betaFeedback

WorksComputer Science

Racing the Clock: Teaching VLNs Budget-Conditioned Object Search

Abstract

A rescue robot entering a building with thirty seconds on the clock should sprint toward the single most promising room; given ten minutes, it should instead explore a broader tour. Yet vision-language-navigation (VLN) models are typically evaluated under a fixed episode limit that the policy never observes, even though time-critical missions require behavior to adapt to the available budget. We study budget-conditioned object search: given a target category and an explicit action budget, an agent must find as many distinct target instances as possible before the budget expires. We investigate how to supervise VLNs for this task by comparing a privileged TSP expert, which plans near-optimal routes using the full map and true target locations, with a room-level Bellman expert that maximizes expected return under partial observability using high-level cross-scene semantic priors. Despite the TSP expert achieving substantially higher expert performance, its demonstrations are less effective for learning exploration: on held-out HM3D scenes, the Bellman-supervised VLN consistently outperforms the TSP-supervised VLN, with the advantage becoming more pronounced at longer budgets. At 2000 steps (500m), the Bellman-supervised VLN finds 71.4% of reachable target instances, compared with 65.5% for the TSP-supervised VLN and 44.8% for the pretrained VLN. These results highlight a distinction between optimal routing and optimal exploration: effective exploration supervision should optimize decisions under partial observability rather than imitate routes that are optimal only once the environment is fully known.

Introduction

An indoor search robot must allocate its budget between nearby targets and distant regions: a short budget may favor immediate collection, while a longer one may justify travel to a promising room. To assess both effective budget use and its influence on decisions, we first ask: how does the available budget change where and how a vision-language navigation (VLN) policy searches?

This motivates budget-conditioned object search: finding as many distinct instances of a requested category as possible within an action budget. The agent receives RGB history, a target instruction, and a remaining-budget countdown, but no complete map or unseen target locations. Travel, scanning, and target approach consume budget. We count an instance as collected when it meets geometric proximity and camera-field-of-view criteria, without physical pickup. Distinct discoveries provide a denser signal than binary success and preserve repeated choices between visible targets and further exploration. Collection across budgets measures search performance; controlled countdown interventions test behavioral sensitivity (Figure 1).

Room-level Bellman-supervised Uni-NaVid trajectories under different stated budgets

Figure 1. The stated budget changes the learned VLN’s search strategy. Room-level Bellman-supervised Uni-NaVid is rolled out from the same start pose in the same held-out HM3D scene with identical deterministic decoding; only the budget in the instruction changes. The resulting trajectories differ substantially rather than simply extending one another, most visibly between B=150B = 150 and B=400B = 400. Across B=50/150/400/800/2000B = 50/150/400/800/2000, the policy collects 4/11/10/12/12 of 13 reachable chairs.

The remaining horizon can change the optimal action despite identical observation histories [1]; omitting it creates state aliasing between short-horizon collection and longer-horizon exploration. Although finite-resource planning is well studied, pretrained VLNs such as Uni-NaVid [2] lack an explicit remaining-budget countdown. In a NaVILA [3] pilot, a shorter stated limit sometimes produced more actions than a longer one, motivating budget-aware demonstrations beyond merely adding a prompt limit.

To assess what supervision best transfers to a partially observed student, we next ask: what kind of expert demonstrations effectively teach budget-conditioned search? A full-map expert can route efficiently yet prescribe different actions in environments indistinguishable to the student, an information mismatch that can hinder asymmetric imitation [4, 5, 6, 7]. Prior work addresses this mismatch by modifying the imitation objective or by adapting the expert to the student. We ask a complementary question: what environmental information should the planning expert itself retain when generating exploration supervision?

Existing exploration methods span different information regimes. Experts with full map and target-location access [8, 9] reduce search to routing among known goals, while receding-horizon planners [10, 11, 12, 13] act from currently available information but may provide weak long-horizon guidance. Scene-prediction methods inject priors about unseen regions [14, 15], but predicting detailed unseen geometry and semantics is itself challenging.

We make this information choice explicit. An expert plans over an abstraction z=φ(M)z = \varphi(\mathcal{M}) of the full environment M\mathcal{M} and knows only the part of zz revealed to it so far; one budget-indexed Bellman recursion over this information state then defines the expert. Varying φ\varphi spans a spectrum of information access. Revealing every target location yields an object-level Bellman expert that routes among known targets: nothing is left to search for. Retaining observed room connectivity, doorway evidence, and cross-scene semantic priors while withholding the contents of unexplored rooms yields a room-level Bellman expert, for which each unentered room is a search opportunity with uncertain yield and delayed reward. Withholding unseen contents, not the room granularity itself, is what turns room selection into search. It also determines what the student is asked to imitate: object-level actions can depend on targets the student has not seen, whereas room-level actions depend only on information the student could acquire by exploring. Multi-room search preserves the key choice between a visible target and a room with uncertain, delayed reward.

Equal target values and unit primitive-action costs isolate these questions from target priorities and cost weights, while collection and uncertain exploration still compete for finite budget. Our focus is supervision and budget use within a fixed video VLN architecture: both Bellman experts supervise Uni-NaVid students that predict primitive actions from RGB history and language, and a greedy frontier expert and its supervised VLN provide coverage-oriented comparisons. Our goal is not to replace explicit-map exploration methods such as frontier planning; we use them as references for what explicit spatial memory achieves.

All three expert-supervised VLNs use matched scenes, budgets, and record counts; the two Bellman students also use identical optimization. Room-level supervision achieves the highest mean collection at every budget in ten held-out HM3D scenes despite lower expert return. Budget-blind controls distinguish countdown gains from supervision-source differences; countdown interventions test behavioral sensitivity. These findings motivate recording demonstration budgets and considering expert search behavior when choosing supervision.

Our contributions are:

  • We formulate budget-conditioned object search, using distinct target collection to measure effective budget use and controlled countdown interventions to assess behavioral sensitivity to the budget.

  • We formalize expert design as a choice of abstraction φ\varphi of the full environment. One budget-indexed Bellman recursion then spans fully informed routing and search under uncertainty, and the abstraction makes explicit which information enters the expert’s actions and must be imitated by a partially observed student.

  • We instantiate a room-level Bellman expert that withholds the contents of unexplored rooms and estimates this recursion from strictly cross-scene transitions, providing budget-conditioned search demonstrations for a VLN.

  • We compare room-level, fully informed object-level, and greedy-frontier supervision for Uni-NaVid on matched scenes, budgets, and record counts. Room-level supervision yields the highest mean collection, while higher expert return need not yield a stronger student.

Privileged and Planning Supervision

A high-performing expert need not provide optimal supervision to a student with partial information. ADVISOR [4], A2D [5], and COSIL [16] reduce reliance on privileged actions or combine imitation with learning under partial observability; recent work studies student realizability and teacher–student alignment [6, 7]. ADVISOR and A2D characterize the resulting imitation gap through the expert’s actions marginalized over information the student cannot observe, the quantity we use in Sec. IV-C. Privileged planning also supervises navigation in Learning by Cheating [17], shortest-path imitation in SPOC [8], and large-scale exploration [18]. These approaches modify the imitation objective or adapt the expert to the student. We instead vary the information available to the planning expert itself through its scene abstraction, and study how the resulting supervision transfers to a fixed video VLN, comparing fully informed object-level routing with room-level search under uncertainty. The two Bellman-supervised students share architecture and training settings. Greedy-frontier supervision provides an additional comparison with coverage-oriented exploration demonstrations.

Large-scale imitation and reinforcement learning support policies such as ViNT and LM-Nav [19, 20], SPOC [8], and video VLNs NaVid [21], Uni-NaVid [2], and NaVILA [3]. Object search exploits geometry and semantics [22, 23, 24, 25, 26], with POMDPs modeling uncertainty over object locations [27, 28]. Exploration uses rooms and doorways [29], incremental frontiers and hierarchical planning [30], or information gain under mapping uncertainty [31]; cooperative systems further combine exploration, mapping, and inspection in unknown environments [32]. MultiON [33], GOAT-Bench [34], and long-horizon VLN [35] extend task horizons. Whereas object search commonly tests single-target success under a fixed limit [36], we measure distinct discoveries and search-strategy changes across budgets. Our comparisons examine supervision for a video VLN; explicit planning methods supply demonstrations and reference trajectories.

Budgeted and Budget-Conditioned Exploration

Finite-resource decision making is classical in orienteering [37], constrained MDPs [38], and informative path planning. ning, where trajectories trade information gain against limited time, distance, energy, or other resources [10, 12, 39, 40, 41, 42, 43]. Our object-level Bellman expert solves an orienteering-style routing problem over known targets; the room-level expert instead plans over expected yields of unexplored regions. Recent planners use predicted maps and pathwise information gain to reason about unseen regions along candidate trajectories [13]. Resource dependence has also been studied directly in learned exploration policies [44], alongside adaptive and frontier-based exploration [45, 46, 47].

Decision Transformer [48] and reward-conditioned or constrained policies [49, 50] condition on returns, horizons, or constraints. Time Limits in RL [1] shows that omitting the remaining horizon can alias states requiring different optimal actions; navigation methods likewise incorporate budget, return, or termination signals [51, 52, 53, 54, 55]. We focus on exploration: which regions to search and in what order as the budget changes, and which expert supervision transfers this behavior to a video VLN.

Budget-Conditioned Object Search

Problem formulation

An episode e=(M,q0,y,B)e = (\mathcal{M}, q_0, y, B) specifies a multi-room environment M\mathcal{M}—its complete geometry, semantics, and object instances—together with the initial agent state, target category, and action budget. Rooms organize exploration; the objective counts object instances and the policy selects primitive actions. The budget can also represent time, distance, energy, or another resource.

At step tt, the policy π(at∣ht,y,bt)\pi(a_t \mid h_t, y, b_t) selects at∈Aa_t \in\mathcal{A} from observation–action history ht=(o0,a0,…,ot)h_t = (o_0, a_0, \ldots, o_t) and remaining budget bt=B−∑k=0t−1c(ak)b_t = B - \sum_{k=0}^{t-1} c(a_k), with positive action cost c(at)c(a_t). Here π\pi denotes any policy using these inputs; our learned instances are Uni-NaVid students trained on expert actions (§ 5-C). Observations reveal M\mathcal{M} only partially, so different environments may be indistinguishable under the same history. M\mathcal{M} is therefore privileged: the simulator uses it to credit discoveries, experts may access it only through the abstraction their design permits (§ 4), and the learned policy never observes it.

Let Iy(M)\mathcal{I}_y(\mathcal{M}) contain the category-yy instances and Ft⊆Iy(M)\mathcal{F}_t \subseteq\mathcal{I}_y(\mathcal{M}) those found by time tt under a policy-independent criterion. An instance is collected when it is first credited by this criterion (§ 5-A). The reward rt=∣Ft+1∖Ft∣r_t = \lvert\mathcal{F}_{t+1} \setminus\mathcal{F}_t \rvert counts only newly credited instances after action ata_t; revisiting a credited instance adds no reward. The objective is

πB⋆∈arg⁡max⁡π Eπ[∑t=0τ−1rt]s.t.∑t=0τ−1c(at)≤B,(1)\pi_B^\star\in\underset{\pi}{\arg\max}\ \mathbb{E}_{\pi}\left[\sum_{t=0}^{\tau-1} r_t\right] \qquad\text{s.t.} \qquad\sum_{t=0}^{\tau-1} c(a_t) \le B, \tag*{(1)}

where τ\tau denotes the termination time.

Changing BB changes the decision problem rather than merely the point at which a fixed trajectory is truncated. Under a tight budget, actions with high immediate expected reward may be preferred; with more budget, the agent can invest in exploration whose reward is realized only later.

The optimal action therefore depends jointly on what has been observed and how much budget remains, and omitting btb_t from the policy input aliases states that require different actions [1].

Heterogeneous target values replace the count with rt=∑i∈Ft+1∖Ftwir_t = \sum_{i\in\mathcal{F}_{t+1}\setminus\mathcal{F}_t} w_i, wi≥0w_i \ge0; Eq. (1) already permits nonuniform positive costs. We use wi=c(at)=1w_i = c(a_t) = 1. Extensions require corresponding expert reward/cost estimates and policy inputs conveying the task specification. They preserve our research questions, but persistence of the supervision advantage requires further evaluation.

Budget-conditioned evaluation

Standard single-object search provides a sparse success signal: once one instance is found, the episode counts as successful regardless of how effectively the agent explored. Counting distinct discoveries as a function of the available budget instead distinguishes policies beyond single-target success and provides a denser evaluation of budget use.

For an episode ee, let Ry(e)\mathcal{R}_y(e) denote the set of target instances considered valid and reachable under the benchmark’s embodiment and discovery criterion. For a fixed evaluated policy, Fe(B)\mathcal{F}_e(B) is the set of distinct instances credited by termination of the rollout with initial budget BB, i.e., Ft\mathcal{F}_t at that rollout’s termination time. Changing BB can change both the trajectory and its termination time, hence the discovered set. We define normalized collection, the fraction of valid reachable instances found, as

Ne(B)=∣Fe(B)∩Ry(e)∣∣Ry(e)∣,(2)N_e(B) = \frac{\lvert\mathcal{F}_e(B) \cap\mathcal{R}_y(e) \rvert}{\lvert\mathcal{R}_y(e) \rvert}, \tag*{(2)}

and report its average over the set EB\mathcal{E}_B of matched evaluation episodes run with budget BB,

N(B)=1∣EB∣∑e∈EBNe(B).(3)N(B) = \frac{1}{\lvert\mathcal{E}_B\rvert} \sum_{e\in\mathcal{E}_B} N_e(B). \tag*{(3)}

Higher collection with more budget does not establish use of the budget input: a fixed strategy may find more objects simply by running longer. We therefore hold scene, target, start, model, and decoding fixed, vary only the countdown, and examine trajectory divergence within the shared horizon. These interventions test behavioral sensitivity; N(B)N(B) measures performance.

Abstract-State Bellman Experts

A strong search expert can derive its advantage in two different ways: by reasoning effectively from information that the learned policy can also acquire, or by exploiting information that the learned policy will never observe. The distinction matters for budget-conditioned search, where demonstrations should teach not only where to go but how to trade immediate collection against exploration whose payoff is uncertain. If an expert observes the complete environment, including unseen target locations, object search collapses to routing: the expert knows where the rewards are and only needs to reach them efficiently. Such trajectories can be near-optimal, yet they never demonstrate a decision made without knowing what lies behind a door. We therefore treat the expert’s information as an explicit design choice. Both of our Bellman experts optimize the same budgeted objective; they differ in the abstraction of the environment they plan over and in how much of it is revealed to them.

Bellman Planning over Scene Abstractions

Recall that M\mathcal{M} denotes the complete, privileged environment (§ 3). An expert does not plan over M\mathcal{M} directly, but over an abstraction

z=φ(M),(4)z = \varphi(\mathcal{M}), \tag*{(4)}

where φ\varphi is a design choice specifying which task-relevant structure of M\mathcal{M} is retained, such as the locations of all target instances or a room graph with room types and contents, and which details are discarded. Like M\mathcal{M}, zz is a property of the environment rather than of what has been observed, and is therefore privileged. At time tt, the expert’s information state sts_t summarizes the part of zz revealed to it so far, together with the robot state and the set Ft\mathcal{F}_t of credited instances. Two choices thus define an expert’s information: what φ\varphi retains, and what sts_t reveals.

Let p(M)p(\mathcal{M}) denote the distribution of environments and p(M∣s)p(\mathcal{M}\mid s) the distribution over environments consistent with information state ss. Because multiple environments can remain consistent with sts_t, the expert evaluates actions through rewards and transitions averaged under this distribution,

Rφ(s,a)=EM∣s[rt∣M,st=s,at=a],(5)R_\varphi(s,a) = \mathbb{E}_{\mathcal{M}\mid s}\left[r_t\mid\mathcal{M},s_t=s,a_t=a\right], \tag*{(5)}
Pφ(s′∣s,a)=EM∣s[Pr⁡(st+1=s′∣M,s,a)],(6)P_\varphi(s'\mid s,a) = \mathbb{E}_{\mathcal{M}\mid s}\left[\Pr(s_{t+1}=s'\mid\mathcal{M},s,a)\right], \tag*{(6)}

where EM∣s\mathbb{E}_{\mathcal{M}\mid s} denotes expectation over M∼p(M∣s)\mathcal{M}\sim p(\mathcal{M}\mid s). Here aa may be a temporally extended action, such as visiting a target or entering and searching a room, with c(a)c(a) its total primitive-action cost and rtr_t the reward accumulated over its primitive steps. For remaining budget bb, the corresponding

Bellman recursion is

Qφ∗(s,b,a)=Rφ(s,a)+Es′∼Pφ(⋅∣s,a)[Vφ∗(s′,b−c(a))],(7)Q_\varphi^*(s,b,a)=R_\varphi(s,a)+\mathbb{E}_{s'\sim P_\varphi(\cdot\mid s,a)}\left[V_\varphi^*(s',b-c(a))\right], \tag*{(7)}
Vφ∗(s,b)=max⁡a:c(a)≤bQφ∗(s,b,a),Vφ∗(s,0)=0.(8)V_\varphi^*(s,b)=\max_{a:c(a)\le b}Q_\varphi^*(s,b,a),\qquad V_\varphi^*(s,0)=0. \tag*{(8)}

Both of our Bellman experts implement Eqs. (7)–(8). They differ in φ\varphi, in what sts_t reveals, and therefore in the distribution p(M∣st)p(\mathcal{M}\mid s_t) that the expectations average over.

A Spectrum of Abstractions

The choice of φ\varphi and of what sts_t reveals determines the uncertainty the expert must reason over, and spans a spectrum of information access. Table I contrasts the two Bellman experts with the student they supervise.

Object-level expertRoom-level expertVLN student
Abstraction zzmap and all target instancesroom graph with room types and contentsnone; raw RGB
Revealed in sts_tall of zz, from t=0t=0observed rooms and connectivity, discovered objects, doorway evidenceRGB history hth_t and countdown btb_t
Withheld—contents of unentered roomseverything unseen
Planningdeterministic routing, Eq. (9)expected return from cross-scene transitions, Eq. (11)none; imitates expert actions
Decision unitnext target instancenext frontier roomnext four primitives
Actions depend on unseen targetsyesnono

Table I. Information available to each Bellman expert and to the student they supervise. M\mathcal{M} is the full environment, z=φ(M)z=\varphi(\mathcal{M}) the expert’s abstraction, and sts_t the part of zz revealed at time tt.

Revealed targets: object-level routing. If φ\varphi retains the map and every target instance, and zz is revealed in full at t=0t=0, then sts_t determines all future rewards. The expectations in Eqs. (5)–(6) collapse, and the recursion can be written as deterministic budgeted routing over known targets, an orienteering problem [37]:

Vobj(s,b)=max⁡(0,max⁡i∈Kb(s)[1+Vobj(s⊕i,b−κi)]),(9)V_{\mathrm{obj}}(s,b)=\max\left(0,\max_{i\in\mathcal{K}_b(s)}\left[1+V_{\mathrm{obj}}(s\oplus i,b-\kappa_i)\right]\right), \tag*{(9)}

where κi=κ(s,i)\kappa_i=\kappa(s,i) is the primitive-action cost of reaching a crediting pose of instance ii from ss, Kb(s)={i∈Iy(M)∖F(s):κi≤b}\mathcal{K}_b(s)=\{i\in\mathcal{I}_y(\mathcal{M})\setminus\mathcal{F}(s):\kappa_i\le b\} contains the uncredited instances reachable within budget bb, F(s)\mathcal{F}(s) is the set of credited instances, and s⊕is\oplus i is the state after crediting ii. Decisions concern which known instance to visit next; “fully informed” refers to knowledge of every target location, including those outside the camera’s view. This is our object-level Bellman expert: it has nothing left to search for.

Observed geometry only: coverage exploration. At the other extreme, if φ\varphi retains only geometry and the expert has no informative prior on what unexplored regions contain, it has little basis for distinguishing them and plans essentially for coverage, as in frontier and receding-horizon exploration [46, 11]. Such planners act only on information a student can also acquire, but give weak guidance about where distant rewards are likely to be.

Intermediate abstractions: search with priors. Between these extremes, φ\varphi can retain coarse structure of unseen regions, namely their connectivity and likely semantics, while sts_t withholds their realized contents. Actions are then valued by expected return over environments consistent with the current information, with p(M∣st)p(\mathcal{M}\mid s_t) supplying cross-scene regularities such as which room types tend to contain the target category. An action with little immediate reward may still be optimal if it reaches a promising region or resolves uncertainty that enables future reward, and whether it is worthwhile depends on the remaining budget bb. Exploration can thus be viewed as acting to resolve the task-relevant uncertainty left by the chosen abstraction. Our room-level Bellman expert (§ 4-D–IV-F) occupies this regime.

Granularity versus information access. These are distinct axes. Planning over rooms does not by itself create uncertainty: a room-level expert given the realized contents of every room would face a deterministic problem like Eq. (9), with rooms rather than instances as decision units. It is withholding unrevealed contents that turns room selection into search. Our two Bellman experts differ along both axes, which our experiments do not separate (Sec. VI).

Abstraction and Expert–Student Information Mismatch

Each expert πE(a∣st,bt)\pi_E(a\mid s_t,b_t) supplies action labels for a separate Uni-NaVid student πθ(a∣ht,y,bt)\pi_\theta(a\mid h_t,y,b_t) trained with the same imitation objective. At inference, students act from RGB history and language, including the countdown, without querying an expert or running its planner. With a cross-entropy imitation loss on expert-visited states, the best student in an unrestricted policy class matches the expert’s actions marginalized over the expert information that its own history does not determine [4, 5]:

πθ∗(a∣ht,y,bt)=Est∼p(⋅∣ht,y,bt)[πE(a∣st,bt)].(10)\pi_\theta^*(a\mid h_t,y,b_t)=\mathbb{E}_{s_t\sim p(\cdot\mid h_t,y,b_t)}\left[\pi_E(a\mid s_t,b_t)\right]. \tag*{(10)}

When sts_t is determined by hth_t, the student can reproduce the expert. Otherwise it learns a mixture of the expert’s actions across environments it cannot distinguish, and for budget-conditioned search such a mixture need not be a coherent search strategy. For example, from an identical start view the object-level expert may turn left in one scene and right in another depending on where unseen targets lie; the student receives identical inputs and can only learn their average.

The abstraction controls which information enters this marginalization. The object-level expert’s sts_t contains every target location from t=0t=0, so Eq. (10) averages routes toward targets the student has not seen. The room-level expert withholds unrevealed contents from its own decisions, so its actions depend only on information a student could in principle acquire by exploring or learn as a prior from training scenes. A residual mismatch remains: the room-level expert maintains an explicit room graph and reads observed room types and objects from simulator annotations, which the student must infer from RGB history. Rather than adapting the learning objective or the expert to a particular student, we use φ\varphi to control the expert’s information, seeking an abstraction that retains enough structure for long-horizon planning while requiring the expert to reason over uncertainty that the student must also resolve through exploration.

This argument motivates the room-level design but does not predict the outcome: a better-aligned expert may also be a weaker one, as the object-level expert’s higher return shows (§ 5-B), and the two experts differ in more than information access. § 5-C tests the resulting supervision empirically.

Room-Level Instantiation

Room entry incurs travel and search costs, reveals targets, and may expose further regions; semantics provide target-yield priors. This captures the collection–exploration tradeoff without requiring a room representation in the task objective or student’s action space. Figure 2 shows entered rooms and adjacent unentered candidates, evaluated by expected yield and the budget needed to reach and search them.

Figure 2. Room-level Bellman expert. (a) “Known” rooms have been entered; their recorded contents include only discovered objects. (b) Frontier rooms are valued using doorway-visible semantic evidence and cross-scene Bellman estimates. (c) Execution remains closed-loop: visible targets can interrupt a sweep, after which the expert replans.

Our room-level map M\mathcal{M} is a room graph with room types and contents. The information state sts_t reveals only its observed part, withholding the geometry, contents, and target locations of unexplored rooms: it contains the observed room graph Gtobs=(Vtobs,Etobs)G_t^{\mathrm{obs}}=(V_t^{\mathrm{obs}},E_t^{\mathrm{obs}}), robot state, observed room types and discovered contents, and frontier-doorway evidence. Observed room types and objects come from simulator ground-truth annotations, exposed as rooms are entered or objects become visible, rather than inferred from RGB. Entering a room does not reveal its complete object inventory; only visible objects become available for direct approach.

The frontier set U(st)\mathcal{U}(s_t) contains adjacent unentered rooms on the current floor with a doorway reachable along a finite geodesic. An unentered room may have objects visible through its doorway; this limited evidence does not reveal its complete contents. Such objects provide a room-type estimate through the cross-scene prior P(ρ∣o)P(\rho\mid o) learned from Matterport3D [56], where ρ\rho is room type and oo is doorway-visible object evidence. The expert uses this estimate to predict discoveries in the unobserved interior. An estimated full sweep exceeding the budget does not exclude a frontier; the budget instead filters transition outcomes below.

Cross-Scene Bellman Estimation

We estimate the unknown reward and transition models from a bank of room-transition records from 181 HM3D scenes. Each record ξi=(si,vi,ri,di,si′)\xi_i = (s_i, v_i, r_i, d_i, s_i') stores the information state before selecting frontier room viv_i, the newly credited target counts rir_i, the realized primitive-action cost did_i, and the information state after execution. Thus, riyr_i^y is a room-transition reward for category yy, accumulating the primitive-step rewards defined in § 3. These outcomes supply empirical samples for Eqs. (7)–(8); the doorway prior supplies the candidate’s estimated room type, not its realized reward. This bank supports expert planning; the video/action records used to train the VLN are described in § 5-C.

For a query (s,v)(s,v), we retrieve transitions with matching abstract context–action signatures and use weighted nearest neighbors when exact matches are sparse. Retrieval is strictly cross-scene. The source scenes are split into five folds, and planning in a scene excludes every transition from that scene’s entire fold. All benchmark scenes are excluded from the evaluation transition bank. Each retrieved transition is thus one outcome of the same abstract action in another environment consistent with the same information state, i.e., an empirical sample from p(M∣s)p(\mathcal{M} \mid s) in Eqs. (5)–(6). Strict cross-scene retrieval ensures that these samples never come from the realized contents of the scene being planned: the prior is learned from other environments rather than read off the current one.

Retrieved transitions jointly sample reward, execution cost, and successor context. Let Nb(s,v)\mathcal{N}_b(s,v) contain those with di≤bd_i \le b, with weights αi≥0\alpha_i \ge0 normalized to ∑i∈Nb(s,v)αi=1\sum_{i\in\mathcal{N}_b(s,v)} \alpha_i = 1. For category yy, the empirical backup combines immediate reward and continuation return:

Q^y(s,b,v)=∑i∈Nb(s,v)αi[riy+V^y(si′,b−di)],(11)\hat{Q}_y(s,b,v)=\sum_{i\in\mathcal{N}_b(s,v)}\alpha_i\left[r_i^y+\hat{V}_y(s_i',b-d_i)\right], \tag*{(11)}

where riyr_i^y is the immediate room-transition reward for category yy and V^y(si′,b−di)\hat{V}_y(s_i',b-d_i) estimates the best subsequent return after spending did_i steps. The state value selects the highest estimated return over frontier rooms and a zero-valued stopping option:

V^y(s,b)=max⁡(0,max⁡v∈U(s)Q^y(s,b,v)),V^y(s,0)=0.(12)\hat{V}_y(s,b)=\max\left(0,\max_{v\in\mathcal{U}(s)}\hat{Q}_y(s,b,v)\right),\qquad\hat{V}_y(s,0)=0. \tag*{(12)}

The hats denote estimates computed recursively under the empirical model. Filtering by di≤bd_i \le b and renormalizing conditions on outcomes that fit the budget; it does not model the probability or reward of exceeding it. A room with no admissible transition has undefined Q^y\hat{Q}_y and is excluded from the maximization, but remains eligible for the fallback below.

Continuation enables non-myopic exploration: a low-reward connector room may expose promising regions, yet be unhelpful under a tight budget. Positive transition costs permit evaluation in increasing order of bb.

Doorway evidence thus estimates room type; retrieved outcomes turn this semantic expectation into reward, cost, and continuation estimates. A room’s ranking depends on budget-dependent returns, not its semantic promise alone.

Closed-Loop Execution

Primitive execution is closed-loop: a newly visible uncredited target within 5 m and line of sight interrupts the room sweep for approach-and-collect. The executor uses observed targets and coverage to determine when its room search is complete; it does not consult a ground-truth room inventory or target count. Figure 2(c) shows a mid-sweep detour and departure with most sweep viewpoints unvisited. Alternating scan directions at successive viewpoints reduces the left:right turning ratio from 2:1 to 1.2:1.

If no frontier has positive estimated value or admissible evidence, the expert enters the nearest unexplored frontier as a curiosity fallback. The zero branch in Eq. (12) is a value-estimation option, not an execution rule requiring termination when positive evidence is absent. After every room transition or interruption it updates sts_t and replans. If all frontiers are exhausted while budget remains, it re-sweeps visited rooms, providing demonstrations at horizons longer than required to explore the scene.

Unexplored contents and target locations are used only after execution to construct transitions, never for action selection. The abstraction merges distinct histories and the model uses finite cross-scene data, so this is an empirical room-level Bellman expert, without a Bayes-optimality claim for the underlying partially observable search problem. In the notation of § 4-A, the expert conditions on the abstract state sts_t rather than on its full history, so p(M∣st)p(\mathcal{M}\mid s_t) can differ from the posterior given that history, and the expectations in Eqs. (5)–(6) are approximated from finite cross-scene samples.

Experiments

We test budget-dependent VLN behavior through collection scores and countdown interventions, and supervision transfer through matched training comparisons. Experts and learned policies share the objective, budgets, and discovery criterion. The two Bellman experts also share the budgeted optimization principle, applied to known targets or expected room yields according to their available information (§ 4); greedy frontier provides a coverage-oriented baseline. We examine demonstrations before learned policies, using the held-out HM3D protocol below.

HM3D Experimental Setup

Environment and action space. We instantiate budget-conditioned object search in multi-room HM3D [57] indoor scenes in Habitat [58] with A={FORWARD,LEFT,RIGHT,STOP}\mathcal{A} = \{\mathrm{FORWARD}, \mathrm{LEFT}, \mathrm{RIGHT}, \mathrm{STOP}\}, where FORWARD moves 0.25 m and LEFT/RIGHT rotate 30∘30^\circ. Each primitive costs one budget unit, and episodes terminate on STOP or budget exhaustion. We evaluate B∈{150,400,800,2000}B \in\{150, 400, 800, 2000\}. Learned VLNs receive the remaining budget only through language, e.g., “Find as many chairs as possible. You have btb_t steps remaining,” with no dedicated numerical budget input. Their observations consist of RGB video history; the full map, room graph, and ground-truth locations of unseen targets are not provided to them.

Discovery criterion. We define an object as found using a policy-independent geometric referee. A target instance is credited when the agent is within 1.5 m of its bounding extent, within 1.5 m vertically, and its bearing is within 39.5∘39.5^\circ of the camera optical axis. This bound is half of the camera’s 79∘79^\circ horizontal field of view. The criterion measures target coverage within the camera’s nearby viewing region. It is applied identically to every policy and requires no separate recognition report or collection action. Each instance is credited at most once. We exclude instances in the starting room and those without any valid crediting pose geodesically reachable from the start. Reachability is computed directly on the HM3D navmesh rather than the room graph, allowing, e.g., an instance visible through a doorway to be credited without entering its semantic room.

Matched evaluation protocol. We evaluate all methods on 10 held-out HM3D-val scenes excluded from both VLN fine-tuning and room-level Bellman transition data. The benchmark contains 50 admission-verified (scene, category) pairs with two deterministic starts each. Scene, category, initial position, heading, and budget are matched across methods, giving 100 episodes per budget. Table II reports mean normalized collection N(B)N(B) (Sec. III) with the half-width of a 95% confidence interval (CI). CIs are percentile bootstraps over evaluation episodes: for each method and budget we resample the matched episode set 10,000 times with replacement and take the 2.5th and 97.5th percentiles of the resampled mean. The episode is the resampling unit, so scene, category, and start are resampled together. Figure 3(a) draws these intervals as bands and Figure 3(b) as error bars, offset horizontally for legibility. We also examine cross-budget trajectory divergence, revisiting, and monotonicity violations, where a larger budget yields fewer targets.

Figure 3. Expert return and supervision transfer on the frozen benchmark. (a) Expert collection N(B)N(B) with 95% bootstrap intervals: object-level Bellman leads room-level Bellman, which leads unprivileged frontier baselines. (b) Mean collection after matched Uni-NaVid fine-tuning reverses the ordering of the two Bellman experts. Greedy frontier is a non-learned reference; Table II additionally reports greedy-frontier-supervised VLN performance. Budget-dependent margins and long-horizon revisiting are discussed in § 5-C.

Expert Supervision

Expert performance. The object-level Bellman expert solves budgeted routing with the full map and target locations; the room-level expert follows the information constraints in Sec. IV. The greedy frontier expert provides a coverage-oriented exploration baseline with explicit spatial memory. We evaluate both its direct execution and the VLN trained on its demonstrations under the same held-out protocol and action budgets.

The object-level expert leads at every budget, most strongly at tight budgets (Figure 3(a)). Both Bellman experts act until the budget is exhausted: after collecting all targets, the object-level expert re-sweeps the environment, while the room-level expert re-sweeps visited rooms after exhausting frontiers. These continuation behaviors are part of the respective supervision schemes. Re-sweeping ensures every demonstration reaches its assigned budget, controlling the distribution of remaining budgets in training records across experts despite different completion times.

The gap between the two Bellman experts measures what revealing target locations is worth to a planner: 45.6 points at B=150B = 150 (91.3% versus 45.7%), shrinking to 8.2, 2.9, and 2.0 points at B=400/800/2000B = 400/800/2000 (Table II). Knowing where targets are matters most when little budget remains to recover from an uninformed first choice. This advantage is largest exactly where decisions depend most on target locations that the student cannot observe (Sec. IV-C), so expert return alone does not determine how useful demonstrations are for learning.

Budget-dependent demonstrations. High collection alone does not establish that the room-level expert changes its strategy with the available budget. Varying only BB at matched scenes, target categories, and starts, 55% of 691 admitted training pairs select a different first room across budgets; 96% are monotone in targets collected.

At equal executed-action counts, mean cumulative collection for B=150B = 150 leads larger-budget demonstrations for about 131 steps, by up to 7.0 points, before saturating. This controls for additional execution time: tight budgets favor immediate yield, while larger ones justify delayed-reward transfers. Thus, the demonstrations contain budget-dependent decisions rather than truncations of a common route.

Learning a Budget-Conditioned VLN

Fine-tuning setup. Bellman-expert demonstrations are converted to Uni-NaVid’s native training format using its fixed inference cadence: after every two executed primitives, the model is queried for the next four actions. Each record contains the video history up to time tt, the expert’s next four primitives, and a natural-language countdown with B−tB - t steps remaining. The budget is provided only through language; we introduce no numerical budget input or architectural modification.

The room-level Bellman supervision has 388,199 records: 214,562 from 89 HM3D and 173,637 from 59 Matterport3D scenes, at budgets 50, 150, 400, 800, 1200, and 2000. A deterministic 20% of records omit the countdown. We train directly on this mixture without contrast mining or reweighting. For each Bellman supervision scheme, we also train a budget-blind control using the same video/action records, with the countdown removed from every record. Architecture, optimization, and training steps are fixed; only the training countdown text differs. The controls retain budget-dependent trajectories but receive no explicit countdown during training.

We otherwise follow the Uni-NaVid stage-2 recipe: the vision encoder is frozen, the language model and projector are trained, the learning rate is 10−510^{-5}, and training lasts one epoch, 3,032 steps at effective batch size 128, taking approximately 62 hours on four RTX PRO 6000 Blackwell GPUs. The object-level, room-level, and greedy-frontier students use the same scenes, budgets, and record count. The two Bellman students also share architecture and optimization; only their expert supervision differs.

Learned-policy performance. Table II compares experts and learned policies. The non-learned greedy frontier uses explicit spatial memory, whereas VLNs retain history implicitly in context. Its direct execution provides a long-horizon coverage reference; its supervised VLN tests transfer of these demonstrations to RGB history. The central comparison concerns supervision within the Uni-NaVid architecture.

PolicyB=150B = 150B=400B = 400B=800B = 800B=2000B = 2000
(a) Expert policies
Greedy frontier expert24.4 ±6.352.1 ±7.185.8 ±4.693.8 ±2.9
Object-level Bellman expert91.3 ±3.499.8 ±0.2100.0 ±0.0100.0 ±0.0
Room-level Bellman expert45.7 ±9.891.6 ±6.497.1 ±4.298.0 ±3.5
(b) Learned VLN policies
Pretrained VLN28.3 ±6.135.5 ±6.238.8 ±6.644.8 ±7.0
VLN trained on greedy frontier expert25.6 ±6.044.5 ±7.354.6 ±7.568.5 ±7.3
VLN trained on object-level expert28.3 ±6.245.0 ±7.358.8 ±7.565.5 ±7.4
    budget-blind training25.5 ±6.042.2 ±7.255.0 ±7.561.8 ±7.5
VLN trained on room-level expert (ours)31.6 ±6.858.0 ±7.464.5 ±7.172.7 ±6.8
    budget-blind training28.5 ±6.251.6 ±7.559.5 ±7.563.5 ±7.5

Table II. Panel (a): experts. Panel (b): pretrained Uni-NaVid or fine-tuned on the named expert; indented rows repeat that supervision with the countdown removed from every training record. Values are mean N(B)N(B) (%) ± the half-width of the 95% confidence interval. N(B)N(B) is the percentage of distinct valid, reachable targets found within BB primitives, averaged across the 100 matched episodes. Higher is better; bold marks the highest mean among evaluated VLNs at each budget.

At B=150/400/800/2000B = 150/400/800/2000, room-level supervision improves over the pretrained VLN by 3.3/22.5/25.7/27.9 points, over object-level supervision by 3.3/13.5/0.5/7.2, and over greedy-frontier supervision by 6.0/13.5/9.9/4.2; its margin over the pretrained VLN widens with budget. At B=2000B = 2000, it yields 72.7% collection, versus 68.5% for greedy-frontier supervision, 65.5% for object-level supervision, and 44.8% for the pretrained VLN. All VLNs beat the non-learned greedy frontier expert at B=150B = 150, only room-level supervision does at B=400B = 400, and the frontier expert leads every VLN from B=800B = 800 onward (Table II).

One plausible explanation for this crossover is that learned semantic priors help prioritize promising regions when only a few actions are available, whereas a coverage-oriented frontier strategy may spend scarce budget on regions regardless of target likelihood. At longer horizons, explicit spatial memory dominates: greedy frontier tracks explored and unexplored regions directly, while the VLN must retain this information implicitly in its growing context. Consistent with this explanation, at B=800B = 800 and B=2000B = 2000, 92–93% of the room-level-supervised VLN’s forward actions revisit previously traversed cells. Short horizons thus appear to favor semantic prioritization, and long horizons reliable coverage and memory.

The ordering of the two Bellman experts therefore reverses after distillation: higher object-level expert return is associated with lower student collection, while room-level performance exceeds greedy frontier before and after distillation. At B=2000B = 2000, the greedy-frontier-supervised VLN also beats the object-level student (68.5% versus 65.5%), further showing that expert-return rankings need not transfer; this comparison does not isolate Bellman planning from other demonstration differences.

The comparison evaluates object-level routing and room-level search as complete supervision schemes, including their abstractions and execution behaviors. Information alignment is one possible explanation for the reversal [4, 5]. By Eq. (10), a student imitating the object-level expert learns routes averaged over target placements it cannot see, whereas room-level demonstrations search for unseen targets using semantic expectations that depend only on evidence and priors the student could also acquire. Under this view, the information that makes the object-level expert stronger is the same information that makes its actions harder to imitate. The empirical result establishes that expert-return rankings need not transfer; it does not attribute the difference to information access alone, because the experts also differ in planning granularity and execution (Sec. VI).

Budget-blind training. Within each expert’s fixed demonstrations, countdown annotations improve mean collection at B=150/400/800/2000B = 150/400/800/2000 by 2.8/2.8/3.6/3.7 percentage points for object-level and 3.1/6.4/5.0/9.2 for room-level supervision Table II. With countdowns omitted from both students’ training, room-level supervision still leads by 3.0/9.4/4.5/1.7 points. Thus, its advantage persists without explicit budget annotations, while countdown gains depend on supervision source. These comparisons distinguish supervision-source and countdown effects while retaining budget-dependent expert trajectories. The next intervention tests countdown use at inference.

Controlled countdown intervention. To test use of the budget input independently of collection scores, matched probes fix the initial observation, model, and deterministic decoding while changing only the countdown sentence.

Changing B=150B = 150 to B=2000B = 2000 changes the room-level Bellman-supervised VLN’s trajectory on all six held-out-scene pairs, with divergence after 2–11 actions. Collection increases on every pair: 2 → 8, 4 → 7, 1 → 10, 3 → 12, 1 → 7, and 4 → 11. Same-budget repeats reproduce the action sequence. Divergence before either budget expires establishes countdown sensitivity; collection gains also reflect additional available actions and do not alone establish more efficient allocation. In the mesh-rendered probe of Figure 1, the B=50/150/800B = 50/150/800 routes diverge within the first few actions from the same start and collect 4/11/12 of 13 chairs.

Across 206 mined contrast pairs, balanced first-action accuracy against the room-level expert’s budget-conditioned actions rises from 0.344 pretrained to 0.572 after room-level supervision. Yet on the 84 pairs where the expert changes its first action, correctly predicting both branches remains near chance. Budget conditioning is thus more reliable over the evolving closed-loop trajectory than as an immediate switch at t=0t = 0.

Failure analysis. A substantial expert–student gap remains, especially at long horizons. At B=800B = 800 and B=2000B = 2000, 92–93% of the VLN’s forward actions revisit traversed cells. Offline imitation trains on expert trajectories, providing limited coverage of states reached after student deviations. This pattern motivates better closed-loop recovery during learned search.

Real-World Deployment

We deploy the room-level Bellman-supervised VLN zero-shot on an indoor quadruped without real-world fine-tuning. Following Uni-NaVid [2], an onboard camera streams RGB to a remote GPU server, which returns four primitives per call in 0.22 s on average; locomotion dominates runtime.

As in simulation, forward and turning actions command 0.25 m translation and 30° rotation. Odometry executes each primitive until its measured displacement reaches the target, maintaining consistency with predictions.

In a multi-room home containing 14 chairs, we fix the start pose, target, and weights and vary only B∈B \in

{50,150,400,2000}\{50, 150, 400, 2000\} (Figure 4). We run each budget five times; Table III reports the mean and standard deviation (SD) of chairs found. We define the success rate in each run as SR = 100×100 \times (chairs found/14) and report its mean.

Real-robot deployment in a multi-room home

Figure 4. Real-robot deployment in a multi-room home. Routes at B∈{50,150,400,2000}B \in\{50, 150, 400, 2000\} share the scene, start, target category, and weights. At B=50B = 50, the robot searches for the closest chairs near its start; with larger budgets, it first makes longer excursions toward the family room and kitchen. Routes overlay the robot’s lidar map; circles mark starts, stars mark endpoints, and photographs show traversed spaces.

Budget BB501504002000
Chairs found3.4±1.145.0±1.007.6±1.529.2±1.30
Mean SR (%)24.335.754.365.7

Table 3. Real-robot chair search over five runs per budget.

Mean discoveries rise from 3.4 to 9.2 and SR from 24.3% to 65.7% as the budget increases. Routes also differ: at B=50B = 50, the robot searches near its start in the living and dining rooms, whereas larger budgets permit longer transfers to the family room and kitchen. The repeated trials and route changes show budget-dependent behavior despite physical-camera observations and locomotion error.

Discussion and Limitations

Choosing an abstraction for a student. Our results suggest treating the expert’s abstraction as a design variable chosen for the student rather than for expert return. Two requirements pull in different directions. The abstraction must retain enough structure for long-horizon, budget-aware planning; without room connectivity and semantic priors, the expert degenerates toward coverage. It should also keep the expert’s decisions from depending on information that the student can never recover: revealing target locations makes the expert stronger but, by Eq. (10), asks the student to imitate decisions it cannot ground in its inputs. Rooms and doorways meet both requirements reasonably well for indoor search: they are visible in RGB, their semantics can be learned as cross-scene priors, and their unseen contents must be discovered by exploring. A heuristic that follows is to retain what the student can perceive or learn from data, and to withhold what it can only discover by acting. Whether this heuristic predicts student performance across tasks and abstractions remains open.

Disentangling granularity and information access. The object-level and room-level experts differ in decision unit (instances versus rooms), in information access (revealed versus withheld targets), and in execution behavior. Budget-blind controls separate countdown annotations from the supervision source, but not these factors from each other. Two additional experts would isolate information access: a room-level expert given the realized contents of every room, and an object-level expert restricted to targets it has already observed. Comparing their students with ours would directly test the information-alignment explanation.

Residual mismatch and memory. Even the room-level expert reads room types and objects from simulator annotations and maintains an explicit room graph, which the student must infer from RGB history and retain in context. The large remaining expert–student gap at long horizons, together with frequent revisiting, suggests that spatial memory rather than supervision limits the student there. Abstractions that the student can track internally, or students with explicit memory, are natural next steps.

action-count budgets. Weighted target rewards and nonuniform action costs extend the formulation, but supervision rankings there remain unknown. Future work can evaluate these settings and select abstractions for downstream student performance.

Acknowledgment

This work used DeltaAI at the National Center for Supercomputing Applications (NCSA) through allocation CIS261244 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. We thank Mason Lisi for generously providing access to their house for real-world robotic experiments. We also thank Tyler Bonnen for thoughtful discussions.

Conclusion

We study budget-conditioned VLN exploration and the transfer of different expert supervision schemes to Uni-NaVid. Budget-conditioned object search counts distinct geometric discoveries for a denser evaluation signal than binary success; controlled countdown interventions show that changing only the stated budget changes trajectories, beyond extending a fixed route.

Framing experts as Bellman planners over abstractions of the full environment makes the information they act on an explicit design choice: the same budgeted recursion yields fully informed routing or search under uncertainty, depending on what the abstraction reveals.

Expert return and supervision quality diverge: despite higher return from the fully informed object-level Bellman expert, room-level supervision achieves the highest mean collection among evaluated VLNs at every budget. Greedy-frontier supervision also beats object-level supervision at the longest budget, reinforcing that expert rankings need not transfer. The matched Bellman comparison supports room-level search as an effective demonstration scheme for the evaluated VLN; information access, abstraction, and execution jointly define that scheme. Budget-blind controls show that room-level supervision retains a mean advantage without countdown annotations during training, while including these annotations improves both students at every tested budget.

For real-world demonstrations, these findings motivate recording budget with observations and actions and considering search under uncertainty. Such metadata both explains choices between nearby targets and other rooms and provides a learning signal; whether it improves real-world learning requires further evaluation.

At 2000 steps, the room-level-supervised VLN remains below its expert (72.7% versus 98.0%). Frequent revisiting and the non-learned frontier expert’s long-horizon advantage motivate better spatial memory and recovery from trajectory deviations. Rooms are an empirical abstraction, and evidence is limited to indoor search with equal-value targets and

References

  1. [1]F. Pardo, A. Tavakoli, V. Levdik, and P. Kormushev, “Time limits in reinforcement learning,” in International Conference on Machine Learning (ICML), 2018.arxiv.org/abs/1712.00378
  2. [2]J. Zhang, K. Wang, S. Wang, M. Li, H. Liu, S. Wei, Z. Wang, Z. Zhang, and H. Wang, “Uni-NaVid: A video-based vision-language-action model for unifying embodied navigation tasks,” in Robotics: Science and Systems (RSS), 2025.arxiv.org/abs/2412.06224
  3. [3]A.-C. Cheng, Y. Ji, Z. Yang, Z. Gongye, X. Zou, J. Kautz, E. Bylk, H. Yin, S. Liu, and X. Wang, “NaVILA: Legged robot vision-language-action model for navigation,” in Robotics: Science and Systems (RSS), 2025.arxiv.org/abs/2412.04453
  4. [4]L. Weihs, U. Jain, J.-L. Liu, J. Salvador, S. Lazebnik, A. Kembhavi, and A. Schwing, “Bridging the imitation gap by adaptive insubordination,” in Advances in Neural Information Processing Systems (NeurIPS), 2021.
  5. [5]A. Warrington, J. W. Lavington, A. Scibior, M. Schmidt, and F. Wood, “Robust asymmetric learning in POMDPs,” in International Conference on Machine Learning (ICML), 2021.
  6. [6]Y. Kim, N. Chin, A. Vasudev, and S. Choudhury, “Distilling realizable students from unrealizable teachers,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.arxiv.org/abs/2505.09546
  7. [7]Y. Li, G. Xie, and Z. Lu, “Guided policy optimization under partial observability,” in International Conference on Learning Representations (ICLR), 2026.arxiv.org/abs/2505.15418
  8. [8]K. Ehsani, T. Gupta, R. Hendrix, J. Salvador, L. Weihs, K.-H. Zeng, K. P. Singh, Y. Kim, W. Han, A. Herrasti, R. Krishna, D. Schwenk, E. VanderBilt, and A. Kembhavi, “SPOC: Imitating shortest paths in simulation enables effective navigation and manipulation in the real world,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.arxiv.org/abs/2312.02976
  9. [9]Y. Cao, Y. Wang, J. Liang, S. Liao, Y. Zhang, P. Li, and G. Sartoretti, “HEADER: Hierarchical robot exploration via attention-based deep reinforcement learning with expert-guided reward,” arXiv preprint arXiv:2510.15679, 2025.arxiv.org/abs/2510.15679
  10. [10]S. He, C. D. Hsu, D. Ong, Y. S. Shao, and P. Chaudhari, “Active perception using neural radiance fields,” in American Control Conference (ACC), 2024.arxiv.org/abs/2310.09892
  11. [11]A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart, “Receding horizon “next-best-view” planner for 3D exploration,” in IEEE International Conference on Robotics and Automation (ICRA), 2016.DOI
  12. [12]S. He, Y. Tao, I. Spasojevic, V. Kumar, and P. Chaudhari, “An active perception game for robust information gathering,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 14 168–14 174.DOI
  13. [13]S. Baek, B. Moon, S. Kim, M. Cao, C. Ho, S. Scherer, and J. H. Jeon, “PIPE planner: Pathwise information gain with map predictions for indoor robot exploration,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.
  14. [14]Y. Tao, E. Iceland, B. Li, E. Zwecher, U. Heinemann, A. Cohen, A. Avni, O. Gal, A. Barel, and V. Kumar, “Learning to explore indoor environments using autonomous micro aerial vehicles,” in IEEE International Conference on Robotics and Automation (ICRA), 2024.arxiv.org/abs/2309.06986
  15. [15]N. Harutyunyan, B. Moon, S. Kim, C. Ho, A. Hung, and S. Scherer, “MapExRL: Human-inspired indoor exploration with predicted environment context and reinforcement learning,” in International Conference on Advanced Robotics (ICAR), 2025.
  16. [16]H. H. Nguyen, A. Baisero, D. Wang, C. Amato, and R. Platt, “Leveraging fully observable policies for learning under partial observability,” in Conference on Robot Learning (CoRL), 2022.arxiv.org/abs/2211.01991
  17. [17]D. Chen, B. Zhou, V. Koltun, and P. Krähenbühl, “Learning by cheating,” in Conference on Robot Learning (CoRL), 2019.arxiv.org/abs/1912.12294
  18. [18]Y. Cao, R. Z. Zhao, Y. Wang, B. Xiang, and G. Sartoretti, “Deep reinforcement learning-based large-scale robot exploration,” IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 4631–4638, 2024.arxiv.org/abs/2403.10833
  19. [19]D. Shah, A. Sridhar, N. Dashora, K. Zachowicz, K. Black, N. Hirose, and S. Levine, “ViNT: A foundation model for visual navigation,” in Conference on Robot Learning (CoRL), 2023.arxiv.org/abs/2306.14846
  20. [20]D. Shah, B. Osiński, B. Ichter, and S. Levine, “LM-Nav: Robotic navigation with large pre-trained models of language, vision, and action,” in Conference on Robot Learning (CoRL), 2022.arxiv.org/abs/2207.04429
  21. [21]J. Zhang, K. Wang, R. Xu, G. Zhou, Y. Hong, X. Fang, Q. Wu, Z. Zhang, and H. Wang, “NaViDe: Video-based VLM plans the next step for vision-and-language navigation,” in Robotics: Science and Systems (RSS), 2024.arxiv.org/abs/2402.15852
  22. [22]H. Yin, X. Xu, Z. Wu, J. Zhou, and J. Lu, “SG-Nav: Online 3D scene graph prompting for LLM-based zero-shot object navigation,” in Advances in Neural Information Processing Systems (NeurIPS), 2024.arxiv.org/abs/2410.08189
  23. [23]Y. Cao, J. Zhang, Z. Yu, S. Liu, Z. Qin, Q. Zou, B. Du, and K. Xu, “CogNav: Cognitive process modeling for object goal navigation with LLMs,” in IEEE/CVF International Conference on Computer Vision (ICCV), 2025.arxiv.org/abs/2412.10439
  24. [24]D. X. Chen, X. Guo, Y. Duan, R. Zhang, and L. Chen, “WMNav: Integrating vision-language models into world models for object goal navigation,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.arxiv.org/abs/2503.02247
  25. [25]P. Han, F. Jia, M. Zhang, Y. Qiu, H. Tang, Y. Zheng, T. Wang, and J. Hao, “MUUVLA: Learning to explore object navigation via map understanding,” arXiv preprint arXiv:2509.25966, 2025.arxiv.org/abs/2509.25966
  26. [26]N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “VLFM: Vision-language frontier maps for zero-shot semantic navigation,” in IEEE International Conference on Robotics and Automation (ICRA), 2024.arxiv.org/abs/2312.03275
  27. [27]A. Wandzel, Y. Oh, M. Fishman, N. Kumar, L. L. S. Wong, and S. Tellex, “Multi-object search using object-oriented POMDPs,” in IEEE International Conference on Robotics and Automation (ICRA), 2019.DOI
  28. [28]K. Zheng, Y. Sung, G. Konidaris, and S. Tellex, “Multi-resolution POMDP planning for multi-object search in 3D,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021.arxiv.org/abs/2005.02878
  29. [29]S. Kim, M. Corah, J. Keller, G. Best, and S. Scherer, “Multi-robot multi-room exploration with geometric cue extraction and circular decomposition,” IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1190–1197, 2024.DOI
  30. [30]B. Zhou, Y. Zhang, X. Chen, and S. Shen, “FUEL: Fast UAV exploration using incremental frontier structure and hierarchical planning,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 779–786, 2021.arxiv.org/abs/2010.11561
  31. [31]S. Suresh, P. Sodhi, J. G. Mangelson, D. Wettergreen, and M. Kaess, “Active SLAM using 3D submap saliency for underwater volumetric exploration,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 3132–3138.DOI
  32. [32]X. Xu, M. Cao, S. Yuan, T.-M. Nguyen, T.-M. Nguyen, and L. Xie, “A cost-effective cooperative exploration and inspection strategy for heterogeneous aerial system,” in IEEE International Conference on Control and Automation (ICCA), 2024.arxiv.org/abs/2403.01225
  33. [33]S. Wani, S. Patel, U. Jain, A. X. Chang, and M. Savva, “MultiON: Benchmarking semantic map memory using multi-object navigation,” in Advances in Neural Information Processing Systems (NeurIPS), 2020.arxiv.org/abs/2012.03912
  34. [34]M. Khanna, R. Ramrakhya, G. Chhablani, S. Yenamandra, T. Gervet, M. Chang, Z. Kira, D. S. Chaplot, D. Batra, and R. Mottaghi, “GOAT-Bench: A benchmark for multi-modal lifelong navigation,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.arxiv.org/abs/2404.06609
  35. [35]X. Song, W. Chen, Y. Liu, W. Chen, G. Li, and L. Lin, “Towards long-horizon vision-language navigation: Platform, benchmark and method,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.arxiv.org/abs/2412.09082
  36. [36]D. Batra, A. Gokaslan, A. Kembhavi, O. Maksymets, R. Mottaghi, M. Savva, A. Toshev, and E. Wijmans, “ObjectNav revisited: On evaluation of embodied agents navigating to objects,” arXiv preprint arXiv:2006.13171, 2020.arxiv.org/abs/2006.13171
  37. [37]P. Vansteenwegen, W. Souffriau, and D. Van Oudheusden, “The orienteering problem: A survey,” European Journal of Operational Research, vol. 209, no. 1, pp. 1–10, 2011.DOI
  38. [38]E. Altman, Constrained Markov Decision Processes. Chapman and Hall/CRC, 1999.
  39. [39]G. A. Hollinger and G. S. Sukhatme, “Sampling-based robotic information gathering algorithms,” International Journal of Robotics Research, vol. 33, no. 9, pp. 1271–1287, 2014.DOI
  40. [40]M. Popović, T. Vidal-Calleja, G. Hitz, J. J. Chung, I. Sa, R. Siegwart, and J. Nieto, “An informative path planning framework for UAV-based terrain monitoring,” Autonomous Robots, vol. 44, no. 6, pp. 889–911, 2020.arxiv.org/abs/1809.03870
  41. [41]Y. Cao, Y. Wang, A. Vashisht, H. Fan, and G. Sartoretti, “CAtNIPP: Context-aware attention-based network for informative path planning,” in Conference on Robot Learning (CoRL), 2022.
  42. [42]H. Zhu, J. J. Chung, N. R. J. Lawrence, R. Siegwart, and J. Alonso-Mora, “Online informative path planning for active information gathering of a 3D surface,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2021, pp. 1488–1494.
  43. [43]B. Moon, N. Suvarna, A. Jong, S. Chatterjee, J. Yuan, M. Cao, and S. Scherer, “IA-TIGRIS: An incremental and adaptive sampling-based planner for online informative path planning,” IEEE Transactions on Robotics, pp. 1–19, 2026.DOI
  44. [44]J. J. Chung, N. R. J. Lawrence, and S. Sukkarieh, “Learning to soar: Resource-constrained exploration in reinforcement learning,” The International Journal of Robotics Research, vol. 34, no. 2, pp. 158–172, 2015.DOI
  45. [45]D. Golovin and A. Krause, “Adaptive submodularity: Theory and applications in active learning and stochastic optimization,” Journal of Artificial Intelligence Research, vol. 42, pp. 427–486, 2011.arxiv.org/abs/1003.3967
  46. [46]B. Yamauchi, “A frontier-based approach for autonomous exploration,” in IEEE International Symposium on Computational Intelligence in Robotics and Automation (CIRA), 1997.
  47. [47]D. S. Chaplot, D. Gandhi, S. Gupta, A. Gupta, and R. Salakhutdinov, “Learning to explore using active neural SLAM,” in International Conference on Learning Representations (ICLR), 2020.arxiv.org/abs/2004.05155
  48. [48]L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch, “Decision transformer: Reinforcement learning via sequence modeling,” in Advances in Neural Information Processing Systems (NeurIPS), 2021.arxiv.org/abs/2106.01345
  49. [49]A. Kumar, X. B. Peng, and S. Levine, “Reward-conditioned policies,” arXiv preprint arXiv:1912.13465, 2019.arxiv.org/abs/1912.13465
  50. [50]N. Carrara, E. Leurent, R. Laroche, T. Urvoy, O.-A. Maillard, and O. Pietquin, “Budgeted reinforcement learning in continuous state space,” in Advances in Neural Information Processing Systems (NeurIPS), 2019.arxiv.org/abs/1903.01004
  51. [51]T. Chen, S. Kumar, J. Xu, G. Pavlakos, J. Biswas, and R. Martín-Martín, “Searching in space and time: Unified memory-action loops for open-world object retrieval,” arXiv preprint arXiv:2511.14004, 2025.arxiv.org/abs/2511.14004
  52. [52]X. Lin, K. Wu, J. Li, J. Tang, Q. He, and L. Zhao, “BAT-Nav: Budget-aware arbitration and termination for long-horizon semantic navigation,” arXiv preprint arXiv:2605.16932, 2026.
  53. [53]J. Li, N. Zhang, X. Qu, K. Lu, G. Li, J. Wan, and J. Wang, “RATE-Nav: Region-aware termination enhancement for zero-shot object navigation with vision-language models,” in Findings of the Association for Computational Linguistics (ACL Findings), 2025, pp. 6564–6574.arxiv.org/abs/2506.02354
  54. [54]D. Nakajima, K. Tanaka, D. Iwata, and K. Terashima, “MOON: Multi-objective optimization-driven object-goal navigation using a variable-horizon set-orienteering planner,” arXiv preprint arXiv:2505.12752, 2025.arxiv.org/abs/2505.12752
  55. [55]D. Lawson and A. H. Qureshi, “Control transformer: Robot navigation in unknown environments through PRM-guided return-conditioned sequence modeling,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023.
  56. [56]S. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y. Zhang, “Matterport3D: Learning from RGB-D data in indoor environments,” in International Conference on 3D Vision (3DV), 2017.arxiv.org/abs/1709.06158
  57. [57]S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. M. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, M. Savva, Y. Zhao, and D. Batra, “Habitat-Matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI,” in NeurIPS Datasets and Benchmarks, 2021.arxiv.org/abs/2109.08238
  58. [58]M. Savva, A. Kadian, O. Maksymets, Y. Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V. Koltun, J. Malik, D. Parikh, and D. Batra, “Habitat: A platform for embodied AI research,” in IEEE/CVF International Conference on Computer Vision (ICCV), 2019.arxiv.org/abs/1904.01201

Paper details

vlnc:W_p77wzvc9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents