acceptodds
Under review as a conference paper at ICLR 2027

Branch-Prox: Joint Arrival-Time and Action Optimization with World Models

Abstract

Goal-conditioned planning with world models often optimizes action sequences by matching the predicted terminal representation to a goal representation at a fixed horizon. However, this objective does not explicitly encourage reaching the goal sooner. We introduce Branch-Prox, a trajectory optimization method that jointly optimizes arrival time and action sequences through a differentiable world model. Each candidate arrival time defines a branch whose objective combines elapsed time with a penalty for exceeding a goal-matching tolerance in latent space. An earlier arrival time may appear unfavorable under the current action sequence yet become preferable after action refinement. Rather than selecting an arrival time based on the current trajectory alone, Branch-Prox solves a convex proximal subproblem for each candidate and jointly selects the arrival time and action update using the optimized local model values. Each box-constrained subproblem reduces to a one-dimensional dual search. Under standard smoothness assumptions on the goal-matching costs and exact branch solves, we establish sufficient decrease of the planning objective with backtracking line search. Branch-Prox integrates into model predictive control without world-model fine-tuning or auxiliary model training. Across three LeWorldModel case studies with 64 matched episodes and three planner seeds, Branch-Prox achieves the lowest mean failure-capped steps among common-horizon reference methods. On Reacher, the mean decreases from the strongest reference baseline's 34.40 steps to 13.57 with 100% native success. On DINO-WM Wall, Branch-Prox reaches all 50 goals with 19.30 mean capped steps, compared with 37.44 for iCEM.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, Goal-Conditioned Planning, Trajectory Optimization, Arrival-Time Optimization, Proximal Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ybGLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Branch-Prox optimizes action sequences together with a discrete arrival endpoint in a frozen world model. Each endpoint defines a cost combining elapsed time and a penalty for exceeding a latent goal tolerance. The method solves a proximal subproblem for every endpoint, selects the best local update, and checks it through nonlinear backtracking. The paper proves sufficient decrease under stated assumptions and evaluates the planner on selected LeWorldModel and DINO-WM tasks.

Critical strengths
  1. Refining endpoints before choosing one addresses a concrete limitation of optimizing only the current winning branch. An earlier endpoint can look worse under the current actions but become preferable after refinement. The analytical example and recorded planning update explain this mechanism clearly.

  2. The box-constrained subproblem has a scalar-dual solution, and the descent analysis supports the proposed backtracking procedure. The authors also distinguish this within-solve guarantee from physical task success and report the additional computation required.

Critical weaknesses
  1. The benefit of the branch update is less clear than the benefit of the arrival objective. Comparisons with current-branch refinement and same-objective Projected Adam favor Branch-Prox most clearly on TwoRoom. However, Branch-Prox computes eight endpoint gradients per update, compared with one for Adam, and the comparison does not match planning time. The results therefore leave open how much of the advantage comes from the update rule rather than the additional computation.

  2. The tasks were selected after inspecting a broader evaluation. The authors disclose this, but without a summary of the omitted results it is difficult to judge where the method helps and where it fails. This matters because the added benefit of the branch update already varies across the displayed tasks.

Questions for authors
  1. Does Branch-Prox retain its advantage over same-objective Projected Adam when planning time is matched? The existing budget sweep is useful, but does not settle this comparison.

  2. Why is the goal tolerance defined relative to the current latent distance? As this scale changes at each MPC decision, how does the choice affect behavior near the goal compared with a fixed tolerance?

  3. Which tasks were included in the broader evaluation, and how were the displayed tasks selected? Please summarize the omitted results and the main failure cases.

Reviewer_46RiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper observes that plans made in latent space could have earlier states in the plan that are closer to the goal than the terminal state in the plan. Based on this information, actions can be refined to minimize the arrival time, which would be the path cost plus cost-to-go. This process can be parallelized with multiple plans.

Critical strengths

N/A

Critical weaknesses

The paper appears to be motivated by the basic insight of heuristic search [1] that the total cost of a path through a given node in a search tree is the path cost (cost from the start state to get to the node) plus the cost-to-go (the estimate of the cost of a shortest path from the state to a closest goal state). The paper comes to this through a roundabout way. The paper is hard to follow.

From what I could understand, the actions are being optimized to follow the approximation of a shortest path. There are many other methods that could be used here, such as a sampled version of heuristic search, that the authors do not consider. Moreover, there appears to be a significant time disparity between the authors’ method and methods to which they compare. Tables in the appendix show the proposed method takes multiple times longer than other methods. Perhaps the other methods could have compared more favorably if just given the same amount of time. This is alluded to in Table 2 “work is unequal”.

Therefore, it is unclear if this method is doing fundamentally better compared to other methods, or if the time spent tweaking of actions could be better spent using other methods.

[1] Hart, Peter E., Nils J. Nilsson, and Bertram Raphael. "A formal basis for the heuristic determination of minimum cost paths." IEEE transactions on Systems Science and Cybernetics 4.2 (1968): 100-107.

Questions for authors

What experiments do you have accounting for the extra time taken by your method?

Reviewer_ref9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies image-goal planning with frozen world models. Given a current image and a goal image, a pretrained latent predictor such as DINO-WM or LeWorldModel is used to optimize an action sequence in latent space. The aim is to bring the predicted endpoint latent close to the goal latent, and the plan is executed in MPC fashion. No policy, critic, or model is trained.

Critical strengths
  1. A precisely stated problem. That terminal objectives do not prefer early arrival is a simple and correct observation; the distinction between “maximum horizon” and “arrival time” is clean, and the fix requires no additional training.

  2. Correct theory, tightly coupled to the algorithm. Proposition 1’s dual characterization yields a cheap branch solver, and Theorem 1 is a correct specialization of the standard prox-linear result to the branch envelope. The analytic example can be re-derived and is backed by an observed instance from a real solve. The authors are also explicit about the limits of the guarantee: a single MPC decision, exact solves, and a documented floating-point exception.

  3. The most rigorous statistics of the five papers. These include an episode-level bootstrap, paired sign-flip tests, Holm correction over a predeclared comparison family, fixed-deadline success rates, and conditional arrival-time analysis.

Critical weaknesses
  1. Tasks were selected after the fact, which is the main concern in this work. The authors state explicitly that the three tasks were chosen after inspecting outcomes from a broader evaluation, and that complete records are kept in an accompanying research archive. The excluded tasks are not reported, which directly weakens the generality of the conclusions.

  2. “Arrival time” is conflated with reliability. Capped steps reward both success rate and speed, and much of the gain actually comes from success rate: On Reacher, when both succeed, Branch-Prox arrives only 2.80 steps earlier than GRASP; the main difference is success rate: 100% vs. 75.5%. On TwoRoom, the gap in the Current branch control comes entirely from success rate +0.04 steps when both succeed.

The “earlier arrival” emphasized in the title and narrative does not fully match where the gains mostly come from.

  1. Shorter effective horizon may be a confound. Baselines improve greatly with shorter H (AdamW on Reacher goes from 82 to 33), which suggests that shortening the effective horizon by itself reduces compounding model error. The Fixed H control uses only H = 8. A terminal-objective control that uses the same optimizer with a development-selected shorter H is missing, so the contributions of “preferring early arrival” and “adaptively shortening the horizon” cannot be separated.

  2. Method-level comparisons confound several factors. External baselines use a different objective (unnormalized, no tolerance) and different initialization (shifted warm starts vs. fresh initialization at every decision). Projected Adam separates part of this, but it was designed after the original test outcomes were seen, and the compute is unequal.

  3. Limited scope of the theoretical guarantee. Theorem 1 guarantees only a decrease of the latent proxy objective within a single decision. It depends on a smoothness constant L that is not estimated and may be large for neural predictors, and on exact solves. It guarantees neither convergence to a stationary point or global optimum, nor physical arrival, and the envelope itself is nonconvex and nonsmooth.

Questions for authors
  1. What were the results on the excluded tasks? Can they be reported in full in the appendix, along with the original selection criterion?

  2. Can you add a control that uses the Branch-Prox optimizer with only a terminal objective at a development-selected shorter H, to separate the early-arrival preference from horizon shortening?

  3. How was the reference configuration for the LeWorldModel tasks chosen? Were test episodes used?

  4. Can you report success rate and conditional arrival time (when both methods succeed) separately, to make clear which accounts for most of the gain?

  5. Can you provide equal-wall-clock comparisons, or an efficient implementation that reduces endpoint-gradient computation?

I am more than willing to reconsider my score if the additional results address my concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.