acceptodds
Withdrawn from ICLR 2027

EXPLICIT ROLES IN GENERATIVE PLANNING; RETRAINING, SELECTION, AND ROBUSTNESS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The contribution of explicit role information to generative planning is difficult to isolate when geometry, demonstrations, and candidate selection can convey related information. We present a controlled intervention framework for vessel planning that separates training-time role availability, inference-time generator inputs, and role-dependent candidate selection. The framework combines architecture-matched retraining, generator–selector factorial interventions, and paired robustness tests to distinguish learning with explicit roles from reacting to their removal or corruption at inference. Across MLP diffusion, temporal U-Net diffusion, and flow matching, we pair 30 newly trained explicit-role-blind policies with 30 preserved conditioned counterparts. Our prospective evaluation comprises 1,608,000 episode executions in controlled synthetic environments, using ten training seeds per architecture, shared scene banks, and crossed resampling. Among six primary retraining contrasts, only MLP diffusion with correct selection exhibits a positive arrival difference supported by multiplicity-adjusted confidence intervals, with a gain of 1.76 percentage points. In contrast, descriptive factorial experiments show that cyclically incorrect generator codes reduce arrival by 7.03–19.07 percentage points relative to correct codes across the three architectures, with eight candidates and a correct selector held fixed. Paired noise–delay and current–lag experiments further reveal non-additive changes in closed-loop outcomes, highlighting the importance of evaluating coupled disturbances. Together, these findings distinguish the benefit of explicit-role training from sensitivity to incorrect role inputs and establish a reproducible evaluation protocol for examining conditioning, candidate selection, and robustness in autonomous vessel planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generative vessel planning, Explicit role conditioning, Diffusion models, Flow matching, Candidate selection, Robustness evaluation.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yvd5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at whether explicit encounter role labels help with generative planning for autonomous vessels when the relevant information is already present via the geometry, demonstrations and candidate selection. The key contribution is a controlled evaluation framework which disentangles the three factors of: role availability during training, role inputs during inference and role dependent candidate ranking. It uses independently retrained models, generator-selector interventions and paired robustness tests across MLP diffusion, temporal U-Net diffusion and flow matching.

Validation is via synthetic episode runs (10 training seeds per architecture, new role-blind models paired with their conditioned counterparts, shared eval scenes, bootstrap confidence intervals for seed and scene variation). The comparisons are against behaviour cloning, Gaussian mixture policies, cross entropy planning, checkpoint sensitivity and joint disturbance.

The key conclusion is that explicit role training delivers very little benefit in practice: only the MLP diffusion case with the right role code shows a positive improvement (and that is only with multiplicity adjusted intervals). Providing incorrect role codes actually reduces arrival across all architectures, which shows that sensitivity to corrupted inputs is not in itself a sign of benefit from training with those inputs. The joint noise-delay and current-lag tests also show non-additive effects. The evidence does support the value of separating out these different interventions, but again the benefit is only in the synthetic environment where the role information can be present in an indirect way; it does not establish that it is in any way superior to conventional planners or in any way safe in the real world.

Critical strengths

Architecture-matched retraining and separate generator-selector interventions deal with a key ambiguity: the fact that performance degrades when a role code is removed or changed does not prove that training with that code leads to improved planning. The small gains from retraining compared to the large losses incurred by using the wrong code is a powerful argument in favour of this approach as it provides much greater methodological clarity and makes the conditioning claims far more interpretable.

The study pairs conditioned and role-blind models across 10 seeds per architecture, evaluates them on common scenes and accounts for both seed and scene variability via multiplicity adjusted intervals for the 6 main contrasts. Baselines, checkpoint checks and joint disturbance experiments are all used to put the results in context. This really does bolster confidence in the paper's bounded conclusions including the finding that the benefits of explicit role are very limited and setting dependent.

Critical weaknesses

A simple rule will recover every initial role label from the remaining geometric features and the role dependent demonstrations are still available to the 'role blind' models. The paper admits this but doesn't systematically vary the amount of role information that can be recovered via these alternative channels. The small retraining differences therefore leave open the question of when explicit roles actually become useful. This in turn limits the significance of the primary finding and undercuts the case for accepting the result beyond this one benchmark.

The noise-delay experiments maintain clean, fixed oracle roles whereas the separate role corruption experiments use a pre-determined cyclic substitution. Neither of these scenarios captures the situation where role estimates become uncertain or change as the agent accumulates noisy observations and experiences. When combined with synthetic dynamics and non-reciprocal targets this effectively rules out any broad conclusions. The large evaluation therefore provides very little evidence as to whether the results would generalise to the kind of autonomous vessel planning we are actually interested in and this is a critical point in my assessment of the papers impact.

Questions for authors

Given that geometry perfectly reveals the initial role, what existing evidence distinguishes limited usefulness of explicit roles from redundancy specific to this benchmark? Do the per-role results show larger conditioning benefits where the geometric recovery rule becomes less accurate later in an encounter? This would affect how broadly I interpret the main finding.

Reviewer_Bv2QOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a controlled-intervention framework for studying the role of explicit navigation-role information in generative collision-avoidance planning for unmanned surface vessels. The framework separates three intervention sites: role availability during training, role inputs to the action generator at inference time, and role-dependent costs used by the candidate selector. The authors compare conditioned and independently retrained role-blind versions of MLP diffusion, temporal U-Net diffusion, and flow-matching planners. Their evaluation uses 30 newly trained role-blind models paired with 30 preserved conditioned models, ten seeds per architecture, shared scene banks, generator–selector factorial experiments, checkpoint analyses, and observation/execution stress tests, totaling 1.608 million episode executions. The main findings are that explicit role training yields a statistically significant arrival improvement only for MLP diffusion with the correct selector, while incorrect generator role codes substantially reduce arrival across all three architectures. The paper concludes that training-time conditioning, generator-side corruption, and selector-side role costs should be evaluated separately.

Critical strengths

Carefully disentangled intervention design.: This paper clearly separates role availability during training, role inputs to the generator, and role-dependent candidate selection.

Critical weaknesses
  1. Weak central empirical evidence.: Among the six primary retraining contrasts, only one—the relatively weak MLP diffusion model with a correct selector—shows a statistically significant improvement after multiplicity adjustment, and the gain is only 1.76 percentage points. No statistically significant benefit is observed for the temporal U-Net or flow-matching models.

  2. Severely limited external validity.: The target vessels in the evaluation are nonreactive, the robustness tests retain clean oracle role labels, and corrupted role inputs are evaluated using only a single asymmetric cyclic permutation.

Questions for authors
  1. What causal effect is being estimated?
    Since role-blind models can partially recover roles from geometry and demonstrations, does the conditioned-versus-blind difference measure the marginal value of the explicit role code or the overall value of role information? How is this distinction justified?

  2. Are the conditioned models valid matched controls?
    The conditioned models are preserved from an earlier phase, whereas the role-blind models are newly trained. Were the codebase, data ordering, training procedure, and checkpoint-selection protocol identical across phases? If not, why should the 1.76-percentage-point difference be attributed to role conditioning?

  3. How general is the incorrect-role result?
    The corruption experiment uses only one asymmetric cyclic permutation. How do the authors distinguish general sensitivity to incorrect role information from sensitivity to this particular, potentially out-of-distribution mapping?

Reviewer_NxeAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies whether explicit role information genuinely benefits generative planning by separating role availability during training, generator inputs at inference, and role-dependent candidate selection. In a synthetic autonomous-vessel planning task, the authors compare explicitly conditioned policies with independently retrained role-blind counterparts across MLP diffusion, temporal U-Net diffusion, and flow matching architectures. They also conduct factorial interventions on generator role codes and selectors, together with paired robustness tests involving observation noise, delays, environmental current, and actuator lag. The evaluation comprises 1.608 million episode executions across shared scene banks, ten training seeds per architecture, and multiple candidate-selection settings. Among six primary retraining contrasts, only MLP diffusion with correct selection shows a statistically supported benefit from explicit role training, with a 1.76 percentage point increase in arrival. The other contrasts remain inconclusive. In contrast, cyclically incorrect generator roles consistently reduce arrival by 7.03 to 19.07 percentage points. The results suggest that explicit role conditioning has architecture and selection dependent benefits, while incorrect role inputs and coupled disturbances can substantially affect closed-loop planning outcomes.

Critical strengths

This paper has an unusually careful intervention design. It separates training-time role availability, generator inputs, and candidate selection, uses architecture-matched retraining, and evaluates shared scene banks with ten seeds per architecture. This substantially reduces common confounds in studies that infer the value of conditioning from inference-time ablations alone.

The validation is broad and statistically disciplined. The study includes 1.608 million episode executions, multiplicity-adjusted intervals for the primary contrasts, three generative policy families, and paired tests under noise, delay, current, and actuator-lag disturbances. The authors also report mostly null results rather than overstating the single significant training effect, which makes the conclusions about when explicit roles help and when incorrect roles are harmful more credible.

Critical weaknesses

The empirical setting is a highly simplified synthetic vessel environment with oracle role labels, fixed dynamics, hand-designed role coordinates, and a finite candidate selector. The robustness tests perturb continuous observations but keep the role, goal, and target masks clean. As a result, the study does not establish whether the conclusions transfer to real vessel perception, uncertain or changing roles, realistic hydrodynamics, or safety-critical deployment.

The central positive effect is weak and architecture dependent. Only one of six primary retraining contrasts supports a benefit from explicit role training, with a modest 1.76 percentage point arrival increase, while the other five intervals include zero. The large degradation under cyclically incorrect role codes mainly demonstrates sensitivity to deliberately corrupted inputs, rather than showing that explicit roles are necessary or broadly beneficial for planning.

Questions for authors
  1. How do the authors justify the generality of conclusions drawn from a synthetic vessel environment with oracle role labels, simplified dynamics, and a hand-designed candidate selector? Are there experiments using real vessel trajectories, noisy role estimates, or more realistic hydrodynamics?
  2. Since the role channels are partly recoverable from geometry and demonstrations, how much of the reported behavior is actually due to explicit role information? Could the authors compare against shuffled, randomly reassigned, or independently predicted role labels while keeping all other inputs fixed?
  3. The robustness experiments keep the role, goal, and target masks clean while perturbing continuous observations. What happens when role labels are misclassified, delayed, missing, or inconsistent with the current scene? This is important for assessing end-to-end robustness.
  4. The role selector uses a hand-chosen penalty weight and a specific cyclic corruption map. How sensitive are the results to the selector weight, alternative role cost definitions, and symmetric or random role corruptions?
  5. Only one of six primary retraining contrasts is statistically significant. Can the authors provide a power analysis and confidence intervals for the practical effect sizes, or an equivalence analysis showing whether the remaining contrasts are meaningfully small rather than simply underpowered?

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

Related papers

Discussion (0)

Sign in to comment.