining. Prior results are taken from their papers unless marked with *, which we re-evaluated with the released checkpoints under our protocol.

MethodParamsRobot PTSpatialObjectGoalLongAvg.
Vision-language-action models
OpenVLA-OFT7B✓97.698.497.994.597.1
π03.3B✓96.898.895.885.294.2
π0.53B✓98.898.298.092.496.9
GR00T-N1.63.3B✓97.798.597.594.497.0
SmolVLA0.45B✗90.096.092.071.087.3
RD-VLA0.5B–92.099.096.084.893.0
World action models
Motus8B✓96.899.896.697.697.7
LingBot-VA5.3B✓98.599.697.298.598.5
Cosmos Policy2B✗98.1100.098.297.698.5
LaWAM2.3B✓99.499.698.497.098.6
LiLa-WAM0.5B✗98.098.897.294.297.1
Light-WAM*1.9B✗96.798.996.690.995.8
Fast-WAM*~6B✗97.599.396.393.596.6
Faster-WAM*~6.3B✗99.099.698.298.198.7
Ours
LoopedWAM0.58B✗96.999.597.990.196.1
LoopedWAM-E (Ka=4, N=2)0.58B✗97.299.397.588.595.6
LoopedWAM-E (Ka=2, N=2)0.58B✗97.899.197.889.296.0

Our models and * rows: 40 tasks × 50 initial states × 3 seeds = 6,000 episodes per model. Params exclude the frozen VAE and text encoder. Robot PT: pretraining on robot data.

MethodParamsRobot PTCleanRandomizedAvg.
Vision-language-action models
π03.3B✓65.958.462.2
π0.53B✓82.776.879.8
X-VLA0.9B✓72.872.872.8
World action models
Motus8B✓88.787.087.8
LingBot-VA5.3B✓92.991.692.2
AHA-WAM7.2B✗93.492.292.8
Fast-WAM~6B✗91.991.891.8
Faster-WAM~6.3B✗92.892.392.6
LaWAM2.3B✓92.689.891.2
Efficient-WAM1B✗86.785.786.2
Light-WAM1.9B✗76.476.376.4
LiLa-WAM0.5B✗90.589.089.8
Ours
LoopedWAM0.58B✗92.191.791.9
LoopedWAM-E (Ka=4, N=2)0.58B✗91.391.691.5
LoopedWAM-E (Ka=2, N=2)0.58B✗90.690.690.6

50 bimanual tasks, one model, 100 episodes per task in each setting. LoopedWAM is the best model below 2B parameters.

MethodParamsAvg. success
GR00T-N1.63.3B47.6
StarVLA-GR00T4B47.8
Being-H0.73B49.2
THAW-VLA0.8B50.5
DiT4DiT–50.8
VP-VLA2×4B53.8
TwinBrainVLA2×4B54.6
LDA-1B1.6B55.4
ABot-M04B58.3
ACE-Ego-04B72.8
Ours
LoopedWAM0.58B52.2
LoopedWAM-E (Ka=4, N=2)0.58B52.4
LoopedWAM-E (Ka=2, N=2)0.58B51.6

24 tabletop tasks for a GR1 humanoid (29-D actions), 50 rollouts per task for our models. Rollout counts of prior methods follow their papers (THAW-VLA: 20; LDA-1B: 51; GR00T-N1.6: not stated).

MethodParamsKaNMem. (GB)Latency (ms)Speed-upLIBERORoboTwin
Prior world action models
Fast-WAM*~6B–1015.25107.51.0×96.691.8
Fast-WAM*~6B–215.3237.12.9×96.090.1
Faster-WAM*~6.3B–1012.7672.61.5×98.792.6
Light-WAM*1.9B––4.3821.84.9×95.876.4
Dense controls (same Wan2.1 donor)
Dense-301.42B–103.7990.51.2×96.3–
Dense-30-E1.42B–24.3333.93.2×95.9–
Dense-120.58B–102.0940.82.6×94.0–
Ours
LoopedWAM0.58B4102.0591.81.2×96.191.9
LoopedWAM0.58B422.0935.03.1×95.391.2
LoopedWAM-E0.58B422.7833.23.2×95.691.5
LoopedWAM-E0.58B222.6827.63.9×96.090.6
LoopedWAM-E0.58B122.5824.74.4×95.289.3

One shared harness: one H100, batch 1, mean over 500 queries, from raw observation to actions. Mem.: peak GPU memory (VAE included, text encoder not). Speed-up relative to Fast-WAM* at its native 10 steps. Light-WAM predicts the chunk in one forward pass.

ModelKvKaParamsLIBERORoboTwin-10
Dense-12 (12 layers)––0.58B90.077.3
LoopedWAM140.58B91.779.6
LoopedWAM220.58B92.578.9
LoopedWAM410.58B94.081.0
LoopedWAM420.58B94.582.4
LoopedWAM (default)440.58B94.981.4
Dense-30 (30 layers)––1.42B95.683.6

Shorter ablation schedule (10 LIBERO epochs). At equal parameters, video loops (4/1) beat action loops (1/4) by 2.3 points at 1.8× lower latency: video depth is paid once per query.

Real world

Four tasks on a real AgileX PiPER arm

A 6-DoF PiPER with a parallel gripper, observed by a front and a wrist RealSense D435i. We collect 100 teleoperated demonstrations per task and train LoopedWAM from the same Wan2.1 donor. The policy outputs 32 absolute joint targets per query at 15 Hz.

Real-robot setup with leader and follower arms, and start and end states of the four tasks
Setup and tasks. (a) Follower arm with front and wrist cameras; demonstrations come from teleoperation with a leader arm. (b–e) Start and end states of the four tasks.

Language-conditioned pressing

Same policy, same table: the instruction alone decides which button the arm presses. Played at 2× speed.

"press the blue button" "press the red button" "press the pink button" "press the orange button"

Rollouts from both cameras

Uncut successful episodes of LoopedWAM (Kv=Ka=4, N=10), front view left and wrist view right, played at 2× speed.

Front
Wrist
"pick up the can and place it on the plate"
Front
Wrist
"pick up the can and place it on the plate"
Front
Wrist
"stack the bowls"
Front
Wrist
"stack the bowls"
Front
Wrist
"put the spoon in the bowl"
Front
Wrist
"put the spoon in the bowl"

Success, LoopedWAM

TaskSuccessRate
Place can on plate17 / 2470.8%
Press color button20 / 2290.9%
Stack bowls11 / 1668.8%
Put spoon in bowl12 / 2157.1%
All60 / 8372.3%

One evaluation session with randomized object placements; every trial starts from the same home pose.

Cost on one GPU

ModelLatencyPeak mem.
Fast-WAM* (~6B)150 ms12.7 GiB
LoopedWAM160 ms3.7 GiB
LoopedWAM-E (4, 2)58 ms3.3 GiB
LoopedWAM-E (2, 2)46 ms3.3 GiB

Full robot query (preprocessing and 32-action chunk) on one RTX PRO 6000, batch 1. *Fine-tuned by us on the same demonstrations.

Citation

BibTeX

@article{dao2026loopedwam,
  title   = {LoopedWAM: Looped World Action Models for Elastic Inference},
  author  = {Dao, Anh and Liu, Xinxin and Yu, Hang and
             Chan Santiago, Jeffrey A. and Shang, Yuzhang},
  year    = {2026}
}