When More Planning Hurts: Learned Planners for World Models

Josef Chen

cs.ROOct 27, 2026

When More Planning Hurts: Learned Planners for World Models

Authors: Josef Chen

Organizations: Independent Researcher

Abstract

On PushT, our fixed action updater (1,505 parameters) solves 8 of 16 scenes reserved from our configuration selection. The comparison planners perform as follows: Selected Adam (7 tasks solved, total planning and encoding time 218.440s), adapted iCEM (2 tasks solved, total planning and encoding time 49.309s), released LePlanner (16 tasks solved, total planning and encoding time 5.652s). The total planning and encoding time for our action updater is 59.079s. Note that our action updater does not satisfy the task-cost criterion we fixed before evaluation. We also conduct another matched experiment where we only replay the first ten commands in each context and then allow the next solve 4, 8 or 12 updates (subsequent solves all get 8 updates). Among all 16 contexts (8 scene clusters), the 3 budgets get 9, 8 and 7 successes, respectively. Interestingly, the extra updates did not rescue any failing task in this experiment. All the changes in success/failure happen in one scene cluster, where we observe that the extra updates decrease the costs of the selected latent world-model trajectories for our updater, while the tasks that used to be successes become failures. The code, weights, evaluators, execution traces and cost accounting accompany the paper.

Figures & tables

Figure 1: Similar block poses can hide a failed goal. In scene 18651, the official predictor gets a ten-command history and follows the plans with four and twelve updates. For the plan with four updates, the task succeeds at command 25. For the plan with twelve updates, the selected score is lower and the task fails at command 50. This is the only solve that differs between these two settings; in all other solves, eight updates were used. Success requires matching not only the block position and orientation, but also the pusher position. The recorded poses at the end of the episode are used to check the pusher’s distance from the goal and the joint position error. The two displayed angle errors are both within tolerance. In the figure, gold marks the goals in the dataset; blue and black show the recorded block and pusher poses, respectively; dotted lines indicate the pusher position error. All results in the matched experiment for every scene/model/effort combination are shown in the bottom panel. The scene in which there is a difference is highlighted.

Question	Observed evidence	Interpretation limit
Task value	Reserved sixteen-scene confirmation: updater $8/16$ , LePlanner $16/16$ ; updater fails the predefined criterion	Since the controllers vary on more than just the controller architecture, we can’t draw conclusions about the controller architecture independently of population effects
Value of effort	Matched $K\in\{4,8,12\}$ : $9/8/7$ successes; all changes in one of eight clusters	We do not test learned allocation policies. Also, since only one scene shows a change in outcome, conclusions about this are limited
Reuse	One frozen head runs across three related predictors and three horizons	While all predictors share the same encoder and controller interface, it’s unknown whether the test scenes appeared in their pretraining data, so it’s difficult to make strong conclusions about transfer

Table 1: Findings and their scope. We report each experiment separately. In particular, additional controller runs within a scene are not additional scene units.

Controller	Success	Commands	P+E (s)	Task (s)	Plan traj.	Backwards
Adapted iCEM500	2/16	725	49.309	54.317	73,177	0
Adam32, lr 0.1	7/16	621	218.440	220.676	4,318	4,064
Fixed learned8	8/16	603	59.079	61.493	1,230	984
Released LePlanner	16/16	243	5.652	6.715	216	0

Table 2: The confirmation study consists of 64 trials in total, using four fixed controllers on sixteen reserved scenes. Plan trajectories and backward calls are also included, but they measure different aspects of work, so they are not directly comparable across controllers. The time cost of planning and encoding (P+E) is summed together. For LePlanner, we sum its inclusive timer only once. All costs, including failures and cold starts, are included. Task time includes P+E.

Figure 2: Sixteen-scene confirmation. Shown are (left) number of successful tasks (out of sixteen total scenes) and (right) total planning and encoding time (in seconds; note that the sum includes unsuccessful tasks and the cold cost), using linear axes with different units, for four controllers that use different computations. In terms of the practical criterion, the updater fails against more traditional control approaches, and note that here LePlanner is being assessed as a complete controller.

First updates	Successes	Forwards	Backward/NN calls	P+E sum(s)
4	9/16	1146	904	55.983
8	8/16	1290	1032	62.309
12	7/16	1404	1136	68.310

Table 3: All three efforts;16 contexts in eight scene clusters per row. Counts/times include one acquired prefix plus each suffix, excluding replay. Every later solve uses 8.

Figure 3: Scene 18651: selected score and eventual outcome. Filled circles indicate success and crosses failure. Scores can only be compared within the context of the same model. All binary changes occur in this one cluster out of eight.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Native goals	Plan + encoder (s)	Episode (s)
CEM	119/216	1.029	2.912
Adam	132/216	6.289	8.227
Shared SFT	140/216	5.832	7.732
Same head, padded	140/216	5.831	7.673

Appendix

Table 4: The operating points for the direct-action methods correspond to 216 tasks per method from 24 scenes, three horizons and three predictors. We report numbers including all failures and cold work. Note that episode time includes the time spent in planning/encoding, so do not add the two.

Figure 4: Success and cost of direct-action at the configured settings, with instrumentation included in the means. Fig. 5 gives the paired intervals.

Figure 5: The figure shows shared-minus-Adam success differences with descriptive 95% scene intervals. The scenes are the same 24 for each cohort. Comparisons across cohorts should be interpreted with caution since the parameterization and implementation vary between cohorts. As a result, observed differences are not necessarily attributable to a single factor.

Method/protocol	Success	Commands	Plan traj.	P+E(s)	Task(s)
Four development scenes: 10646,2690,6916,2424
Released LePlanner	4/4	60	56	6.072	6.931
Owned learned8	2/4	154	310	16.713	17.267
Projected Adam	2/4	148	420	22.073	22.576
Bounded one-step CEM	1/4	167	2210	2.341	2.875
Common two-scene subset: 10646,2690

Appendix

Table 5: Two additional probes were also run for development purposes, on a four-scene and on a two-scene cohort. In all cases, the numbers are cumulative over the whole set of completed tasks. The number in the P+E column includes planning and encoding (and transfer when available). Plan trajectories is the number of plans (i.e. action sequences) considered and not the number of module calls (i.e. the computational cost can be different even when the numbers are the same). Module and transition counters are still present in this release. Note that P+E is included in the task service number. Hence, no extra column for it should be added. Cold-start, method-order and pipeline differences remain.

Setting	Success	P+E (s)	Task (s)	Plan traj.	Backwards
Calibration: four exposed scenes
Adam8, lr 0.1	1/4	31.187	33.402	350	280
Adam16, lr 0.3	1/4	30.782	31.398	630	560
Adam32, lr 0.1	2/4	57.278	57.902	1,122	1,056
iCEM100	0/4	4.897	5.621	3,988	0
iCEM300	0/4	5.566	6.331	11,908	0

Appendix

Table 6: Note that Plan trajectories is the number of action plans evaluated during planning, but does not necessarily correspond to the amount of computation, as different algorithms make a different number of calls to the backward method and/or batch in different ways. Task service includes P+E. P+E is total planning plus encoding time across all decisions the agent makes while executing the task (including cold starts and instrumented work). Each row corresponds to four scene units, with calibration and validation on different scenes, in a matched experiment with conventional controls.

Figure 6: Bounded conventional calibration (left) and separate development validation (right), four scenes per panel. The horizontal axis is the completed-task sum of planning plus encoding in seconds, on a logarithmic scale; the vertical axis is the task success count out of four scenes. All nine settings and all failure cases are presented, with the points representing instrumented executions. The two scene sets remain separate.

Method	Native goals	Plan + encoder (s)	Episode (s)
Cost-only A115	56/216	1.056	4.247
Halton–Boltzmann	116/216	0.532	2.053
CEM	106/216	0.550	1.928
Adam	122/216	5.324	6.795
Shared SFT	129/216	5.362	6.711
best65	114/216	0.690	3.828

Appendix

Table 7: Original fixed-five-knot matrix, showing the mean time (in seconds) for each episode. Note that the definition of an episode varies by source, and also that the instrumentation is not the same in all cases.

Figure 7: For each cohort, the panel summarizes the range of task-cost. The details of parameterization, recording and timing vary between cohorts. Note that these intervals correspond to resampling of source scenes and do not represent resampling of decisions or model fits.

Method	Native goals	Plan + encoder (s)	Episode (s)
CEM	106/216	1.780	3.218
Adam	122/216	1.640	2.984
Shared SFT	128/216	3.321	4.658

Appendix

Table 8: Compiled controls, run to completion, with 216 trials recorded for each method.

Method	Native goals	Plan + encoder (s)	Episode (s)
CEM	119/216	1.029	2.912
Adam	132/216	6.289	8.227
Shared SFT	140/216	5.832	7.732
Same head, padded	140/216	5.831	7.673

Appendix

Table 9: Direct-action H-dimensional controls. Padding retains the same fitted head.

Figure 8: For all 24 aligned and 192 offdiagonal partitions, intervals are computed conditional on the fitted heads and related predictors. Note that since all comparisons are made to the same historical shared reference, these comparisons are not independent. Positive values favor the specialist. Each fixed specialist is compared to a common reference (in a predeclared order). These intervals are computed for all 216 cases.

Method	Native goals	Plan + encoder (s)	Episode (s)
Official H3	110/216	5.424	6.402
Official H5	135/216	5.324	6.377
FT17 H3	141/216	5.158	6.068
FT17 H5	127/216	5.150	6.087
Shared pooled	140/216	5.832	7.732

Appendix

Table 10: All cells (task/model/H) reported according to the source records. There are 216 evaluations for each fixed specialist and the reused pooled reference model.

Fixed role	Target goals/48	Retention goals/48	Target P+E (s)	Retention P+E (s)
Pooled unchanged	35	34	3.773	3.752
Official unchanged	34	32	3.901	3.806
Official to FT17 warm	35	33	3.854	3.867
FT17 fresh	35	34	3.718	3.673

Appendix

Table 11: Complete fixed-role adaptation, with four roles; P+E indicates the recorded mean for planning and encoding. Each role/panel includes 48 task episodes.

Figure 9: All eight fixed modifications relative to the one shared seed13 reference model, with 216 paired tasks each. As before, we use the same 24 scenes and cached teacher lineage, and report unadjusted, descriptive 95% scene intervals. Notably, removing gradient magnitude and scale still retains the gradient-sign channel.

Fixed head	Native goals	Plan + encoder (s)	Episode (s)
Shared seed13	140/216	5.832	7.732
Shared seed23	131/216	5.408	10.162
Shared seed37	141/216	5.262	9.958
No gradient/scale	110/216	5.589	11.120
No cost	120/216	5.437	11.049
No budget	135/216	5.351	10.441

Appendix

Table 12: Controls are finite and complete, with only a single shared reference being reused. Mean P+E per episode is left unchanged (preserving the instrumentation and cold costs specific to its source).

Actual frozen policy	ID goals/16	ID planning (s)	OOD goals/20	OOD planning (s)
Fixed8	9	50.158	8	94.635
Threshold4/8	9	46.518	8	95.835
TRAIN-marginal random	9	31.847	5	67.341
Failed ridge gate	4	17.373	5	26.596
Fixed8 + ignored gate	9	46.519	8	94.213

Appendix

Table 13: All frozen EVAL policies. Note that the ID/OOD success denominators (16/20) count contexts that repeat over four source scenes. Planning times include the gate.

Native method	Decisions	Mean (s)	p50 (s)	p95 (s)	Maximum (s)
CEM	1666	0.133	0.075	0.132	15.551
Adam	1574	0.863	0.770	1.143	41.354
Shared SFT	1547	0.814	0.786	1.177	10.221
Same head, padded	1547	0.814	0.778	1.199	14.763

Appendix

Table 14: These numbers are not confidence intervals, and are not meant to be taken as measurements of latencies on real robots (nor deadline measurements). These numbers are meant to represent the distribution of latencies (planning plus encoder) per decision, taking into account all decisions including cold starts and decisions where the task ultimately failed. In total there are 6,334 decisions. The percentiles were computed by sorting the values, and then performing linear interpolation between values as necessary.

Compiled method	Mean (s)	p50 (s)	p95 (s)	Maximum (s)
CEM	1.780	0.391	2.171	123.361
Adam	1.640	1.122	1.900	82.226
Shared SFT	3.321	1.234	1.996	269.791

Appendix

Table 15: The compiled planning + encoding time (in seconds) for the full task. For each of the methods we ran, across all 216 tasks we ran, we report the median, p95 and maximum time. Note that this includes tasks that were unsuccessful (and cold work, if we logged it), but does not include the eager reference row (since it is not an additional compiled task).

Controller	Success	Commands	P+E (s)	Task + recorded artifacts (s)
Adapted iCEM500	2/16	725	25.525	30.115
Adam32, lr 0.1	7/16	621	207.673	211.419
Fixed learned8	8/16	602	57.649	61.385
Released LePlanner	16/16	243	6.008	7.679

Appendix

Table 16: Separate-machine repetition of the same 64 confirmation tasks; each success label matched the reference result. Time columns report totals for all T4 trials and include unsuccessful trials; setup is given separately below. Task time runs through saving the recorded artifacts and includes P+E. The original confirmation measurements are in Table 2 .

World Model-Based Planning

World model-based planning uses models of how environments change to compare imagined actions and choose goal-directed behavior.

See all 263 papers

How Much Planning Is Enough? Reducing Search and Computation in World-Model Planning
Oct 6, 2026Changbai Li, Sirui Li, Yichen Yang +4World Model-Based Planning Efficient Inference for World Action Models
Beihang University · Chengdu University of Technology · Nanyang Technological University
Reinforced Planning with Latent World Models
Aug 19, 2026Armin Sommer, Jannik SchillingWorld Model-Based Planning Robotic RL
Pantheon Industries
Adapt for Planning, Not just Prediction: Optimistic Test-Time World Model Adaptation
Oct 26, 2026Akshay Mete, Dileep Kalathil, P. R. KumarTest-Time Adaptation World Model-Based Planning
ECE Department, Texas A&M University, College Station, Texas, USA.

Oct 6, 2026J/K move · Enter open · S save

How Much Planning Is Enough? Reducing Search and Computation in World-Model Planning

Changbai Li, Sirui Li, Yichen Yang, Tongfei Chen+3

Beihang University · Chengdu University of Technology · Nanyang Technological University

World Model-Based Planning Efficient Inference for World Action Models

Open paper AskPDF arXiv

Visual world models enable goal-directed control through decision-time action search, but their deployment efficiency is often limited by conservatively large planning budgets. We show that competitive task performance can be achieved without agreement with the Full-budget action, that sufficient budgets vary across model--task pairs, and that iterative planners repeatedly encode solve-invariant context. To address these inefficiencies, we propose {SufficientPlan}, a simple deployment framework that requires no modification to pretrained world models or planner updates. Its {Paired Sequential Budget Certification (PSBC)} component uses paired closed-loop evidence to search for and certify a reduced model--task-specific budget within a predefined Full-performance tolerance. Its {Static-Context Reuse (SCR)} component caches observation and goal representations across search iterations while preserving candidate-dependent planning and selected actions. Experiments across multiple world-model backbones and visual-control tasks show that SufficientPlan substantially reduces search budgets and planning latency while maintaining competitive control performance.

Abstract

World Model-Based Planning

Explore similar work

How Much Planning Is Enough? Reducing Search and Computation in World-Model Planning

Reinforced Planning with Latent World Models

Adapt for Planning, Not just Prediction: Optimistic Test-Time World Model Adaptation