Organizations: University of the West of England, Bristol, UK · University of Oxford, Oxford, UK · Imperial College London, London, UK · University of Exeter, Exeter, UK · University of Cambridge, Cambridge, UK
A world model predicts how a scene changes after an action, and a planner uses it to search for actions that reach a goal. If the camera moves or the controller changes, the same command moves the robot differently: when its commands are turned by 90 degrees, a frozen planner with the LeWM world model that succeeds in 77.5% of episodes succeeds in only 7.8%. Current fixes retrain the model, need training data, or forget a set-up when it returns. We propose Library of Hands (LoH), a training-free method that keeps the world model frozen and corrects only the map from commands to applied actions, which we call the hand. LoH reads the hand through the frozen model, with no extra sensor: it finds the action that best explains the change between two camera images (a reading). It stores every hand it identifies, so a returning set-up is recognised rather than re-learnt. In closed-loop tests, LoH more than doubles success on three simulated tasks, from 28.6% to 62.3%, and loses only 0.6 percentage points when nothing changes. It also works, unchanged, with two other world models and on two of three models pretrained on real robot data, and its gain grows with the quality of its readings.
Figures & tables
Figure 1: One method closes most of the gap, in simulation and on real-data models. Success (%) in the first two episodes after a known hand returns; bottom row: mean over the 12 cells. Right: the share of the gap to a planner told the true hand that LoH closes. LeWM: S1 and S2, on the 12 seeds on which all three planners ran; DINO-WM and JEPA-WM: S1, 6 seeds; RoboCasa: S1, 8 seeds ( Table 5 ). ∗ Trained by us on the data of LeWM.
Approach
(1)
(2)
(3)
(4)
(5)
(6)
Test-time training JEPA-TTT, AdaJEPA
✗
∘
∘
✗
✗
✓
Trained in-context reader SyncWorld
✓
✗
∘
✗
✗
✗
Calibration per episode V-JEPA 2
✓
✓
✓
✗
✗
✓
Coarse shift classifier ShiftWM-Bench
✓
✗
✗
✗
✗
✓
One kept correction one memory
✓
✓
✓
∘
✗
✓
LoH (ours)
✓
✓
✓
✓
✓
✓
Table 1: Only LoH has all six properties. (1) keeps the weights frozen, (2) needs no training, (3) estimates a continuous map (any linear map, not a few fixed classes), (4) recalls returning hands, (5) weighs each reading by its reliability, (6) works with the released model weights (checkpoints); ∘ : partly.
Hand
Change
Part
Real cause
N
none
none
the training set-up
A, D
turn 90 ∘ , 37 ∘
direction
a moved camera
T
random turn
direction
a manually set camera
E
mirror of x
handedness
an inverted axis
C
half gain, delay 2
scale, timing
a slower controller
O
offset on x
translation
an actuator bias
Table 2: The main hands. Each changes one part of Equation 1 . Delays are in control steps. Table 8 gives the others.
Method
PushT
Reach.
2-Room
All
Returns
S4
Frozen planner
20.6
28.6
36.6
28.6
26.6
78.1
ShiftWM-style
36.2
41.0
42.0
39.7
35.8
77.3
No memory †
48.4
44.3
47.0
46.6
43.8
76.9
One memory †
54.4
48.7
48.6
50.6
45.0
76.3
MBCD-style †
61.5
53.7
48.1
54.4
52.0
74.2
Fixed Share †
63.4
58.7
47.1
56.4
54.9
75.0
Table 3: Locked test on 20 new seeds. Success (%) on S1–S3, after returns, and with no change (S4). Bold : best; underlined : second (oracle excluded). † Earlier sensor ( Section 4.6 ). ‡ Run on 12 of the 20 seeds; on these 12 seeds, the frozen planner and LoH reach 28.4 and 62.4 over all episodes, and 26.1 and 60.3 after returns.
Figure 3: A hand that returns is recalled, not learned again. Success on S1 over the three LeWM tasks (8 seeds, 24 runs per method). (a) By visit of a hand (blocks of five episodes). (b) By episode after a known hand returns. Bands: 95% CIs over runs.
Model
Task
R2
S1
1 mem.
S4
S2
1 mem.
LeWM
PushT
0.98
+52.2
+11.4
−0.8
+36.9
+9.0
Reacher
0.70
+48.9
+17.3
−0.8
+28.7
+11.5
2-Room
0.90
+23.5
+10.0
−0.1
+13.0
+6.0
DINO-WM
PushT
0.998
+58.3
−8.3
−0.8
+46.0
+5.5
Reacher ∗
0.981
+56.2
+4.2
−2.5
+57.9
+5.5
2-Room ∗
0.950
+31.3
+6.2
−0.8
+26.3
+11.5
Table 4: One method, three world models. Values in pp: LoH minus the frozen planner (S1, S2: after returns; S4: no change) or minus one memory (1 mem.). LeWM: 20 seeds; others: 6 seeds (DINO-WM PushT S2: 12) and the first 35, 70 and 20 episodes of S1, S2 and S4. R2 : quality of the sensor, the share of the variance of the applied block action that the readings explain. ∗ Trained by us on LeWM’s data.
Model
R2
Frozen
True hand
LoH
Gap
1 mem.
No change
DINO-WM
0.80
7.8
35.9
26.6
67%
29.7
+7.5
JEPA-WM
0.70
14.1
35.9
21.9
36%
25.0
−6.2
V-JEPA 2-AC
0.68
21.9
35.9
35.9
100%
31.2
+8.8
Table 5: World models pretrained on real robot data. RoboCasa, place task (S1: 8 seeds; no change: 4 seeds): success after returns on S1 (%), share of the gap to the true hand closed, and LoH minus frozen with no change (S4).
Figure 4: The same returning hand, two planners. DINO-WM in RoboCasa (episode 10 of one run): hand A, a turn of 90 ∘ , returns after five episodes of the mirror E; robot camera at planning steps 0 to 60. The frozen arm swings away and fails; LoH recalls hand A and places the object.
Figure 5: The quality of the sensor sets the size of the gain. (a) Gain over the frozen planner after returns, with noise of standard deviation σ (in units of the standard deviation of the commands) added to each reading (LeWM: 8 seeds; DINO-WM: 6 seeds). (b) Share of the gap closed against the R2 of the readings (18 data points; Two-Room has no unsaturated blocks). Bands and bars: 95% bootstrap intervals.
Figure 6: Ablation study of framework components and the selection rule. Results are aggregated over 48 runs across multiple random seeds. Values denote changes in success rate relative to LoH , with error bars indicating 95% intervals. Component ablations are defined as follows: Sense , replacing the sensor with its earlier variant, with the floor disabled in both configurations; Explain , using a single fixed map; Remember , removing memory; and Act , using a frozen planner.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Constants
Sense
Adam, 100 steps, learning rate 0.1, nine starts: the command turned by 0∘ , ±45∘ , ±90∘ and 180∘ , the mirrored command, the mirrored command turned by 180∘ , and zero. A block counts if erra<θ , or if erra<3θ and erra<0.25errc ( errc : the error of the command). Jacobian: central differences, step 0.01. R~t=Ct/rˉ , eigenvalues clipped to [0.1,1000] ; rˉ is the median of tr(C)/2 over the last 400 blocks. g : least squares on 1,200 nominal blocks.
Explain
Class weights 1.0, 1.0, 0.3, 0.5 and 0.25 (identity, similarity, reflection, linear, affine); delay weight 1 for d=0 and 0.1 for d>0 ; prior 0.5 on the identity with d=0 . Noise prior σ2∼IG(2,σ0) with σ0=0.01s2 , where s2 is the expected squared norm of a command block sum. Map prior β−β0∣σ2∼N(0,σ2V0) with V0=τ2/σ0 and τ=1 (offset: 0.12s2/σ0 ). Clip model: at most 12 refits, tolerance 0.002.
Remember
h and q : Beta(1,1) posterior means over the last 20 episodes and the last 20 switches. Change point: the last 4 blocks, posterior above 0.6. New entry: rnew>0.5 and at least 4 blocks. Learning: entries with rk≥0.5 . Merge: log Bayes factor above 3. Use counts decay by 0.98 per episode. Removal: nk<0.2 and no win for 50 episodes, or no win within 20 episodes of creation. At most 64 blocks per entry and 16 entries. Floor: smallest singular value at least 0.2.
Act
CEM: 300 samples, 30 iterations, 30 elites, initial standard deviation 1.0 (normalised), horizon 5 blocks, warm start. θ : the 95th percentile of the frozen one-block error on 20 unshifted episodes (9.02 on PushT, 39.7 on Reacher, 59.0 on Two-Room, 272 on Cube, 0.100 on DINO-WM PushT).
Appendix
Table 6: Constants of LoH .
Task
Raw action
Clip
Success
g
R2
PushT
2-D pusher action
none
agent and block within 20 px, block angle within 20 ∘
0.940
0.98
Reacher
2-D joint torques
[−1,1]
every joint within 0.05 rad of the target
0.957
0.70
Two-Room
2-D velocity
[−1,1]
agent within 16 px of the target
0.967
0.90
Cube
5-D end effector
none
cube within 4 cm of the target
1.054
0.80
Appendix
Table 7: Tasks, with the sensor gain g and the sensor R2 on nominal LeWM data.
Code
Definition
Real change with the same effect
B, H
turn by −90∘ ; turn by 180∘
a moved camera; a camera on the other side
M
mirror of x , then a turn by 90∘
swapped and turned axes
OA
turn by 90∘ , then the offset of O
a moved camera and a bias
Z, X
dead band; saturation (non-linear)
a worn or limited actuator
ZA, XD
turn by 90∘ , then dead band; turn by 37∘ , then saturation
both of the above
F, G
PushT friction; centre of gravity moved by 20 px
a new surface or object
Appendix
Table 8: Further hands. F and G change the world, not the interface.
Schedule
Content
Episodes
S1
A, C, A, C, D, A, D in blocks of 5 episodes, 4 times
140
S2
keep the last hand with p=0.6 ; else a new random turn with p=0.1 , else uniform over A, C, D, E, N
140
S3
a new random turn in every episode
140
S4
N in every episode
60
S5, S6
as S1 and S2 with the non-linear hands
140
S7, S8
PushT: hands and world changes in blocks of 5
140, 160
Appendix
Table 9: Schedules. S1 to S4 are the main schedules.
ID
Prediction
Result
Held
Failed
P1a
The size of R~t predicts the residual: slope in [0.5,1.5]
1.19 / 0.92 / 0.97
3
0
P1b
The direction of R~t is informative
Reacher: ratio 1.93
1
0
P1c
The calibrated gain is 1±0.03 through the origin
0.740 / 1.018 / 1.042
1
2
P1d
R~t does not flag clipped blocks
trace ratio 1.01 (Reacher), 10.0 (Two-Room)
0
2
P2a
Angle error falls as Iθ−1/2 : slope in [−0.75,−0.25]
−0.20 / −0.42 / −0.58
2
1
P2b
Posterior scale calibrated: median ∣z∣ in [0.45,1.0]
0.96 / 1.32 / 0.65
2
1
Appendix
Table 10: Measurable predictions and their checks (PushT / Reacher / Two-Room where per task).
LoH minus
Measure
PushT
Reacher
Two-Room
Pooled
SD
Frozen planner
after returns
+44.5
+38.8
+18.2
+33.9 [31.1, 36.6]
15.5
Frozen planner
S4
−0.8
−0.8
−0.1
−0.6 [ −0.9 , −0.2 ]
1.4
ShiftWM-style
after returns
+29.6
+29.1
+15.1
+24.6 [22.6, 26.6]
11.3
No memory †
after returns
+15.2
+25.9
+8.9
+16.7 [14.6, 18.7]
11.6
One memory †
after returns
+14.0
+23.3
+9.1
+15.5 [13.6, 17.4]
10.7
MBCD-style †
after returns
+2.4
+14.4
+8.4
+8.4 [6.7, 10.0]
9.2
Appendix
Table 11: Locked test: LoH minus each arm, per task (20 seeds; true hand: the 12 seeds on which it ran, paired). Runs per pooled cell: 120 after returns, 60 on S4 (oracle: 72 and 36). Pooled: mean with 95% CI. SD: standard deviation of the pooled per-run differences. † Earlier sensor.
Arm
A ( 90∘ )
C (gain, delay)
D ( 37∘ )
E (mirror)
N (no change)
T (new turn)
Frozen planner
7.8
34.5
31.4
9.1
77.5
28.6
One memory †
41.4
51.6
54.2
36.6
74.0
49.4
Fixed Share †
60.5
56.6
62.9
54.6
63.1
51.8
LoH
66.1
64.6
69.1
59.7
68.2
57.0
True hand
77.9
72.6
80.0
78.3
76.7
76.1
Appendix
Table 12: Success for each hand (S1 and S2; 20 seeds; true hand: 12 of them). † Earlier sensor.
Stage
Comparison
PushT
Reacher
Two-Room
Sense
Full minus earlier sensor (sticky prior)
+2.5 [ −0.8 , 5.4]
+8.6 [5.3, 11.9]
+6.6 [3.5, 9.5]
Sense
Full minus earlier sensor (Fixed Share)
+1.4 [ −1.3 , 4.1]
+7.9 [3.5, 12.5]
+9.3 [6.0, 13.1]
Explain
LoH minus one model (linear, no delay)
+12.7 [8.8, 16.4]
+11.2 [7.7, 14.3]
+3.5 [0.5, 6.6]
Explain
LoH minus five classes with no delay
+11.4 [7.8, 15.0]
+5.0 [0.3, 9.7]
+1.8 [ −0.7 , 4.5]
Remember
LoH minus no memory
+19.3 [15.3, 23.6]
+22.9 [18.7, 27.2]
+7.6 [3.3, 12.1]
Remember
LoH minus one memory
+13.1 [8.9, 17.1]
+15.6 [12.2, 18.9]
+8.3 [5.8, 10.8]
Appendix
Table 13: The ablation per task, after returns (8 seeds; 16 runs per cell). The first two rows compare two sensors without the floor; the other rows use the floor on both sides. On 12 further seeds (72 runs each), LoH minus one model is +8.3 [6.5, 10.1], and LoH minus no delay on hand C is +21.4 [18.1, 24.8].
Test
Result
Offset hand (S9)
The offset costs the frozen planner 41.8 pp [30.1, 52.7] on hand O. LoH minus frozen on O: +23.4 [14.0, 32.6] (56% of the gap); after returns: +37.1 [32.0, 42.4]. Fresh seeds: +42.2 , +21.8 , +34.0 . On Two-Room, the offset costs only 3.1 pp, and LoH is 6.2 [3.1, 9.4] below the frozen planner, mostly in the first episode after hand A (post hoc).
Offset hand on DINO-WM and JEPA-WM
Partial: the grid reached its time limit with 15 and 13 of 18 complete LoH runs, so we give no verdict. On these runs, after returns: +60.8 [52.5, 69.2] and +58.7 [49.0, 68.3]; on O: +22.7 [4.7, 40.0] and +22.3 [4.6, 38.5]. On Two-Room, LoH again loses on O ( −20.0 [ −34.0 , −4.0 ] and −16.0 [ −24.0 , −10.0 ]).
One hand (L1)
LoH minus the true hand, episodes 5–139: −0.3 [ −1.9 , +1.2 ]; Two-Room −0.3 [ −0.9 , +0.4 ]. The library ends with one entry.
Long stay (L2)
First five episodes on N after 120 episodes of B: −10.0 [ −15.0 , −3.8 ]; Two-Room −12.5 [ −22.5 , −2.5 ]. On PushT and Reacher, the whole loss is the first episode on N (37.5% against 81.2%).
One switch (L3)
After a switch from N to B: +76.6 [73.3, 80.0]; Two-Room +55.4 [50.7, 60.0]. Before the switch, the cost is 0.7 pp.
2, 4, 8 hands
All episodes (PushT and Reacher, 32 runs per cell): LoH 76.0, 71.3, 67.6; Fixed Share 75.2, 71.5, 67.2; one memory 71.7, 66.5, 60.7; frozen 45.3, 27.1, 20.4. The library ends with 2.4, 4.3 and 8.6 entries.
Appendix
Table 14: Further tests on LeWM, unless a row names another model.
PushT
Reacher
Two-Room
DINO-WM
σ
gain
V
gain
V
gain
V
gain
0
+47.6
+19.3
+39.7
+22.9
+18.2
+7.6
+58.3
0.3
+43.7
+22.4
+33.7
+21.7
+13.6
+8.1
–
0.6
+38.8
+27.0
+24.2
+15.7
+7.4
+5.2
+54.2
1.0
+28.8
+23.8
+12.6
+8.5
+3.4
+2.2
+45.8
1.5
+18.8
+17.0
+5.1
+1.9
+0.1
+2.2
+29.2
Appendix
Table 15: The sensor-noise sweep, after returns: gain over the frozen planner, and the value of the library V ( LoH minus no memory, same noise). LeWM: 16 runs per cell (8 seeds). DINO-WM: 6 runs, S1.
Locked test
Streams (AUC)
Arm
S1–S3
After returns
S4
No change
Rotation
LoH
62.3
60.4
77.5
82.5
78.1
LoH without the floor
61.9
60.3
77.4
52.9
64.5
Frozen planner
28.6
26.6
78.1
83.3
2.0
Appendix
Table 16: LoH with and without the floor. Locked test (20 seeds): success (%) over S1–S3, after returns and with no change (S4). 500-episode streams of the JEPA-TTT protocol on PushT (three seeds): AUC of held-out success with no change and under the action rotation.
Task
Shift
Frozen
AdaJEPA
JEPA-TTT
LoH
PushT
action rotation
2.0
4.3
68.0 / 50.9
82.3 / 78.1
contact rotation
19.3
19.3
19.3 / 14.6
19.3 / 18.3
Two-Room
wave
42.3
40.3
63.3 / 54.4
49.7 / 45.9
grid
46.0
48.3
67.0 / 59.5
49.0 / 45.2
Reacher
phase
23.3
18.0
22.0 / 17.4
25.7 / 23.2
harmonic
20.7
20.3
22.3 / 17.7
23.0 / 20.3
Appendix
Table 17: The JEPA-TTT protocol (three seeds). Held-out success (%), Best / AUC; one value where the two are equal. No change: AUC with no shift.
Arm
After returns
S4
LoH minus arm, after returns
Arm minus frozen, S4
Continual test-time training
26.1
74.3
+34.1 [31.3, 36.9]
−3.8 [ −5.6 , −2.0 ]
AdaJEPA (reset each episode)
26.6
76.5
+33.7 [30.2, 37.2]
−1.6 [ −3.2 , 0.0]
JEPA-TTT (persistent)
26.4
71.4
+33.8 [30.0, 37.6]
−6.7 [ −8.7 , −4.9 ]
Calibration in every episode
41.8
64.7
+18.4 [15.7, 21.2]
−13.4 [ −16.7 , −10.2 ]
LoH
60.3
77.7
–
−0.4
Frozen planner
26.1
78.1
+34.2
–
Appendix
Table 18: Learned and calibration methods on our schedules (12 seeds; 72 runs after returns, 36 on S4).
DINO-WM
JEPA-WM
V-JEPA 2-AC
Frozen encoder
DINOv2 ViT-S/14
DINOv3 ViT-L
V-JEPA 2 ViT-g
Calibration: R2 / gain / turn of the map
0.80 / 1.08 / 39.3 ∘
0.70 / 0.73 / 22.3 ∘
0.68 / 0.95 / 35.2 ∘
S1 schedule
A, E, A, E, D, A, D
A, B, A, B, D, A, D
A, E, A, E, D, A, D
Frozen / true hand / LoH / one memory
7.8 / 35.9 / 26.6 / 29.7
14.1 / 35.9 / 21.9 / 25.0
21.9 / 35.9 / 35.9 / 31.2
P0: sensor R2 (bar 0.5)
pass
pass
pass
P1: LoH − frozen (bar +10 )
+18.8 [ +4.7 , +34.4 ] (8), pass
+7.8 [ +3.1 , +10.9 ] (8), fail
+14.1 [ −3.1 , +29.7 ] (8), pass
Appendix
Table 19: Results in RoboCasa, place. Success after returns on S1 in %. Differences with 95% bootstrap intervals; seeds in parentheses.
Figure 7: The frame of each model turns with the camera. (a) The view of V-JEPA 2-AC at each camera yaw θ (degrees), with the turn of the frame in which the model reads the commands and the R2 of the sensor (grey: below 0.5). (b) RoboCasa place, 8 calibration episodes per yaw, no injected hand. Filled marks: R2≥0.5 . Shaded: the camera yaws of the 261 DROID views. Dashed: the line that the frame of V-JEPA 2-AC follows on DROID, fixed before the runs.
Rule
Measure
Bar
Value
n
Verdict
Locked test, 20 seeds
L1
ours minus frozen, after returns
≥+10
+33.9
[31.1, 36.6]
120
pass
L2
ours minus four baselines † , after returns
>0 , CI
+5.5 to +24.6
lower ≥+4.1
120
pass
L3
ours minus frozen, S4
≥−1
−0.6
[ −0.9 , −0.2 ]
60
pass
L4
ours minus ShiftWM-style, S1–S3
≥+5
+22.6
[20.6, 24.7]
180
pass
L5
ours minus no memory † , S1–S3
≥+5
+15.8
[14.1, 17.5]
180
pass
Appendix
Table 20: Pre-registered rules on the final method. † Earlier sensor.
Post-training is essential for turning pretrained generalist robot policies into reliable task-specific controllers, but existing human-in-the-loop pipelines remain tied to physical execution: each correction requires robot time, scene setup, resets, and operator supervision in the real world. Meanwhile, action-conditioned world models have been studied mainly for imagination, synthetic data generation, and policy evaluation. We propose Human-in-the-World-Model (Hi-WM), a post-training framework that uses a learned world model as a reusable corrective substrate for failure-targeted policy improvement. A policy is first rolled out in closed loop inside the world model; when the rollout becomes incorrect or failure-prone, a human intervenes directly in the model to provide short corrective actions. Hi-WM caches intermediate states and supports rollback and branching, allowing a single failure state to be reused for multiple corrective continuations and yielding dense supervision around behaviors that the base policy handles poorly. The resulting corrective trajectories are then added back to the training set for post-training. We evaluate Hi-WM on three real-world manipulation tasks spanning both rigid and deformable object interaction, and on two policy backbones. Hi-WM improves real-world success by 37.9 points on average over the base policy and by 19.0 points over a world-model closed-loop baseline, while world-model evaluation correlates strongly with real-world performance (r = 0.953). These results suggest that world models can serve not only as generators or evaluators, but also as effective corrective substrates for scalable robot post-training.
Yaxuan Li, Zhongyi Zhou, Yefei Chen +5
Current Robotics · Tsinghua University · Peking University +1
World models such as DINO-WM and LeWM specify the goal with an image, which is difficult to obtain in advance for novel tasks. We present the Grounded World Model (GWM), a latent world model that enables zero-shot planning in the real world from language goals alone. Given a candidate action sequence and the current observation, GWM predicts the future in the visual space of a pretrained video-language embedding model. The frozen readout of this embedding model maps this imagined future and the task description into the same embedding space, where their negative cosine similarity serves as the planning cost. Training GWM requires only offline and task-agnostic video-action pairs and no language labels. In simulated experiments on WISER, planning with GWM, which executes the candidate action of lowest cost, solves 87% of 288 tasks with unseen instructions and visual signals, while ten fine-tuned VLAs average 22%. We then scale GWM up with real robot data, and use it for zero-shot planning in realistic simulation and real scenes. In the IsaacSim evaluation, planning with GWM completes all 70 trials across 14 tasks that require reasoning over referring expressions, matching a modular planner grounded by a frontier VLM, while pi0.5 reaches 37/70. Deployed on a real Franka, the same stack completes 55/60 separately evaluated pick-and-place sub-tasks, comparable to the modular planner's 52/60, with the full system running locally on a single consumer GPU. Project website: https://quanyili.github.io/gwm-wiser/.
World models offer a promising way to help robots understand how the physical world evolves and plan complex behaviours through imagination. Yet existing studies mainly demonstrate what these models can accomplish, leaving unclear when their predictions remain useful for planning and where they fail. We study this question using action-conditioned predictors built on five frozen self-supervised visual backbones: V-JEPA 2, V-JEPA 2.1, VideoMAEv2, VideoPrism, and DINOv2. We use frozen backbones to test representations intended to transfer across environments. We evaluate these models on diverse Meta-World manipulation tasks and real-robot interactions from BridgeData V2. We find that a world model guides action selection reliably only when the goal lies within, or slightly beyond, the trajectory it imagines during planning. With five-step rollouts, the length the predictor was trained on, the world model ranks actions reliably only for targets five to ten control steps ahead, whereas task goals lie 16 to 53 steps away. Neither an 81-fold larger predictor nor longer-rollout training extends this range; the encoder affects both range and closed-loop success, with V-JEPA 2.1 performing most consistently. More fundamentally, the limit persists under perfect prediction: using the real simulator, success falls from 92% to 41% as the target moves from five to twenty steps ahead of a five-step rollout. Planning therefore requires either longer imagined trajectories or closer subgoals. For distant goals, pure imagination succeeds in 23% of episodes, planning with feedback (MPC) raises success to 30%, imagining as far as the goal to 47%, and nearby expert subgoals to 76%. Used within its plannable range, a world model can also improve a vision-language-action (VLA) policy: choosing among eight actions the VLA proposes raises its success from 65% to 77% across 16 different tasks.