acceptodds
Under review as a conference paper at ICLR 2027

LIFT: Learning Long-Horizon Team Feasibility for Heterogeneous Robot Exploration

Abstract

Heterogeneous robot exploration can depend on interactions whose value appears only later. Opening a door or operating an elevator may add little immediate coverage, yet change which joint plans become executable for the team. We call this changing set future team feasibility and present , a model-based approach for learning its evolution from partial observations. At each decision segment, a deterministic planner constructs a compact set of executable joint plans. represents the team state as a relational graph, encodes each plan by its physical semantics, and learns latent dynamics that predict future state, facility status, reward, and plan legality. These predictions are used in legality-masked imagined rollouts to evaluate long-horizon consequences. Feasibility-changing interactions are rare in ordinary exploration experience, so we augment replay with trajectories collected by a rule-based teacher and use temporary policy supervision during training. An event-aware history retains interaction information that remains relevant over longer decision horizons. In Habitat, temporary supervision changes the ranking of the enabling elevator action from 0/55 to 55/55 fixed states, and the preference persists after the supervision is removed. completes all 10 frozen evaluation episodes with 97.4% average coverage and achieves a 0.90 completion rate on a 50-episode extension, where it also records fewer collisions than PPO. Ablations show that teacher supervision is important for cross-floor interaction, while removing history degrades long-horizon prediction and performance on manipulation tasks.

Primary area:
reinforcement learning
Keywords:
Heterogeneous Multi-Robot Exploration, Model-Based Reinforcement Learning, Future Team Feasibility, Long-Horizon Coordination, World Models, Multi-Robot Planning

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

This paper studies long-horizon heterogeneous robot exploration where some actions may provide limited immediate reward but improve future task feasibility. The authors introduce the concept of future team feasibility and propose LIFT, a world-model-based framework that predicts how candidate joint plans affect future executable behaviors. LIFT represents robot plans using semantic action embeddings, learns latent dynamics models to predict future states, rewards, and plan feasibility, and performs imagined rollouts to select actions with higher long-term value. Experiments in simulated multi-robot exploration environments with heterogeneous robots demonstrate that LIFT can improve exploration performance and reduce failures compared with several reinforcement learning baselines.

Critical strengths
  1. Interesting problem formulation. The paper identifies an important limitation of conventional exploration objectives: actions that are valuable for future decision making may not provide immediate information gain or reward. The notion of future team feasibility provides an intuitive perspective for long-horizon robot decision making.

  2. Well-designed system architecture. The combination of semantic action representations, latent world modeling, and imagination-based planning is technically coherent. The action semantic embedding is a reasonable design choice to handle dynamically changing plan spaces.

  3. Meaningful benchmark scenarios. The proposed environments include heterogeneous robots and enabling interactions such as opening doors and operating elevators, which better reflect long-horizon coordination challenges than standard exploration benchmarks.

4.Comprehensive empirical analysis. The paper provides several ablations investigating the role of feasibility prediction, teacher demonstrations, and different components of the proposed framework.

Critical weaknesses
  1. The method relies heavily on a highly structured planning abstraction. The proposed approach assumes that a deterministic planner can generate a set of executable joint plans, and LIFT only needs to select among these candidates. This shifts a substantial part of the decision-making difficulty into the handcrafted planner and limits the applicability to settings where such structured action spaces are available.

  2. The generality of future team feasibility remains unclear. Although the proposed concept is interesting, the current formulation is tightly coupled to predefined skills and symbolic plan representations. It is unclear whether the same idea can extend to continuous manipulation tasks or scenarios where feasible actions are not explicitly enumerated.

  3. The experimental evaluation is limited to simulation. The experiments are conducted in Habitat-based simulation environments. While the benchmark is carefully designed, the paper does not demonstrate whether the proposed approach transfers to real robotic systems or more realistic uncertainties.

  4. Comparison with modern model-based RL methods is insufficient. The baselines mainly include classical reinforcement learning approaches. Since LIFT is based on latent world models and imagined rollouts, stronger comparisons with recent model-based RL methods would be necessary to better establish the advantage of the proposed approach.

  5. The performance improvements are relatively modest. Although LIFT improves some metrics, the gains over strong baselines are not consistently large, and in some settings baseline methods achieve comparable or better exploration performance. More evidence is needed to demonstrate that future feasibility modeling provides substantial practical benefits.

Questions for authors
  1. How does LIFT extend beyond the predefined joint plan space generated by the deterministic planner? The current approach appears to depend on enumerating feasible plans beforehand. Could the authors clarify whether future team feasibility can be applied when the action space is continuous or when new skills emerge?

  2. How sensitive are the results to the quality of the planner and skill library? Since the planner determines the available candidate actions, could the authors provide analysis of how performance changes with weaker or incomplete plan generation?

  3. How computationally expensive is the imagined rollout procedure? Could the authors report the inference-time overhead introduced by feasibility prediction and multi-step planning compared with standard exploration policies?

1 comment

Thank you. We added direct interventions, prediction diagnostics and complete decision timing. Revisions are blue; full protocols/results are in Appendix F and Supplement Sec. 11.

1. Planner dependence. The constructor handles current admissibility, precedence, assignment and ordering. LIFT learns forecasts and a selection policy within this interface; it cannot choose omitted actions. We added all-valid, current-mask persistence and training-only slot-frequency baselines. Original Full exceeds these at validation h=5/10 under the conditioning limits below, supporting predictive information, not planner independence.

2. Discrete versus continuous formulation. Sec. 3.1 now separates history-conditioned admissibility from its bounded discrete implementation. A continuous extension would need an action-conditioned feasibility model, a sampler/optimizer and grounded execution checks. Our task-level PICK/PLACE does not demonstrate continuous manipulation; this remains an extension.

3. Simulation/real robots. Quantitative evidence is from simulation. We corrected the comparison table that could imply a demonstrated real-robot advantage over PPO. Controlled misses, delays and noise do not establish physical robustness; hardware/interface descriptions are compatibility records, not comparative trials.

4. Modern model-based baselines. We strengthened within-architecture evidence with three-seed component removals and a separate prediction study: 12 models, 4,000 updates each, identical fixed replay (10,627 loaded items), matching initialization and sampled batches within each seed, final checkpoints only. Arms are original Full, class-balanced Full, matched balanced mask-off, and no validity loss/mask. Balanced variants are separately labeled. An unsupervised head is only a diagnostic control, so we include strong occupancy baselines. These fixed-replay studies do not replace the still-missing matched online DreamerV3/TD-MPC2 comparison.

5. Modest gains. Full/mask-off/loss-and-mask-off coverage is 95.90/94.87/94.69%, and minimum-floor coverage is 95.21/92.87/94.65%. Completion remains similar (128/127/129 of 150); return favors removing both components, and task/seed effects are mixed. These are three seeds on 50 reused starts. Separately, original Full's validation invalid AP at h=5/10 is 0.499/0.723 versus slot-prior 0.268/0.390; Brier errors are 0.103/0.173 versus 0.117/0.212. Primary h=1 AP is below the prior (0.238 vs 0.255), and h=5/10 use only 27/15 shared windows. Later descriptors are predicted, but recorded future action slots remain inputs; targets are future menu-slot validity. Thus the study supports bounded multi-step prediction, not semantic-action counterfactual validity or proven closed-loop necessity. All arms/modes and horizons are reported.

6. Q1: New skills and learning. New skills need execution/admission contracts, representations and transition data; we do not demonstrate autonomous skill discovery. Existing acquisition/retention evidence is narrower: the same 55 elevator SELECT states improve from 0/55 to 55/55, and two fixed rollouts change from timeout to success, retaining behavior over 32 further updates with teacher policy supervision effectively removed. This concerns enabling behavior within existing skills, not online test-time learning or open-ended self-evolution.

7. Q2: Incomplete planners. On 50 matched starts, Full completes 44/50 with the original menu, 43/50 with 20% random deletion, 44/50 with 40% deletion, and 41/50 after slot shuffling. The loss/mask-off branch gives 44/47/45/43 of 50. Deletion protects status quo and rounds down; shuffling retains the candidate set. Controls, contexts and masks move together, with selected-slot mapping checked. K32/16/8 prefix restrictions each yield 45/50 for Full. These quantify dependence, not superior robustness, invariance, or recovery of essential missing actions; they use earlier short-continuation checkpoints on five reused assets.

8. Q3: Computation. H=15 imagination occurs during actor training, not each evaluation decision. Three serial 50-episode repeats on TITAN RTX give complete decision medians 1163/1172/1174 ms and P95 about 2255 ms; policy medians are 8.79/8.80/8.81 ms, planner-core about 724-727 ms. Timing runs from observation acquisition after the prior skill to next-plan receipt, including processing, proposals, encoding, IPC and synchronization, excluding skill execution/reset/startup/training. Each episode's first decision and ten global warm-ups are excluded. Component quantiles are not additive. Learner-step medians are 0.314/0.300/0.306 s for Full/mask-off/loss-and-mask-off, with overlapping ranges, not isolated head cost. No paired speedup over Priority is claimed.

Reviewer_pFw4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes to select plans based on how well they enable future feasible choices. A latent space model, RSSM, is used to predict future team feasibility. Experiments show the method is beneficial relative to some existing methods in a Habitat-based experiments.

Critical strengths

I like the core concept of future team feasbility and the formulation of the idea as sequence of executable sets. The idea itself is novel to my knowledge and merits exploration.

Critical weaknesses

Some details about the method are currently unclear and would benefit from elaboration:

  • The executable set of joint plans Mt\mathcal{M}_t is conditioned by the current observation but given partial observability, the system would generally require an observation history?
  • the teacher module and demonstrations seems important to the method; can you clarify how these were generated?

Experiments appear to be only performed in a single domain and thus, generalization of the learnt model is unclear.

  • How different were the held-out scenes from the training scenes? Also, the main experiments involve a "frozen" 10-episode split; what aspects of the episodes are "frozen"?
  • The closed-loop comparison against PPO (which is the more realistic setting) seem to suggest that LIFT may not be necessary, especially given it is more complex relative to PPO. Can the authors elaborate on these comparisons?
  • What planner did you use in the experiments? I couldn't seem to find this in the main text.

Writing could be improved to organize the information.

  • Concepts are used before being properly introduced. For example, what is a "menu index"? This is mentioned in the introduction but is not defined up-front. It was only later in Section 3.2 that I understood that the planner generates a "menu" of joint plans and this set is executable set in 3.1.
  • The physical semantics vector could be better described. To clarify, is each element of M\mathcal{M} a joint plan or single robot plan (since ϕ(a)\phi(a) describes "its robot, skill, target" rather than multiple robots).
  • Also, what is a "delivered" observation? In what way is the observation "delivered"? What is a "delivery model" and why is it required?

Minor:

  • Why ∣Mt∣≤64|\mathcal{M}_t| \leq 64? why this arbitrary limit?
  • The paper could benefit from a limitations section (there are a couple of sentences at the end of the conclusion but I think the limitations of the approach extend beyond just scale and real-world eval)
Questions for authors

please see weaknesses above. The most important aspects to me are about partial observability and the generalization of the approach.

2 comments

Thank you for the helpful feedback. We address each point below; revisions are blue. Full protocols/results are in Appendix F and Supplement Sec. 11.

1. Partial observability. We replace M(o_t) by Mt=PK(bt)M_t=P_K(b_t), where bt=B(ht)b_t=B(h_t) summarizes delivered observation/action history. The policy also maintains recurrent RSSM state and selected event-history nodes. This is a bounded approximation, not a sufficient-statistic guarantee; admission can fail under hidden/stale conditions.

2. Teacher generation. Using the current menu/task progress, the rule teacher selects the earliest unfinished, precedence-ready task. It ranks candidates by task-serving controls, frontiers, active controls and fewer unrelated interactions, with hash-based tie-breaking. It never queries future states/menus. The calibration fork has 26 successful episodes and 1,211 unique teacher rows expanded to 4,844 copies. Teacher-containing sequences constitute 50%; BC updates the policy head. These are calibration records, separate from later checkpoints.

3. Domain/generalization. Results concern Habitat indoor tasks with fixed skills. New diagnostics use 20 development/50 validation episodes from new reset seeds on five existing assets. These test new episodes, not new scenes or skills. Active-team checks use padded subsets of three slots.

4. Train/test scenes. The supplied 45-scene inventory excludes the five general-suite scenes but contains four cups-scene IDs. We corrected scene-disjoint wording and list the IDs in the appendix. One inventory cannot establish every historical checkpoint's membership. The new diagnostic's exact initial-observation exclusion audit does not prove all-state disjointness.

5. Frozen split. Scene, task variant, episode seed and initial observation are fixed; subsequent trajectories can differ by policy. Three training seeds reuse 50 starts: 150 rollouts, not 150 independent scenes. New prediction windows at h=1/5/10/20 number 50/27/15/5 and are shared across seeds; post-terminal padding is excluded. Earlier studies retain separate checkpoints/counts.

6. Benefit beyond PPO. The historical 50-case comparison gives LIFT/PPO 45/44 successes and mean collisions 91.82/124.8, while PPO has higher return/coverage and shorter duration. We therefore do not infer necessity from PPO alone. Direct three-seed removals give Full coverage 95.90% versus 94.87%/94.69% for mask-off/loss-and-mask-off, with similar completion (128/127/129 of 150). Separately, original Full's validation invalid AP at h=5/10 is 0.499/0.723 versus training slot-prior 0.268/0.390, with lower Brier error. Primary h=1 is below the prior and h=10 has 15 windows. The supported benefits are average exploration progress and multi-step predictive information, not uniform success superiority.

7. Planner. It is our deterministic constraint-based constructor (planning/action_space.py), not an external solver. It assembles ready assignments, checks admissibility/conflicts, includes status quo, deduplicates, orders and truncates. All selectors share it; Priority is a separate rule selector. Missing proposals cannot be recovered by the policy.

8. Definitions first. Menu/slot are defined on first use. Sec. 3.1 separates a one-segment assignment, an H-step plan, admitted actions and the proposed subset.

9. Team descriptor. Each candidate assigns the whole team, including KEEP/HOLD where appropriate. Controls have one row per robot; context describes the whole candidate. RSSM receives both slot identity and descriptor, so semantic augmentation does not prove permutation invariance.

10. Delivered observations. These are observations available by the current information cutoff. The simulated interface uses visibility gating, 15% misses, 0.08 m position-noise SD, latency up to two segments and expiry after eight. Maintained maps, observation ages and history interpret delayed evidence. The teacher sees the same boundary; this is not demonstrated real-sensor robustness.

11. K=64. This is an implementation budget. At one continued checkpoint, K64 yields 44/50 successes and K32/16/8 each 45/50; doors improve while cups worsen. Restricting the selectable prefix retains all 64 descriptors, so it does not measure smaller-planner cost or establish an optimal K. Random deletion/reordering results are also included.

12. Limitations. Added explicitly: fixed teams/skills, reused scenes, mixed task/seed effects, sparse long-horizon windows, and no matched online model-based baseline or real-robot validation. Prediction targets menu-slot validity and still receives recorded future action slots. Complete decisions take 1.16-1.17 s median, excluding skill execution; the 8.8 ms policy call is only one component.

We thank the reviewers for identifying where our definitions and evidence needed strengthening. We have revised the manuscript in blue and provide a separate point-by-point response to each reviewer. The revision centers LIFT's contribution on an explicit learned team-action interface: an interaction by one robot can change the choices later available to its teammates.

1. Clearer formulation and novelty. Sec. 3.1 now distinguishes a one-segment joint team action, an H-step plan, current admissibility under observation history, and the conditional distribution of future admissible sets. The deterministic constructor supplies a bounded current menu; the latent model predicts its future descriptors/validity for imagined actor training. Temporary slot identity and physical descriptors are distinguished. Delayed enabling effects are already representable in conventional planning; our contribution is this concrete predictive interface, supported by directed experience and event history.

2. Direct component comparison. Across three training seeds, Full has mean coverage 95.90%, versus 94.87% without imagined masking and 94.69% without validity loss/masking. Minimum-floor coverage is 95.21%, 92.87% and 94.65%. Task completion remains similar: 128/150, 127/150 and 129/150. The same 50 starts are reused across seeds; task/seed effects are mixed. We retain the original success predicates and report coverage over successes and failures. The supported improvement is average exploration progress, not universal success-rate superiority.

3. Prediction beyond occupancy baselines. A separate fixed-replay diagnostic trains 12 models for 4,000 updates each, with matched initialization/batches within seeds and final checkpoints only. Original Full's validation invalid-class AP at h=5/10 is 0.499/0.723, versus 0.268/0.390 for a training-only slot-frequency prior; Brier errors also improve (0.103/0.173 vs 0.117/0.212). It also exceeds all-valid and current-mask persistence at these horizons. The designated primary h=1 AP remains below the slot prior (0.238 vs 0.255); h=5/10 have 27/15 shared windows. Later action descriptors are predicted, while recorded future action slots are still supplied. These are measurements of future menu-slot validity, not new closed-loop successes or semantic-action counterfactual evidence. All configurations, horizons and both conditioning modes are disclosed; class-balanced variants do not replace original Full.

4. Sensitivity, cost and protocol transparency. We added candidate-prefix restriction, random deletion/reordering, active-team compatibility, and complete decision timing. These delimit the fixed proposal interface rather than establish planner independence or arbitrary-team scaling. Complete decisions take about 1.17 s median and 2.25 s P95, including proposal construction and communication; the policy call is about 8.8 ms. Physical skill execution is excluded. Teacher generation, partial observations, frozen starting cases and scene-overlap boundaries are now explicit. The added studies reuse existing assets and do not establish new-scene transfer.

Existing paired elevator evidence additionally shows acquisition and short retention of enabling behavior within the fixed skill library. We use that limited learning claim rather than assert autonomous discovery of skills. The original PPO comparison remains competitive; a matched online model-based baseline and quantitative real-robot validation remain open. Historical comparisons, short continuations, new closed-loop summaries and prediction diagnostics are reported separately.

The complete new protocols, results and limitations are in manuscript Appendix F and Supplement Sec. 11; the manuscript body through the conclusion remains 10 pages. We hope these revisions make the technical contribution and the scope of the supporting evidence clearer.

Reviewer_52TJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The work studies a 3-robot heterogeneous team, solving a task in scenes where certain discrete actions like "open-door" or "use elevator" are available to change navigability for the multi-robot team. The authors aim to create a learning-based multi-robot policy to effectively explore spaces with such actions by estimating long-horizon multi-robot action feasibility influenced by deliberate reachability-altering actions. The presented work touches upon a range of challenges spanning policies for long-horizon tasks, scene representations, rollout scoring, latent representations/world models, multi-robot coordination, rare transitions.

Critical strengths

The presented work touches upon a range of challenges spanning policies for long-horizon tasks, scene representations, rollout scoring, latent representations/world models, multi-robot coordination, rare transitions.

Critical weaknesses

Unfortunately, there are several questions about the generalizable insights that may be derived from the paper, beyond the communication of a working example. Namely, several subsets of the investigated challenges by themselves are challenges -- what makes the specific combination the focus of the paper and have the authors ablated the subsets (for example, single robot cases etc); multi-robot team composition, robot capabilities, scalability are all common perspectives in multi-robot problems; the number of actions available or affordances and which ones pose more challenges or how new actions can be accurately classified or composed into new systems are of interest beyond the specifics of the currently presented system. The closed loop results even for the specific system (3-robot team) is somewhat underwhelming. The lack of general observations is the key weakness. More details of the method and problem formulation and task will benefit the exposition.

Questions for authors

Specific comments and questions follow.

In the abstract the authors refer to "facility status". This is not clear at this point. "Legality" is interchangeably with "feasibility". Is there a difference?

Introduction: It should be "later executable plans" is a natural consequence of long-horizon tasks. The authors have not made a convincing argument why this aspect appears to be unique in their proposed strategy of investigation.

The paragraph with three practical difficulties goes into specifics of the authors' proposed study and formulation, rather than generalizable observations. This transition has not been made yet. For e.g., what is "menu"?

A distinction is made between "temporary menu indices" vs "physical semantics". In "receive value through the plans they enable later", it is not clear at this point how plans are enabled or disabled.

"Rare feasibility-changing events" -> do the authors refer to actions/transitions?

The problem formulation is not general but a specific description of a 3-robot system.

Sec 3.1 should be amended to be more consistent with the multi-robot problem. A plan refers to a sequence of actions from the current timestep and spanning some number of future timesteps (horizon H). A multi-robot action would be an action assigned per robot for a timestep, yielding a multi-robot coordinated plan.

The way the authors frame \mathcal{M}_t as a "complete joint plan" is inconsistent with a plan needing to be defined over multiple steps. Assuming joint in "joint plan" refers to a multi-robot, then \mathcal{M}_t at timestep t is closer to one multirobot action. This makes the sequence of sets used to define \mathcal{F} somewhat ill posed. You can have a set of H-length plans, say 64 such plans. This set will be in the object \mathcal{F} defined. However, the possible plans encoded in \mathcal{F} will be |\mathcal{M}_t|^H = 64^H.

The choice of 64 should not be commmunicate as a part of problem formulation. It is not a general property. It is also unclear why and how such a number can be chosen, especially if the team size and number of actions increases.

Legal joint-plan interface: What are slot indices?

The relationship between Sec 3.1. to the exploration task or actions such as "opening a door" or "rare" events is unclear.

Sec 3.3: "For example, opening a door can be preferred because the imagined rollout predicts that a teammate gains access to a previously unreachable region, even when the opening action itself contributes no coverage." -> It is unclear what the importance of no coverage is. In the specific team being studied, G1 can open the door and also cross into the newly opened space, i.e., cover it. The motivation makes sense if a) the heterogeniety forces coordination, i.e., the robot that can open a gate at the base of stairs cannot climb the stairs but the robot that can climb stairs cannot open the gate. b) a long-horizon element where there is an inner atrium with multiple doors and a door-opening robot proceeds to open all the doors in the beginning but the exploration of the rooms may happen over multiple steps into the future. The authors can improve how their setting and USP is motivated to be amenable to their contributions.

The authors describe a very specific "vocabulary" of the action set. It is unclear by what the authors refer to be "interactions that are rarely encountered during ordinary exploration"? The action set is part of the underlying MDP the policy learns over. Some parts of the state space express feasible actions like "CLIMB STAIRS", "PRESS BUTTON" etc. Some actions like "PRESS BUTTON" imply that most states in the state space will not have such an affordance. The problem of learning over such unbalanced transition systems is a common problem in long-horizon planning (learning or other classical methods like PDDL-based AI planning). Why is this being presented as a novel problem by the authors in the context of this specific 3-robot team and for "future feasibility"?

Sec 4. What exactly is the task? The authors refer to a "cup". A purely exploration task may be very different from an object search task. It would be useful to describe the task(s) and benchmarks.

My understanding is that the closed loop numbers are underwhelming. In conjunction with the lack of a strongly motivated problem and methodological description, this raises doubts about the practical efficacy of the system itself.

While some of the questions about generalizability are understandable trade-offs in exposition and brevity; too many of them are missing to make this a strong draft in its current form. The paper will substantially benefit from a set of discussions and arguments about that discuss the general problem.

2 comments

Thank you for the detailed feedback. We address each point below; revisions are marked blue in the manuscript, with full protocols/results in Appendix F and Supplement Sec. 11.

1. Contribution. LIFT learns a changing team-action interface: an interaction by one robot changes later choices for its teammates. Candidate validity is explicitly predicted alongside state/reward and used in imagined actor training.

2. Team size. Added checks give G1 alone 3/4, G1+Go2 4/4, three robots 4/4, and three robots without G1 navigation 4/4. These are padded subsets of the existing architecture, not evidence of arbitrary-size scaling or indispensable heterogeneity.

3. Skills. The vocabulary is fixed; admission checks capability, routes, semantics, precedence and conflicts. New skills would require execution/admission contracts and transition data. Pick/place are task-level effects, not learned arm trajectories.

4. Facility/feasibility. Facility state includes door status and elevator floor/mode. Legality means admission under available evidence; modeled feasibility is future availability, not guaranteed physical executability or task solvability.

5. Long-horizon planning. Delayed enabling effects are familiar in planning. Our specific target is the changing bounded proposal interface under partial observation, with descriptors linking temporary slots to robot controls.

6. Menu. Now defined on first use: the finite candidate list for the current decision segment. Motivation precedes implementation details.

7. Semantics. A slot can change meaning after replanning; its descriptor carries controls, targets and bindings. We corrected the text: RSSM uses both slot identity and descriptor, so invariance to reordering is not established.

8. Rare events. These are interaction-caused transitions, e.g., OPEN changes a facility state, rather than new action classes. Rarity is relative to the collected exploration trajectories.

9. General formulation. Sec. 3.1 now starts with N robots, observation/action history, a maintained snapshot and admissibility; three robots and K=64 are implementation choices.

10. Joint plan. A segment action is at=(ut1,…,utN)a_t=(u_t^1,\ldots,u_t^N). An H-step plan is (at,…,at+H−1)(a_t,\ldots,a_{t+H-1}), with later decisions generally observation-dependent. The implemented "joint plan" assigns all robots for one segment.

11. Future feasibility. Mt=PK(bt)⊆L(bt)M_t=P_K(b_t)\subseteq L(b_t), where L is the admitted action set. Future feasibility is the conditional distribution of future admitted sets given history, the chosen action and a continuation policy. LIFT approximates proposed sets with descriptors/masks; it samples latent trajectories, not all KHK^H action sequences.

12. Why 64. It is a tensor/output budget. On 50 matched starts, K64 completes 44/50; K32/16/8 each complete 45/50. Cups worsen as doors improve. All arms retain 64 descriptors and restrict selectable prefixes; this neither optimizes K nor measures smaller-planner speed.

13. Slot index. It is a temporary row in the padded candidate table; unused rows are masked. Selecting a row dispatches its complete team assignment, not a stable global action ID.

14. Exploration link. A facility interaction can admit later cross-floor movement for a teammate. Reward measures task/exploration progress, not the number of menu entries.

15. Heterogeneity example. We replaced the door example with G1 operating an elevator for a wheeled robot that cannot manipulate or climb stairs. This motivates capability complementarity without claiming universal elevator superiority.

16. PDDL/rare-event novelty. Neither enabling effects nor imbalanced visitation is new. The contribution is their concrete learned interface: semantic descriptors, future-validity prediction and imagined masking, supported by directed experience/event history.

17. Tasks. General tasks require coverage; doors add opening/inspection; elevators add ordered operations/floor visits; cups add search or task-level manipulation. Coverage alone does not replace these success predicates.

18. Evidence. Across three seeds on 50 reused starts, Full coverage is 95.90% versus 94.87%/94.69% for mask-off/loss-and-mask-off; completion is similar (128/127/129 of 150), with mixed task/seed effects. Separately, original Full's validation invalid AP at h=5/10 is 0.499/0.723 versus slot-prior 0.268/0.390, with lower Brier errors. Primary h=1 AP is below that prior; h=10 has 15 shared windows. Future descriptors are predicted, but recorded action slots are supplied. This supports multi-step menu prediction, not causal control benefit. Historical PPO results remain competitive.

I thank the authors for their response.

At the time of writing this comment, I do not note any highlighted changes in the main draft. This makes it difficult to track whether or what has changed in the main draft. I do note changes in the supplementary document.

Moreover, the primary concern about generalizable scientific insights remains. The information provided in the response concisely inform or define specific points but a broader discussion and clarification is required. The tests over subsets of the team are a good start.

Overall, I cannot recommend a change of my score at this point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.