acceptodds
Under review as a conference paper at ICLR 2027

DiG-Nav: Learning Differentiable Goal Allocation for Multi-Object Navigation

Abstract

Multi-object navigation requires an agent to prioritize unvisited targets while learning continuous actions under partial observability. We introduce DiG-Nav, which encodes the remaining goal set with masked self-attention and computes a differentiable allocation over goals. The weighted goal representation conditions the action policy, while a short-horizon latent rollout supplies an auxiliary value objective for learning target priorities. Alternating model fitting and policy improvement preserves gradients through imagined actions without optimizing the predictor toward inflated values. We characterize the sensitivity of allocation probabilities and distinguish temperature smoothing from changes in the highest-ranked goal. On Gibson, Matterport3D, and Flightmare, DiG-Nav achieves episode success rates of 83.6%, 67.1%, and 69.5%, respectively, exceeding the strongest compared baseline by 3.1, 5.3, and 4.2 percentage points. Component ablations support the contributions of learned allocation and predictive supervision. Real-world quadrotor demonstrations illustrate continuous execution across multiple targets.

Primary area:
applications to robotics, autonomy, planning
Keywords:
multi-object navigation, reinforcement learning, differentiable goal allocation, model-based reinforcement learning, embodied AI, robotic navigation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_gVxWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DiG-Nav learns soft weights over the remaining goal categories and uses their weighted representation to condition a continuous-action policy. Training combines PPO with an auxiliary value objective from short latent rollouts. At deployment, the actor generates actions directly. The paper reports results on Gibson, Matterport3D, and Flightmare, together with real quadrotor demonstrations.

Critical strengths

+Learning goal priorities together with actions is a sensible approach to unordered multi-object navigation. +The method achieves higher mean episode success than the strongest compared baseline by 3.1, 5.3, and 4.2 percentage points across the three simulation settings. The MP3D ablations also support the usefulness of learned allocation and the imagined-value objective.

Critical weaknesses

-The allocation objective needs a clearer justification. In Eq. (8), shifting weights toward a high-value goal can increase the score even if the imagined trajectory stays unchanged. Summing future values can also count the same confirmation reward several times. I understand that this is an auxiliary objective, but its connection to completing all goals remains unclear. The ablations show that it helps performance, without separating the benefit of better actions from the effect of reweighting predicted values. -The policy has no explicit memory of past observations or explored areas. The completion mask records confirmed goals, but cannot retain the location of a goal seen earlier and not yet confirmed. Different exploration histories can therefore produce the same input but require different decisions. This is a substantial limitation for partially observed multi-object navigation, and its practical impact is not evaluated. -The baseline adaptations and evaluation settings lack important details. PONI needs a category-selection module, while EGO-Planner and FUEL need semantic search and goal ordering. Appendix E.3 describes their roles but does not explain how these modules were implemented and trained. Scene splits, episode budgets, and the actual reward and auxiliary-loss coefficients are also missing. These details matter for assessing and reproducing the comparisons.

Questions for authors
  1. Can you test a variant where the imagined-value objective does not update the allocation scorer, while PPO still trains it and the remaining training setup stays unchanged? This would help isolate the contribution of this gradient.
  2. Please explain how the baselines were adapted to multi-goal navigation, especially their category-selection and goal-ordering modules. Please also provide the scene splits, evaluation episode counts, episode budgets, and reward and auxiliary-loss coefficients.
1 comment

Thank you for the additional experiments. The both-route detachment control directly addresses my request: the reported ESR difference of 67.1% versus 63.7% supports a contribution from imagined-value updates to allocation. The occlusion probe also quantifies the practical impact of the memory limitation. These results address substantial parts of my review.

My main remaining concern is the baseline comparison. The response explains the roles of PONI, EGO-Planner, and FUEL, but the actual category-selection and goal-ordering rules, training sources, scene splits, episode counts and budgets, and reward/loss coefficients remain unspecified. These details are needed to assess whether the compared systems were configured appropriately.

I also note from the response to Reviewer kGYY that the hard selector performs close to the full method: ESR is 66.3% versus 67.1%, while mSPL is 58.9% versus 58.8%. This limits the evidence for a clear advantage of soft differentiable allocation over hard selection. Uncertainty estimates for the new controls would help interpret these differences.

The rebuttal strengthens the empirical support, but the comparison details remain important to my assessment. I will keep my score unchanged for now.

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

The paper presents a method towards multi-object navigation by particularly learning a differentiable goal allocation to guide navigation. The primary contribution is the use of learned allocation (attention) over CLIP embedding of remaining goals to arrive at a conditioning vector for the navigation policy. Additionally, the paper also incorporates model-based (learned latent dynamics) RL style auxiliary update to improve policy learning. The results demonstrate improvements over several baselines adapted to the sequence-agnostic navigation task (i.e. no pre-specified ordering of goals),

Critical strengths

The paper presents an interesting attempt to jointly learn goal prioritization and navigation and demonstrates empirical improvements across simulated navigation tasks.

Critical weaknesses
  • I don’t understand the advantage of learning goal allocation along with a navigation policy. This seems to unnecessarily complicate the learning problem and the final policy might not generalize better than factorized policies that have a high-level semantic navigator selecting goals for a navigation policy. The paper does not provide controlled comparisons against such factorized policies, making it harder to evaluate the value proposition of joint learning.
  • I’m also not fully convinced about the need for a differentiable goal attention mechanism.
    • The ablations in Table 2 do not necessarily isolate the benefit or need of differentiable allocation. Perhaps a high-level goal-allocation policy learned through RL or other hard goal allocation + goal-conditioned navigation policy can serve as a useful baseline to highlight the need for differentiability here?
    • For the overall objective, I don’t understand the motivation for encouraging higher allocation entropy (Eqn 6,11). Intuitively, wouldn’t sharper goal allocations encourage more consistent navigation with lesser jitters? Some empirical characterization of this regularization would be useful.
    • The analysis presented in Section 4 follows from the properties of softmax and doesn’t seem to provide much insights into the goal allocation/behavior of the policy.
  • While the paper presents statistics at an aggregate level for the simulated task, I did not find any discussion about the goal allocation behavior (potentially comparing/contrasting with baselines). I believe it is critical to provide more insights into this as the benefits observed in the aggregate metrics could be a consequence of poorer navigation policies in the baseline.
    • Also for the real world trials, it would have been nice if the authors are able to present some aggregate stats in the appendix E, the interpretations seem rather shallow for these trials and makes it difficult to assess the significance of the real robot experiments.
Questions for authors
  • Could you provide more details on how fairness of comparison with baselines were ensured? (compute-matching, perception modules used etc.,)
  • More qualitative evidence for how the learned soft-goal allocation attention changes over a task would be appreciated.
  • What is the justification for holding the goal mask/allocation fixed in the imagined-value objective (Eqn 8), I don’t fully understand how this works for assigning credit for appropriate goal sequencing + accomplishing a goal during the rollout.
Reviewer_ZRH1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DiG-Nav for multi-object navigation. The method encodes the remaining categories using frozen CLIP text features and masked self-attention, computes soft allocation weights, and uses the weighted goal representation to condition a continuous-action policy. Training combines PPO with an auxiliary objective that evaluates goal-conditioned values along short imagined latent trajectories. The predictor and critics are fitted separately and frozen during policy improvement while retaining gradients through their inputs. The paper also analyzes the sensitivity of the allocation probabilities and distinguishes probability smoothing from changes in goal rankings.

Critical strengths

The paper addresses a meaningful coordination problem between target prioritization and continuous navigation. Using the same soft weights to condition the actor and aggregate imagined goal values creates an explicit learning pathway between predicted outcomes and target priorities, while deployment avoids additional rollout search. In the reported results, improvements are reported across three simulation settings with five seeds, and the MP3D ablations show substantial performance drops when learned allocation or the imagined-value objective is removed. These results provide useful evidence that the components contribute within the evaluated architecture. The paper is also careful about the scope of its claims, particularly in distinguishing soft allocation from discrete goal commitment and probability smoothing from reduced switching.

Critical weaknesses

My main concern is that the method contribution appears incremental, and the experiments do not yet establish the distinctive benefit of the proposed coupling. Masked attention over goal embeddings, softmax weighting, PPO, and latent-imagination gradients are established techniques; the primary contribution is their integration through an explicit allocation interface. However, uniform allocation is a relatively limited control, and removing the imagined-value objective changes both the supervision and the gradient pathways. Without a capacity-matched policy baseline or a control that selectively detaches the allocation weights on the imagination branch, it remains difficult to determine whether the gains specifically arise from the proposed differentiable allocation mechanism. My second concern is that the behavioral interpretation and reliability of this interface are insufficiently validated. Since the actor receives a weighted mixture of goal features, allocation weights do not directly specify which goal the executed behavior pursues. The imagined objective also keeps the goal allocation and completion mask fixed while accumulating predicted values, making it a surrogate whose practical alignment with complete-mission success warrants closer examination. Its usefulness depends on the learned predictor and goal critic, but no held-out prediction or return-estimation errors are reported to assess whether the training data adequately support these decisions. This leaves the dependence on data coverage and model accuracy unresolved. Finally, the qualitative physical trajectories provide limited additional evidence about behavioral reliability, as aggregate completion and intervention outcomes from the reported 100 trials are not provided.

Questions for authors

Could the authors clarify how much of the improvement is specifically attributable to gradients through the allocation weights, rather than additional predictive supervision or model capacity? A comparison retaining the same predictor, critics, and fitting objectives while detaching the allocation weights only on the imagined-value branch would directly inform this assessment. Could the authors also provide representative time-aligned allocation weights, commands, and goal-confirmation events, together with prediction or goal-value errors from the existing evaluation settings, to establish whether the learned priorities correspond to useful and consistent navigation behavior? Regarding experimental comparability, please clarify the concrete multi-goal adaptations of the principal baselines and which observations, rewards, controllers, and training budgets are shared. Finally, are aggregate completion and intervention outcomes available for the 100 physical trials? These clarifications would help distinguish a useful integration of existing components from a more substantial methodological advance and assess the reliability of the proposed allocation mechanism. Due to time limitation, I fully understand if the authors cannot meet all of above requirements. I encourage the authors to choose the most significant points to respond.

Reviewer_38XMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses unordered MultiON, where an embodied ground or aerial agent must locate and confirm a set of N distinct target object categories in an arbitrary order within a single episode budget. The authors propose DiG-Nav, which couples target prioritization with continuous control and model-based auxiliary value learning through three components: (a) Masked Remaining-Goal Encoding and Soft Allocation (b) Latent Rollout Auxiliary Objective (c) Alternating Optimization and Sensitivity Analysis; along with enabler components.

The authors have put in a significant effort in this paper. While DiG-Nav addresses an interesting problem and reports consistent aggregate gains across three simulators, the experimental validation and technical reporting fall short of the expected quality in its current state. This work seems to be incremental in terms of sheer novelty.

Critical strengths

Coupling target prioritization with continuous control for unordered multi-object navigation is a well-motivated problem.

The consistent empirical gains across Gibson, Matterport3D, and Flightmare over five seeds alongside the core MP3D ablations.

The paper is candid about its theoretical boundaries --- explicitly noting that soft weights do not enforce discrete goal commitment, that temperature scaling does not prevent argmax switching.

Using masked self-attention over unvisited CLIP category embeddings provides a clean, permutation-equivariant interface for variable-sized goal sets.

Critical weaknesses

The ablations do not isolate the core claim of differentiable allocation gradients from auxiliary supervision or score reweighting.

Essential reproducibility parameters (all reward/loss weights, scene splits, episode counts, and baseline multi-goal wrapper implementations) are omitted.

The policy lacks spatial/temporal memory for unconfirmed sightings under partial observability.

The empirical reporting omits quantitative results for the 100 real-world trials.

Encouraging high allocation entropy pushes weights toward uniformity (blending multiple goal embeddings into an ambiguous average vector).

There are typos and some technical issues that is overlooked for now, till other things are corrected. May be paper finalized in hurry.

Questions for authors

"hard-allocation control", "capacity-matched PPO control", and "detached-imagination control" are required to separate gradient pathways from model capacity, but I could not locate the authors reporting any of these controls in their tables or some analysis.

How does DiG-Nav handle revisiting goals that were seen earlier in an episode (outside the 0.5 m confirmation radius) once they leave the single-frame ResNet-18 view?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.