Parallel Asynchronous Embodied Questions (PAEQs)

• PAEQs is a benchmark setting for embodied question answering that handles multiple asynchronous queries with varying urgency under a shared exploration budget.
• It integrates urgency-aware scheduling, shared memory reuse, and parallel processing to minimize exploration steps while enhancing responsiveness.
• The benchmark uses 40 indoor scenes with 200 questions and evaluates performance using metrics like accuracy, DAR, NS, and NUWL to balance efficiency and timeliness.

Searching arXiv for the specified work and closely related embodied QA papers to ground the encyclopedia entry. {"query":"id:(Wang et al., 15 Sep 2025) OR \"Parallel and Asynchronous Embodied Questions Scheduling and Answering\" OR PAEQs embodied question answering", "max_results": 10} {"query":"\"MA-EgoQA\" OR \"Question Answering over Egocentric Videos from Multiple Embodied Agents\" OR (Kim et al., 10 Mar 2026)", "max_results": 10} {"query":"\"Explore-EQA\" embodied QA Ren 2024 OR \"Memory-EQA\" Zhai 2025", "max_results": 10} Parallel Asynchronous Embodied Questions (PAEQs) denotes a benchmarked setting for embodied question answering in which multiple questions arrive over time, carry different urgency levels, and must be handled under a shared exploration budget rather than as isolated single-query episodes. In the formulation introduced alongside ParaEQsA, PAEQs is the evaluation benchmark for Embodied Questions Answering (EQsA), with 40 indoor scenes and five questions per scene, including asynchronous follow-up questions and urgency labels (Wang et al., 15 Sep 2025). The central shift is from classical Embodied Question Answering, which is typically formulated as answering one single question by actively exploring a 3D environment, to a workload in which an embodied agent must jointly optimize exploration efficiency and response timeliness.

1. Formal task structure

Within EQsA, a single scenario is drawn from an unknown distribution $\mathcal D$ and specified as

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

where $e$ is a 3D environment, $T\in\mathbb N$ is the total time budget, $g^0$ is the agent’s starting pose, $\mathcal Q_{\rm init}$ is the set of initial questions presented at $t=0$, and $\mathcal Q_{\rm follow}$ is a stream of follow-up questions arriving at times $t_i>0$ (Wang et al., 15 Sep 2025).

Each question is represented as

$q_i = \bigl(s_i,\;u_i,\;y_i,\;t_i\bigr),$

where $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$0 is the natural-language text, $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$1 is its urgency score, $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$2 is the hidden ground-truth answer, and $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$3 is its arrival time. In this setting, $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$4, so every question is multiple choice.

Agent–environment interaction proceeds over discrete time steps $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$5. At each step, the agent chooses a motion action $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$6, obtains a new pose $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$7, captures RGB $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$8 and depth $$\xi \;=\;\bigl(e,\;T,\;g^0,\;\mathcal Q_{\rm init},\;\mathcal Q_{\rm follow}\bigr),$$9, updates an internal memory $e$0, and may emit answers for any active questions. The system state can be summarized as

$e$1

where $e$2 contains all questions with $e$3 that have not yet been answered.

The resulting optimization target is explicitly dual. The agent must minimize total exploration steps while also answering high-urgency questions with minimal delay. If question $e$4 is answered at time $e$5, its response latency is

$e$6

The scheduling problem is therefore to find a policy $e$7 over states and question-selection actions that balances overall exploration cost against urgency-weighted latency. This formulation makes question ordering a first-class control variable rather than a post hoc reporting artifact.

2. Benchmark design and data characteristics

The PAEQs benchmark is constructed from 40 diverse indoor scenes drawn from Habitat-Matterport 3D (HM3D) (Wang et al., 15 Sep 2025). Each scene contains exactly five questions: three initial questions at $e$8 and two follow-up questions arriving at fixed delays, each 120 seconds after the previous answer. The assignment of which questions are initial versus follow-up was randomized with equal probability to avoid bias.

Urgency annotation is part of the benchmark definition rather than an auxiliary label. Each question receives an urgency score $e$9 via a deterministic few-shot prompt to GPT-OSS-20B with temperature 0. These continuous scores are bucketed into three levels:

Urgency level	Range	Share of data
Low	$T\in\mathbb N$0	57.5%
Medium	$T\in\mathbb N$1	26.5%
High	$T\in\mathbb N$2	16.0%

The benchmark contains 200 total questions distributed across five semantic types: Existence (34.5%), Counting (23.0%), State (21.5%), Identification (14.5%), and Location (6.5%). Each question is manually authored, proofread, and paired with four choice options, with dummy distractors “D) Do not choose this option” used where necessary.

The paper reports that semantic correlation emerges organically in the urgency labels: safety-critical questions such as “Is anyone trapped on the stairs?” receive high scores, whereas “What color is the couch?” tends to be low urgency. The benchmark therefore evaluates not only perceptual and navigational competence but also the ability to treat heterogeneous question importance as operationally meaningful.

3. Evaluation criteria

PAEQs is evaluated with four metrics that jointly capture correctness, exploration efficiency, and responsiveness (Wang et al., 15 Sep 2025).

Accuracy is defined as

$T\in\mathbb N$3

This is the conventional correctness measure over the full question set.

Direct Answer Rate (DAR) measures the fraction of questions answered without any exploration steps: $T\in\mathbb N$4 DAR operationalizes whether prior exploration has been successfully consolidated into reusable memory.

Normalized Steps (NS) quantifies exploration cost relative to each question’s budget. For each $T\in\mathbb N$5,

$T\in\mathbb N$6

Lower NS indicates higher exploration efficiency, and $T\in\mathbb N$7 exactly when the question is answered from memory.

Normalized Urgency-Weighted Latency (NUWL) is the benchmark’s key responsiveness measure. With latency $T\in\mathbb N$8,

$T\in\mathbb N$9

Smaller NUWL indicates that the system tended to answer high-urgency questions sooner.

Taken together, these metrics prevent a degenerate optimization toward either pure accuracy with excessive delay or fast answering without adequate exploration. DAR and NS measure the value of memory reuse; NUWL measures whether scheduling respects urgency; Acc measures whether these efficiencies preserve answer quality.

4. ParaEQsA architecture

ParaEQsA is a microservice-based, parallel architecture with seven principal components that communicate via asynchronous queues (Wang et al., 15 Sep 2025). Its design centers on shared memory reuse and urgency-aware question scheduling.

The Generator loads scenario files and emits initial and follow-up questions at their prescribed times. The Parser invokes an LLM to extract semantic metadata for each incoming question: urgency estimate $g^0$0, scope type (“local” versus “global”), and keyword objects or regions of interest.

The Finishing Module then attempts immediate answering by querying the Group Memory $g^0$1. If confidence exceeds a threshold, the question is answered on the spot and contributes to DAR; otherwise it is forwarded to the Question Pool. The Question Pool consists of an Updater and Buffer that maintain all unanswered questions and recompute a scalar priority whenever a question is added or removed: $g^0$2 Here, urgency is transformed by $g^0$3 to spread out high-urgency scores; local questions receive a scope bonus; reward counts other pending questions that share the same region or object; dependency indicates whether prerequisite questions have been answered. The highest-priority question is selected by the Planner.

The Planner executes targeted exploration inspired by frontier-based semantic mapping. From the current pose $g^0$4, the agent captures $g^0$5, uses a VLM to detect regions or objects relevant to the selected question, scores map frontiers, navigates toward the highest-value frontier, updates a 2D semantic map, stores new observations in $g^0$6, and queries a Stopping Module to determine whether enough evidence has been gathered. If so, the question is forwarded to the Answering Module; otherwise exploration continues until stopping or budget exhaustion. Exploration for each question resumes from the final pose of the previous question, mimicking continuous real-world operation.

The Answering Module invokes a VLM over the question and the current memory $g^0$7 to produce the final answer $g^0$8. The Group Memory persists for the entire scenario and indexes every captured image, depth map, and VLM-extracted caption or detection. Both the Finishing Module and Planner retrieve from this shared store, enabling direct answers and reducing redundant exploration across related questions.

5. Reported performance and ablations

On the 40-scene PAEQs benchmark, ParaEQsA is compared against two sequential baselines adapted from recent single-question EQA systems: Explore-EQA and Memory-EQA (Wang et al., 15 Sep 2025).

Method	Acc ↑	DAR ↑	NS ↓	NUWL ↓
ParaEQsA (ours)	0.65	0.09	0.321	0.204
Explore-EQA	0.62	0.00	0.472	0.551
Memory-EQA	0.64	0.00	0.410	0.474

The reported interpretation is specific. ParaEQsA matches or slightly exceeds the baselines in Accuracy, while being the only method with nonzero DAR. Its NS indicates that it requires approximately 32% of the allowed steps on average, which is described as 24–32% fewer than the baselines. Its NUWL is less than half that of either sequential method, indicating substantially improved responsiveness to urgent questions.

Ablation studies examine the relative contributions of priority, urgency modeling, spatial scope, reward estimation, and dependency reasoning. The reported result is that every component of the priority score—urgency transform, local-scope bias, reward clustering, and dependency checking—meaningfully contributes to NS and NUWL improvements. In the benchmark’s terms, the key empirical claim is not merely that shared memory helps, but that urgency-aware scheduling changes the latency profile in a measurable way without sacrificing accuracy.

6. Relation to adjacent embodied QA lines

PAEQs belongs to a broader movement toward embodied QA beyond the single-agent, single-question setting, but related works use different modalities and problem formulations.

“MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents” formalizes a novel problem of understanding multiple long-horizon egocentric videos simultaneously collected from embodied agents and explicitly introduces Parallel Asynchronous Embodied Question Answering over multiple egocentric video streams (Kim et al., 10 Mar 2026). Its benchmark, MultiAgent-EgoQA, is built on EgoLife and contains 1,741 human-verified multiple-choice questions unique to multiple egocentric streams, spanning five categories: social interaction, task coordination, theory-of-mind, temporal reasoning, and environmental interaction. Its baseline, EgoMAS, is a training-free, retrieval-augmented agent system with event-based shared memory and agent-wise dynamic retrieval. The paper reports that current approaches are unable to effectively handle multiple egocentric streams. This establishes a related but distinct asynchronous setting in which the core bottleneck is multi-stream video memory and retrieval rather than navigation-time scheduling.

By contrast, “Multi-LLM QA with Embodied Exploration” does not introduce or use the terminology of “Parallel Asynchronous Embodied Questions (PAEQs)” (Patel et al., 2024). Its setting uses multiple independent LLM-based agents that explore or are given observations from different parts of a household scene, answer the same binary object-presence query, and feed those answers into aggregation mechanisms such as majority voting, debate, or a learned Central Answer Model. The summary provided for that work states that it instantiates a truly parallel and asynchronous multi-agent QA protocol because each LLM explorer operates independently and, at inference time, no inter-agent communication is needed. A plausible implication is that PAEQs is best understood not as a synonym for all parallel embodied QA, but as a specific benchmarked formulation within a larger family of multi-agent and asynchronous embodied reasoning problems.

The principal distinction is therefore structural. In PAEQs, multiple questions arrive over time for a single embodied agent operating in a shared environment with a shared memory and an explicit urgency-aware scheduler. In MA-EgoQA, multiple embodied agents contribute asynchronous egocentric observations that must be aligned into system-level memory. In the Multi-LLM household setting, multiple agents independently answer the same query from partial observations. These formulations are adjacent, but they operationalize “parallel” and “asynchronous” at different loci: question arrival, sensor stream collection, or answer aggregation.