CoRL 2026 Submission

Finder: Agentic Closed-Loop Object Finding for Embodied Grounding

A reusable closed-loop primitive that turns language-based object grounding into query-conditioned evidence gathering, candidate verification, and explicit accept/continue/abort control.

Anonymous Author(s)

Anonymous Institution

Finder overview

Abstract

Finding the object referred to by language in a partially observed 3D scene is a core capability for embodied agents. Existing approaches either couple object search with online exploration, which can be costly when relevant observations have already been captured, or query pre-built open-vocabulary maps and scene graphs in a largely one-shot manner.

We present Finder, an agentic closed-loop object-finding primitive for embodied grounding. Instead of treating grounding as passive retrieval from a fixed scene representation, Finder maintains a typed loop state that links query-conditioned planning, scoped evidence gathering, candidate verification, and accept/continue/abort control. When evidence is incomplete or ambiguous, the loop can redirect subsequent perception and comparison rather than simply returning the top retrieved object.

On open-vocabulary embodied Object Retrieval in Habitat/HM3D and real-world RGB-D scenes, Finder improves the averaged 1m success rate by 15.75 points over strong baselines. The same primitive also transfers to sequential object grounding and embodied object-centric question answering, improving spatial and temporal localization without changing the inner grounding protocol.

Closed-Loop Finder

Finder organizes one grounding episode as a typed loop: plan what to search for, route perception to scoped evidence, detect candidates, verify candidate-centric context, and judge whether to accept, continue, or abort.

Finder core execution protocol
Core Finder execution protocol: a typed closed loop with explicit request, plan, detection, context, selection, decision, hint, and snapshot objects.
Query-conditioned planning

Finder allocates perception budget to explicit room or scene scopes and expands compact phrase sets for open-vocabulary detection.

Candidate-centric evidence

Candidate shortlist construction keeps local visual context and support-object relations available for verification.

Loop judgment

A structured judge decides whether current evidence is sufficient or whether the next loop should redirect search.

Interactive Demo

These full-loop replays show Finder using SAM3 only as the mask proposal backend, while planning, evidence construction, candidate comparison, and final judgment follow the Finder protocol.

Choose a case. The embedded replay exposes the internal loop state: query planning, candidate detection, evidence construction, selection planning, VLM comparison, and final judgment.

Main Results

Finder is evaluated on open-vocabulary Object Retrieval, sequential object grounding, and embodied object-centric question answering. The same inner grounding primitive is reused across tasks.

+15.75Avg. S@1 improvement over the strongest retrieval baseline
66.24Avg. S@1 on Object Retrieval
25.91SG3D sequential grounding s-acc
39.04mOC-NaVQA spatial localization error

Object Retrieval

Method Sim S@1 Real S@1 Avg S@0.5 Avg S@1 Avg Err
HOV-SG50.9042.6832.9650.490.408
DualMap51.9419.5136.8350.310.348
FSR-VLN48.9024.3931.6247.670.422
Finder67.2946.3459.9366.240.193

Sequential Grounding

Methods-acct-acc
DAAAM+GPT22.1611.22
Finder25.9112.20

Spatio-Temporal QA on OC-NaVQA

MethodQA Acc.PosErrTempErr
DAAAM0.71141.751.792
Finder0.54039.041.433

Analysis

Ablations show that blind repetition is not enough: later loops need carried feedback and targeted follow-up. Design-space studies further show that Finder benefits from balanced perception and phrase budgets rather than simply maximizing either axis.

Finder design-space studies
Design-space studies for Finder's default perception and phrase budgets.
Finder qualitative examples
Qualitative examples: support-aware disambiguation, judge-guided cross-loop refinement, and failure modes.

Dataset

The Object Retrieval benchmark contains 1,632 queries over 197 categories across 9 simulated scenes and 4 real-world scenes.

Object Retrieval dataset statistics
Dataset overview: query distribution, scene counts, and simple/hard split.

Citation

BibTeX will be updated after de-anonymization.

@inproceedings{anonymous2026finder,
  title     = {Finder: Agentic Closed-Loop Object Finding for Embodied Grounding},
  author    = {Anonymous Author(s)},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}