ParaEQsA: Multi-Question Embodied QA

• ParaEQsA is a framework that extends traditional embodied QA by handling asynchronous multi-question tasks with dynamic, urgency-aware scheduling and shared memory.
• It employs a Group Memory Module that reuses previous sensory observations to reduce redundant exploration and enable immediate responses for some queries.
• A Priority-Planning Module computes scores using urgency, scope, reward, and dependency, thereby optimizing exploration efficiency and reducing response latency.

ParaEQsA is a framework for Embodied Questions Answering (EQsA), a problem setting in which an embodied agent must answer multiple questions in a 3D environment under asynchronous arrivals, continuous-valued urgencies, and possible inter-question dependencies. It extends classical Embodied Question Answering (EQA), which is typically formulated as answering one single question by actively exploring a 3D environment, toward a multi-question regime intended to reflect realistic deployments. The framework couples a shared group memory with an urgency-aware scheduling mechanism, and it is evaluated on the Parallel Asynchronous Embodied Questions (PAEQs) benchmark introduced alongside the formulation (Wang et al., 15 Sep 2025).

1. Problem formulation

EQsA is defined as a setting in which an agent must handle a dynamic set of questions that arrive at different times, carry continuous-valued urgencies $u_i\in(0,1)$, and may depend on one another (Wang et al., 15 Sep 2025). A scenario is written as

$$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$

where $e$ is a 3D environment, $T$ is the total budget of time steps, $g^0$ is the agent’s start pose, $\mathcal Q_{\rm init}$ arrives at $t=0$, and $\mathcal Q_{\rm follow}$ is a sequence of follow-up questions arriving at times $t_i>0$.

Each question is represented as

$q_i=(s_i, u_i, y_i, t_i),$

with text $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$0, urgency $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$1, ground-truth label $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$2, and arrival time $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$3. The agent’s goal is to plan both motion and question-answering order so as to minimize overall exploration cost, measured in time steps, while also minimizing per-question latency weighted by urgency (Wang et al., 15 Sep 2025).

This formulation distinguishes EQsA from classical EQA in two technically consequential ways. First, the question set is not fixed at episode start, because follow-up questions may arrive later. Second, the optimization target is not solely answer correctness or navigation efficiency; responsiveness is explicitly coupled to urgency and question timing. This makes scheduling part of the task definition rather than a peripheral systems concern.

2. System architecture

ParaEQsA is implemented as a collection of asynchronous microservices. Within that architecture, two modules are identified as especially novel: a Group Memory Module and a Priority-Planning Module (Wang et al., 15 Sep 2025).

The Group Memory Module is a retrieval-based knowledge base that retains all past RGB + depth observations together with symbolic scene annotations, including object labels and semantic map features. Memory is never reset between questions within one scenario. Subsequent questions first query this memory for relevant facts. The intended effect is redundancy reduction: previously observed visual or symbolic evidence can be reused rather than reacquired, allowing some questions to be answered directly with zero additional exploration or restricting new exploration to novel regions.

The Priority-Planning Module organizes unanswered questions in a central Question Pool. Whenever a question is added or selected, an Updater recomputes every question’s priority score, and a Planner selects the highest-priority question for targeted exploration (Wang et al., 15 Sep 2025). Because the module re-evaluates priorities as new questions arrive, it supports dynamic interleaving rather than a fixed sequential order.

Taken together, these components target two distinct inefficiencies in embodied QA. Shared memory addresses repeated sensing of the same environment regions, while dynamic scheduling addresses the mismatch between question order and urgency. This suggests that ParaEQsA is designed around joint control of information reuse and latency management rather than around navigation alone.

3. Priority planning and question scheduling

The central scheduling rule in ParaEQsA assigns each pending question a score

$$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$4

and the Planner selects the highest-$$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$5 question (Wang et al., 15 Sep 2025).

The urgency term is defined as

$$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$6

which the paper describes as a concave boost to high $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$7. Scope is a binary indicator equal to $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$8 if the question can be answered with a local view and $$\xi=(e, T, g^0,\mathcal Q_{\rm init},\mathcal Q_{\rm follow}),$$9 if it requires global exploration. Reward estimates overlap benefit, namely how many other questions refer to the same object or area. Dependency is $e$0 if all prerequisite questions have been answered and $e$1 otherwise (Wang et al., 15 Sep 2025).

These terms operationalize different scheduling pressures. Urgency favors time-sensitive questions. Scope favors questions that may be resolved locally. Reward biases the schedule toward exploration that is likely to amortize across multiple questions. Dependency prevents selection of questions whose prerequisites are not yet satisfied. The module therefore implements a multi-factor scheduler rather than a single-criterion prioritizer.

A plausible implication is that the framework treats embodied question answering as a resource-allocation problem over shared perceptual evidence. In that view, the most valuable next action is not simply the one that advances the current question, but the one that best trades off urgency, locality, overlap, and prerequisite structure.

4. PAEQs benchmark

The Parallel Asynchronous Embodied Questions benchmark is introduced to evaluate EQsA under controlled asynchronous workloads (Wang et al., 15 Sep 2025). It contains 40 photo-realistic indoor scenes from Habitat-Matterport 3D, with 5 questions per scene for a total of 200 questions.

The benchmark structure fixes 3 initial questions at $e$2 and 2 follow-up questions arriving 120 seconds later. Questions span five semantic types: Existence, Counting, State, Identification, and Location. Urgency scores $e$3 were auto-generated by a frozen LLM, specifically gpt-oss-20b with temperature 0, via few-shot prompting, and then bucketed into low $e$4, medium $e$5, and high $e$6 (Wang et al., 15 Sep 2025).

The benchmark is structured to expose failure modes that are largely absent from single-question EQA episodes. Because follow-up questions arrive after exploration has already begun, an agent must decide whether to continue its current plan or reallocate effort. Because questions may share objects or scene regions, memory reuse can materially affect efficiency. Because urgencies vary, not all latency is equivalent.

A recurrent source of confusion is to regard this benchmark as merely a batched version of EQA. The task definition does not support that interpretation. Arrival times, urgency labels, and dependencies are explicit parts of the scenario specification, so the benchmark evaluates online scheduling behavior in addition to answer generation and embodied exploration.

5. Evaluation protocol and reported results

The paper introduces two metrics intended to jointly characterize efficiency and responsiveness: Direct Answer Rate (DAR) and Normalized Urgency-Weighted Latency (NUWL) (Wang et al., 15 Sep 2025).

DAR is defined as the fraction of questions answered immediately from Group Memory without exploration: $e$7 NUWL is defined from the urgency $e$8 of each question, the response time $e$9, the total episode horizon $T$0, and $T$1. The paper states that NUWL jointly measures efficiency and responsiveness and that answering high-urgency questions sooner drives NUWL toward 1 (Wang et al., 15 Sep 2025).

On the main comparison over the PAEQs benchmark, averaged over 40 scenes and 200 questions, ParaEQsA is compared with Explore-EQA and Memory-EQA.

Model	Acc	DAR
ParaEQsA	0.65	0.09
Explore-EQA	0.62	0.00
Memory-EQA	0.64	0.00

Model	NS	NUWL
ParaEQsA	0.321	0.204
Explore-EQA	0.472	0.551
Memory-EQA	0.410	0.474

The reported pattern is that ParaEQsA attains the highest accuracy, the only nonzero DAR, the lowest Normalized Steps (NS), and the lowest NUWL among the compared systems (Wang et al., 15 Sep 2025). The nonzero DAR is specifically tied to the use of shared group memory, while the reductions in NS and NUWL are attributed to urgency-aware scheduling and reduced redundant exploration.

6. Ablations, interpretation, and significance

The paper reports ablations on the priority module while retaining group memory. Removing the entire Priority component changes NUWL from 0.204 to 0.232. Removing the urgency term changes NUWL to 0.223, removing the scope term to 0.213, removing the reward term to 0.226, and removing the dependency term to 0.218 (Wang et al., 15 Sep 2025).

These ablations are used to assess the relative contributions of priority, urgency modeling, spatial scope, reward estimation, and dependency reasoning. The reported behavior indicates that each term contributes to scheduling quality, with the full priority formulation producing the strongest result among the listed variants.

A plausible misconception is that shared memory alone resolves the multi-question setting. The reported results argue against that interpretation. Memory-EQA matches ParaEQsA more closely on accuracy than Explore-EQA does, but it does not achieve a nonzero DAR advantage over ParaEQsA’s scheduling gains, nor does it match ParaEQsA on NS or NUWL. This suggests that knowledge reuse and question prioritization are complementary rather than interchangeable mechanisms.

In the paper’s framing, the broader significance of ParaEQsA is that urgency-aware, parallel scheduling is key to making embodied agents responsive and efficient under realistic, multi-question workloads (Wang et al., 15 Sep 2025). Within the scope of the reported evidence, the contribution is therefore twofold: a reformulation of embodied QA around asynchronous multi-question episodes, and a systems design that couples shared perceptual memory with dynamic, multi-factor scheduling.