World-action models (WAMs) couple action generation with predictions of future states. An initial action chunk may be flawed. A future prediction that is not conditioned on the chunk may fail to reflect the changes that the proposed actions would induce in visual representations, limiting its usefulness for refinement before execution. We introduce ThinkWAM, a latent WAM that establishes bidirectional feedback before execution through a proposal-rethink-refine loop: the policy first proposes a complete action chunk, the proposal corrects multi-horizon future latents through a learned residual, and the corrected latents are injected into a second pass of the same action head, which generates the executable chunk. An affordance-guided observation-action memory and a near-fine, far-coarse multi-horizon prediction scheme, both motivated by predictive coding, support this loop, and all prediction stays in latent space. ThinkWAM achieves 98.70% success on LIBERO, 84.76% on LIBERO-Plus without OOD fine-tuning, 85.75%/65.85% on RoboTwin 2.0 Clean/Random, and 65.17% on RoboCasa24. ThinkWAM achieves 64.4% mean success across three real-world OOD conditions. The mean two-pass inference latency per action chunk is 153.48 ms on LIBERO, 164.61 ms on RoboTwin, and 101.2 ms on the Galaxea A1Z. Ablations show complementary gains from rethinking and affordance guidance. Diagnostics indicate that proposal-conditioned corrections change both predicted futures and action trajectories, showing that the corrected future representations improve the final action.
Figures & tables
Figure 2: Architecture of ThinkWAM. The temporal JEPA predicts base future latents from the visual prefix and memory, and the affordance decoder adds a memory residual. The proposal At(0) from the first ActionDiT pass conditions the residual predictor, and the corrected futures condition the second pass, which outputs At(1) .
Method
Size (B)
Action P.T.
Spatial
Object
Goal
Long
Avg.
# Non-JEPA methods
π0.5 [ 6 ]
3.3
✓
98.8
98.2
98.0
92.4
96.9
ABot-M0 [ 50 ]
4.6
✓
98.8
99.8
99.0
96.6
98.6
ResVLA [ 64 ]
4.0
✗
96.0
100.0
97.4
92.8
96.6
FastWAM [ 52 ]
6.0
✗
98.2
100.0
97.0
95.2
97.6
ImageWAM [ 59 ]
4.5
✗
97.2
99.2
98.8
98.4
98.4
Table 1: LIBERO success (%) across four suites, with model size in billions (B) and action-supervised policy pretraining (Action P.T.).
Method
Camera
Init.
Language
Light
Texture
Noise
Layout
Avg.
# Non-JEPA methods
π0.5 [ 6 ]
64.0
58.0
88.5
96.6
81.4
87.5
85.9
80.3
ABot-M0 [ 50 ]
60.4
67.9
86.4
96.2
91.6
86.4
82.6
81.6
ResVLA [ 64 ]
53.2
57.5
88.2
94.5
96.3
81.8
78.3
78.5
FastWAM [ 52 ]
16.4
44.5
68.9
78.2
53.7
37.7
60.7
51.4
ImageWAM [ 59 ]
80.8
50.3
91.4
98.1
85.5
93.8
80.5
82.9
Table 2: LIBERO-Plus success (%) by perturbation category. For every method, Avg. is the unweighted mean of the seven categories, recomputed from the listed values.
Table 4: RoboCasa24 average success (%). Demos denotes training demonstrations per task.
Method
P.T.
Two targets
+ Color distractors
+ Purple light
Mean success ↑
Mean latency ↓
π0.5
✓
66.7
50.0
53.3
56.7
119.9
FastWAM
✗
46.7
30.0
36.7
37.8
296.9
ThinkWAM
✗
73.3
56.7
63.3
64.4
101.2
Table 5: A1Z success (%; 30 trials per model/condition) and mean latency (ms/chunk). P.T.: action pretraining.
Figure 3: ThinkWAM rollouts on the A1Z: two targets, color distractors, and purple light (top to bottom). Each row shows four chronological frames from a fixed third-person view.
ID
Configuration
R
A
LIBERO
LIBERO-Plus
# Multi-horizon future targets
–
Full ThinkWAM
✓
✓
98.70
84.76
1
w/o rethink
✗
✓
96.35
80.51
2 †
w/o latent correction
✗
✓
96.55
81.59
3
w/o affordance
✓
✗
97.10
82.01
4
w/o both
✗
✗
95.00
76.41
Table 6: Ablations on LIBERO and LIBERO-Plus (success, %). R: proposal–rethink–refine; A: affordance guidance. † Variant 2 retains both passes but uses base futures in the second pass.
Figure 4: Two paired open-loop examples at h=0,7,15 , with h=0 the observed input. (a) Base and residual-corrected futures decoded by a shared frozen decoder; circles and arrows mark changes. (b) Proposal (red) and refined (teal) executions; dashed lines retain proposal paths and arrows indicate time.
Figure 5: Closed-loop comparison of proposal (red) and refined (teal) actions from matched initial states on LIBERO (top) and LIBERO-Plus (bottom); each branch replans from its own observations. (A) Scene trajectories. (B) Snapshots at the refined branch’s first success. (C) Horizontal end-effector paths; bottom-row curves are slightly offset near their shared start for visibility. (D) End-effector height and gripper commands. Time excludes inference latency.
Method
LIBERO ↓
RoboTwin 2.0 ↓
LaWAM [ 11 ]
222.73
215.62
π0.5 [ 6 ] (RLinf SFT [ 51 ] )
286.62
254.59
OpenWAM- α [ 42 ]
413.97
426.83
FastWAM [ 52 ]
506.48
496.16
ThinkWAM
153.48
164.61
Table 7: Mean inference latency per action chunk (ms) on one A100 80GB.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Component
LIBERO
RoboTwin 2.0
RoboCasa24
Model size
1.507B
1.501B
1.505B
External action pretraining
None
None
None
Replan
14
32
16
Action chunk
16 steps × 7 dims
32 steps × 14 dims
16 steps × 7 dims
Future horizons
t+7 , t+15 , t+31
t+8 , t+16 , t+32
t+7 , t+15 , t+31
Query token counts
288 / 144 / 72
288 / 144 / 72
288 / 144 / 72
Appendix
Table 8: ThinkWAM architecture and benchmark-specific configuration.
Module
Width
Blocks
Heads
FFN
V-JEPA 2.1 ViT-L/16 (frozen)
1,024
24
16
4,096
Qwen2.5-0.5B
896
24
14
4,864
Memory
1,024
2
8
4,096
Temporal and latent-residual predictors
1,024
2
8
4,096
Affordance decoder
256
2
8
1,024
Shared ActionDiT
1,536
16
32
6,144
Appendix
Table 9: Module specifications. FFN is the intermediate feed-forward dimension. The proposal context has 80 tokens (64 action placeholders and 16 memory slots); refinement appends 504 corrected future tokens.
Figure 6: Multi-horizon targets and token budgets. (a) Recorded observations at t+7 , t+15 , and t+31 . (b) One crop at increasing blur, illustrating fine-to-coarse detail. (c) The 288/144/72 query tokens, one cell per token.
Figure 7: Affordance-head outputs on LIBERO. Each example pairs an RGB observation with the predicted affordance scores on a shared 0–1 scale. Dashed boxes identify task objects, and circles locate score maxima.
Stage
Steps
LR
GPU-h
LIBERO P0
40k
10−4→10−5
75.4
LIBERO P1
40k
10−4→10−5
116.5
RoboTwin
90k
5×10−5→5×10−6
211.7
RoboCasa24 P0
20k
10−4→10−5
43.0
RoboCasa24 P1
40k
10−4→10−6
88.0
Appendix
Table 10: Training budgets and estimated GPU-hours.
Method
LIBERO
RoboTwin
LaWAM
8×32 (7 native)
36×32 (16 native EEF)
π0.5 (RLinf SFT)
50×7
50×14 (32 padded)
OpenWAM- α
32×10 EEF
32×20 EEF
FastWAM
32×7
32×14
ThinkWAM
16×7
32×14
Appendix
Table 11: Action chunk shapes. Shapes are chunk length × output dimension; parentheses indicate native or padded dimensions.
Figure 8: Real-robot failure cases of ThinkWAM w/o rethink. Each row shows four chronological third-person frames from one episode: (a) manipulation of a non-target object amid color distractors, and (b) an incomplete transfer under purple illumination.
Task ID
Task instruction
SR (%)
Long-0
put both the alphabet soup and the tomato sauce in the basket
100.00
Long-1
put both the cream cheese box and the butter in the basket
100.00
Long-2
turn on the stove and put the moka pot on it
98.00
Long-3
put the black bowl in the bottom drawer of the cabinet and close it
98.00
Long-4
put the white mug on the left plate and put the yellow and white mug on the right plate
94.00
Long-5
pick up the book and place it in the back compartment of the caddy
96.00
Appendix
Table 12: Per-task LIBERO success (50 initial states per task, 2,000 records).
Task ID
Task instruction
SR (%)
Object-0
pick up the alphabet soup and place it in the basket
100.00
Object-1
pick up the cream cheese and place it in the basket
100.00
Object-2
pick up the salad dressing and place it in the basket
100.00
Object-3
pick up the bbq sauce and place it in the basket
100.00
Object-4
pick up the ketchup and place it in the basket
100.00
Object-5
pick up the tomato sauce and place it in the basket
100.00
Appendix
Table 13: Per-task LIBERO success (continued).
Suite
Category
Successes / total
Rate (%)
Long
Camera Viewpoints
367 / 419
87.59
Long
Robot Initial States
277 / 393
70.48
Long
Language Instructions
335 / 383
87.47
Long
Light Conditions
260 / 274
94.89
Long
Background Textures
143 / 289
49.48
Long
Sensor Noise
417 / 449
92.87
Appendix
Table 14: LIBERO-Plus success counts by suite and perturbation category.
Task
Clean (%)
Random (%)
Adjust Bottle
100.00
95.00
Beat Block Hammer
52.00
24.00
Click Alarmclock
95.00
76.00
Click Bell
100.00
88.00
Dump Bin Bigbin
97.00
85.00
Grab Roller
99.00
78.00
Appendix
Table 15: Per-task ThinkWAM success on the fixed RoboTwin 20-task subset (100 episodes per task and setting).
Task
SR (%)
PnPCounterToCab
34.00
PnPCabToCounter
32.00
PnPCounterToSink
26.00
PnPSinkToCounter
40.00
PnPCounterToMicrowave
26.00
PnPMicrowaveToCounter
34.00
Appendix
Table 16: Per-task ThinkWAM success on RoboCasa24 (24 tasks).
Figure 9: RoboTwin rollouts from matched starts. For each task, the left pair shows the proposal-only interaction and the right pair the interaction and final success with rethink. Circles mark the interaction region.
Figure 11: Additional paired open-loop comparisons, following the visual conventions of Figure 4 . A1–A8 and B1–B8 show decoded futures and action executions for the same examples at h=0,7,15 ( h=0 : observed input). For B1–B8, action MSE decreases by 96.98%, 46.20%, 98.32%, 58.18%, 39.86%, 40.91%, 40.39%, and 40.98%, respectively.
Figure 12: Additional closed-loop comparisons following the protocol and layout of Figure 5 ; proposal in red, refinement in teal. (A) End-effector paths on the initial scene. (B) Snapshots at refinement’s first success, when the proposal has not yet succeeded; circles and labels mark task-relevant differences. (C) Horizontal end-effector paths; black dots mark starts, squares and triangles mark proposal and refined endpoints. (D) End-effector height and gripper commands (white: open; filled: close). Rollouts stop at first success or the task budget (hatched: execution ended). Time is at 20 Hz after initialization, excluding inference latency.
World action models (WAMs) improve robot control by modeling how observations evolve, but generating future observations at test time incurs substantial latency. Fast-WAM removes this process for efficiency; however, our matched implementations show lower generalization for Fast-WAM than for future-aware alternatives, especially with scarce robot demonstrations and in out-of-distribution scenarios. To bridge this gap, we introduce LAWA, a WAM architecture that uses compact latent actions as an operational representation of future intentions, enabling efficient test-time future imagination without generating future observations. Specifically, a discrete tokenizer enhanced by action-free pre-training produces manipulation-centric codebook targets. LAWA jointly denoises a continuous latent state anchored to these targets with executable action chunks while omitting the future-video branch at inference. On RoboCasa, LAWA achieves state-of-the-art average success rates of 65.6% and 80.8% in the few-shot and full data settings, improving over the matched Fast-WAM baseline by 9.6 and 4.5 points, respectively. It also preserves the performance level of the matched Joint-WAM variant while requiring 42.9% lower inference latency. LAWA also demonstrates competitive zero-shot robustness on LIBERO-Plus and superior performance on real-world tasks. These results show that future imagination need not be discarded: retaining it with compact latent actions yields an effective trade-off among performance, generalization, and latency. Code and models will be released.
World Action Models (WAMs) can condition robot actions on generated visual futures, but this additional context is neither cost-free nor uniformly beneficial. We examine what a predicted future actually changes: the action delta Δa between the future-conditioned and future-free execution paths of the same model. Under a shared WAM checkpoint, this delta barely shifts average benchmark success (+0.8 points on LIBERO, −0.8 on RoboTwin2.0), while generating visual futures increases inference latency by 1.7--1.9×. These aggregate metrics conceal a concentrated failure mode that we term future-induced negative transfer: while predictive deltas benefit long-horizon tasks (+2.4 points on LIBERO-Long) and remain negligible on most routine manipulations, they induce severe degradation on precision tasks---with Press Stapler and Place Can Basket losing 54.0 and 34.0 points, recurring as top failure modes across independently trained WAM variants. We introduce Utility-WAM for in-place correction of future-conditioned control. Its primary learning component is a residual adapter that repairs future-induced velocity deltas under a frozen shared WAM by aligning them with ground-truth dynamic guidance and penalizing excess error beyond the future-free baseline. The adapter selectively eliminates the majority of severe task drops (reducing losses to 6.0 and 14.0 points) while preserving long-horizon foresight. To eliminate superfluous video computation, a lightweight pre-generation gate reuses counterfactual action outcomes to selectively bypass video generation on 75.5% of RoboTwin queries, cutting latency by 38% while improving average success to 93.3% on RoboTwin2.0 and 98.9% on LIBERO. On physical bimanual manipulation across three real-world tasks, Utility-WAM succeeds in 53 of 60 trials, compared to 43 for FAST-WAM.
Feihong Zhang, Xi Lin, Yulong Shi +7
Tsinghua University · Johns Hopkins University · Northeastern University +4
World Action Models (WAMs) connect visual prediction with robot control, but supplying predictive context often requires expensive future-video generation. Direct policies avoid this cost but lack an explicit interface for accessing future-indexed predictive information. We introduce ForeWAM, a World Action Model that separates forecasting from rendering to expose and shape latent predictive context for efficient control. Its core mechanism, Future-KV, performs a single Video DiT prefill over the current visual latent and noise-initialized future slots, then reuses the resulting key-value states throughout action denoising. To make this context relevant to control, we introduce dynamics registers supervised by latent actions from a frozen teacher during training, encouraging representations of interaction-induced transitions. This reusable context supports a lightweight, single-layer action decoder. We evaluate ForeWAM on LIBERO, LIBERO-Plus, RoboCasa, and real-world manipulation tasks. Without additional policy-level embodied pretraining, ForeWAM improves RoboCasa success by 9.7 percentage points over Fast-WAM at the same budget of 50 demonstrations per task, reaching 59.2%. With a single-layer decoder, it achieves 77.6% success on LIBERO-Plus and reduces policy-query latency to 88.7 ms on an NVIDIA A800, delivering a 6.27-fold speedup over Fast-WAM. These results show that latent predictive computation provides useful foresight for robust, efficient control without explicit future-video generation.
Jiakai Huang, Zhongbo Wu, Siyu Xu +5
Shanghai Jiao Tong University · ACE Robotics · Nanyang Technological University