LePlanner: An Iterative Amortized Controller For World Models
Saksham Bansal, Om Naphade, Chayan Aggarwal, Vrishin M
cs.RO · cs.AI
Quick verdict
Amortized controller refines plans through a frozen JEPA world model; efficiency gains are solid, but the 98% PushT headline is a selected maximum, not a mean.
Summary
LePlanner is an amortized iterative controller for goal-conditioned visual control that plans through a frozen JEPA world model (LeWM). A transformer maintains plan tokens and refines them over K learned steps, reading predicted latent-rollout consequences, replacing test-time search (CEM/MPPI/iCEM). Training uses an arrival-and-hold loss that supervises goal arrival at the hindsight-sampled offset q rather than at the fixed plan end, plus a frozen Gaussian-mixture penalty keeping action blocks inside the offline support; the stated motivation is the diagnosed 'horizon-reset procrastination' under receding-horizon replanning. On four simulated benchmarks (PushT, Reacher, TwoRooms, OGBench-Cube; 50 trials each, public LeWM checkpoints and datasets), it reports 98%, 100%, 100%, and 92% success; against matched CEM (300x30) on one L4 GPU, wall-clock is 3-49x lower and predictor transitions 20 vs 45,000 per plan. Ablations show that re-indexing arrival by q is the entire fix, with the hold term only reducing drift. Limitations: the PushT headline is the best of eleven task draws (sweep mean 94.5, disclosed by the authors); only four simulated environments on a single world-model family are tested; Figure 3 baseline protocols are not fully specified; the refinement loop is non-contractive beyond its trained depth.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LePlanner, an amortized iterative controller over a frozen JEPA world model, matches or exceeds test-time search planners on four visual goal-reaching benchmarks at 3-49x lower wall-clock and 20 vs 45,000 predictor transitions per plan.
- Supervising arrival at the hindsight goal offset q (arrival-and-hold) eliminates 'horizon-reset procrastination'; the q-reindexing alone is the fix, with the hold term only reducing post-arrival drift.
- LePlanner is not behavior cloning: dataset actions only fit a frozen support density and never serve as regression targets, yet its plans reach goals far better than replaying expert actions (terminal latent distance 0.0211 vs 0.3930, winning 99.6% of samples).
- Success is roughly invariant to the number of executed blocks m (a time-consistent closed-loop policy), whereas matched CEM swings 34% to 90% on PushT between m=1 and m=5.
Strengths
Transparent statistical self-audit
Appendix A discloses that the PushT headline (98/92) is draw 45, the best of eleven draws under a pre-stated stopping rule no draw met, giving sweep means (94.5/87.5) and ranges (88-100/78-96); paired exact McNemar tests and percentile bootstrap run on a shared trial manifest with recorded digest, plus an explicit noise rule (one episode = 2 points, <8-point differences are not results).
Convincing mechanism study
Appendix C measures and eliminates six rival explanations for the 88%-full-plan vs 52%-replan-every-block pathology, then fits a closed-loop recurrence with fixed point D*=b/(1-c) (c: 0.579 to 0.392, b: 0.041 to 0.024 after the fix); the lambda_h sweep (Fig 8, p=0.625/1.0) shows q-reindexing is the entire fix and the hold term only reduces drift - a conclusion honestly narrower than the design story.
Exact efficiency arithmetic plus matched timing
Reproduction-grade specification
Honest diagnostic tooling
Weaknesses
Selected headline number and internal inconsistency on PushT
Baseline protocol ambiguity and an undiscussed baseline win
Generality untested
Novel components do less work than the framing suggests
m-invariance claim is overstated for Cube
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a method paper (arXiv 2609.13845, preprint) proposing LePlanner, an amortized iterative controller that plans through a frozen LeWorldModel (LeWM) JEPA. Three claims: (1) learned iterative refinement replaces test-time search at a fraction of the cost - 20 predictor transitions per plan versus 45,000 for CEM (300 samples x 30 iterations x 5 blocks), and 3-49x lower measured wall-clock per episode; (2) an arrival-and-hold objective, supervising arrival at the hindsight goal offset q rather than at the fixed plan end, cures a failure the authors name horizon-reset procrastination; (3) a frozen Gaussian-mixture support penalty keeps plans on-manifold without behavior cloning. Headline results: 98% PushT, 100% Reacher, 100% TwoRooms, 92% OGBench-Cube, 50 trials each.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'horizon-reset procrastination'?
A failure mode named by this paper: a controller trained to reach the goal at the end of its planning horizon, when deployed with receding-horizon replanning, has its terminal deadline reset after every replan, so it keeps approaching the goal without ever arriving. The fix supervises arrival at the goal's data-derived time offset instead of at the fixed plan end.
How much faster is LePlanner than CEM?
On the paper's matched protocol (single L4 GPU, same session, shared trials), LePlanner's wall-clock per episode is 3-49x lower depending on environment and replanning interval, and it issues 20 latent transitions per plan versus CEM's 45,000 (300 samples x 30 iterations x 5 blocks).
Is LePlanner a behavior-cloning policy?
No. Dataset actions are used only to fit a frozen Gaussian-mixture density that penalizes plans outside the data's support region. The controller's first predicted action block differs substantially from the expert action (mean cosine similarity 0.552), yet reaches goals far better than replaying expert actions would (terminal latent distance 0.0211 vs 0.3930).
What success rates does LePlanner report and how reliable are they?
The paper reports 98% on PushT, 100% on Reacher, 100% on TwoRooms, and 92% on OGBench-Cube over 50 trials each. The PushT figure is the best of eleven task-draw seeds (sweep mean 94.5), a selection the authors disclose themselves; Reacher and TwoRooms are at ceiling, so PushT and Cube carry the discriminative weight.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 3/4 values arrive in the extraction as unlabelled number strings; the method-to-value mapping follows the stated method order and was cross-checked against Section 4.1 and Table 8 (consistent where comparable, and the inconsistencies noted are between the paper's own sections, not extraction artifacts).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.