Embodied Questions Answering (EQsA)

• Embodied Questions Answering (EQsA) is a formulation where an agent handles multiple asynchronous queries in 3D environments, emphasizing scheduling, memory reuse, and urgency-aware decision making.
• The ParaEQsA framework employs a parallel, microservice-style architecture that leverages shared group memory to reduce redundant exploration and efficiently process overlapping questions.
• EQsA systems use priority scheduling based on urgency, scope, reward, and dependency to address questions in a timely manner, balancing exploration efficiency with immediate answering needs.

Searching arXiv for papers on Embodied Questions Answering (EQsA) and closely related Embodied Question Answering work. Embodied Questions Answering (EQsA) is a formulation in which an embodied agent answers multiple questions that may arrive asynchronously while operating in a 3D environment, rather than solving a single isolated query per episode. In this setting, questions may carry different urgencies, may share overlapping spatial information, and may include dependencies, so correctness alone is insufficient: the system must also schedule, reuse memory, and respond efficiently. The formulation is introduced explicitly in “ParaEQsA: Parallel and Asynchronous Embodied Questions Scheduling and Answering” (Wang et al., 15 Sep 2025), and it is best understood as a realistic extension of Embodied Question Answering (EQA), a broader family of tasks in which an agent must explore, remember, reason, and answer grounded questions from egocentric observations (Wang et al., 15 Sep 2025).

1. From EQA to EQsA

Embodied Question Answering (EQA) is the task of answering natural-language questions by grounding them in visual evidence from an environment the agent observes from an embodied, often egocentric, perspective (Li et al., 27 May 2025). In the classic formulation, the agent may need to navigate, gather evidence, remember what it saw, and reason about the scene before answering (Li et al., 27 May 2025). Standard EQA is typically formulated as answering one single question by actively exploring a 3D environment (Wang et al., 15 Sep 2025).

A central limitation of this classical formulation is that it assumes a single, isolated query and sequential completion (Wang et al., 15 Sep 2025). Real deployments, however, often demand handling multiple questions that may arrive asynchronously and carry different urgencies (Wang et al., 15 Sep 2025). EQsA addresses this mismatch by modeling a situation where the agent receives an initial set of questions, additional follow-up questions arrive asynchronously during execution, each question may have a different urgency, and questions may have dependencies and overlapping spatial information (Wang et al., 15 Sep 2025).

This shifts the objective from question answering alone to workload management under embodied constraints. A plausible implication is that EQsA imports scheduling, latency, and shared-memory considerations into embodied QA, making it not merely a larger instance of EQA but a distinct systems-level problem.

2. Formalization and task structure

EQsA is formalized as a scenario

$$\xi := (e, T, g^0, \mathcal{Q}_{init}, \mathcal{Q}_{follow})$$

where $e$ is the 3D environment/scene, $T$ is maximum number of time steps, $g^0$ is the initial pose, $\mathcal{Q}_{init}$ are initial questions at time 0, and $\mathcal{Q}_{follow}$ are follow-up questions arriving asynchronously (Wang et al., 15 Sep 2025). Each question is represented as

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

where $s_i$ is question text, $u_i \in (0,1)$ is urgency, $y_i$ is the ground-truth answer, and $e$0 is arrival time (Wang et al., 15 Sep 2025). The paper uses four-way multiple choice answers $e$1 (Wang et al., 15 Sep 2025).

This formulation differs from prior EQA variants along several axes. In standard EQA, an agent is typically given one question per episode and must actively explore a 3D environment to answer it (Wang et al., 15 Sep 2025). Other extensions broaden the reasoning demand without introducing asynchronous multi-question scheduling. “Multi-Target Embodied Question Answering” (Yu et al., 2019) generalizes EQA to questions with multiple targets in them, such as “Is the dresser in the bedroom bigger than the oven in the kitchen?”, requiring navigation to multiple locations and comparative reasoning (Yu et al., 2019). “Knowledge-based Embodied Question Answering” (Tan et al., 2021) further generalizes EQA by requiring external knowledge, logical reasoning, and multi-turn memory (Tan et al., 2021). These formulations increase semantic or compositional complexity, but they do not redefine the task around urgency-aware asynchronous question handling.

The distinction is important because EQsA introduces performance criteria that do not arise in one-question EQA. A system may answer all questions correctly yet still be inadequate if it delays urgent questions, duplicates exploration, or fails to exploit cross-question overlap. This suggests that EQsA is partly a scheduling problem over embodied evidence acquisition.

3. Core architectural ideas in EQsA systems

The canonical system proposed for EQsA is ParaEQsA, a framework for parallel, urgency-aware scheduling and answering (Wang et al., 15 Sep 2025). Its overall idea is to process multiple embodied questions in a shared, parallelized pipeline instead of answering each question in isolation (Wang et al., 15 Sep 2025). The system is designed as a microservice-style asynchronous architecture with several modules communicating through a Redis stream queue (Wang et al., 15 Sep 2025).

The main modules are: Generator, Parser, Finishing Module, Question Pool, Planner, Stopping Module, Answering Module, and Group Memory (Wang et al., 15 Sep 2025). The Generator loads question groups and scenes and emits initial and follow-up questions according to the temporal schedule (Wang et al., 15 Sep 2025). The Parser uses an LLM to extract semantic metadata from each question, specifically urgency and scope, computed once per question (Wang et al., 15 Sep 2025). The Finishing Module checks whether the question can be answered directly using Group Memory and computes a confidence score from retrieved evidence; if confidence is enough, the answer is produced immediately, otherwise the question goes to the Question Pool (Wang et al., 15 Sep 2025). The Planner picks the highest-priority question and performs targeted exploration to gather information for that question (Wang et al., 15 Sep 2025). The Stopping Module periodically decides whether exploration should stop, and the Answering Module uses a VLM plus memory context to produce the final answer (Wang et al., 15 Sep 2025).

The defining architectural contribution is Group Memory. It stores observations gathered during exploration, is shared across all questions in the same scenario, and supports retrieval when a new question arrives or when the system attempts a direct answer (Wang et al., 15 Sep 2025). In classical sequential EQA, each question is often solved independently, causing repeated exploration of the same areas (Wang et al., 15 Sep 2025). ParaEQsA avoids this by reusing accumulated knowledge across questions, enabling direct answering of some questions without exploration and reduced redundant navigation (Wang et al., 15 Sep 2025).

This shared-memory design connects EQsA to a broader line of memory-augmented embodied QA. MEIA introduces a Multimodal Environment Memory (MEM) that combines object IDs and coordinates as environmental language memory and visual observation pictures as environmental image memory, plus an environmental floor plan derived from accumulated 3D observations (Liu et al., 2024). GraphEQA uses real-time 3D metric-semantic scene graphs and task relevant images as multi-modal memory (Saxena et al., 2024). Long-term Active EQA introduces a structured memory system for robots in the form of a Robotic Mind Palace $e$2, where each $e$3 is a scene-graph world instance corresponding to one episode in time (Ginting et al., 17 Jul 2025). EQsA differs in that its memory is explicitly shared among concurrently pending questions, but it inherits the same premise: embodied QA requires persistent, structured external memory rather than frame-local reasoning alone.

4. Scheduling, urgency, and dependency reasoning

The priority-planning mechanism in ParaEQsA is implemented through the Question Pool and its Updater (Wang et al., 15 Sep 2025). Instead of FIFO order, questions are assigned a priority: $e$4 where $e$5 are weights for the four components (Wang et al., 15 Sep 2025).

Urgency is modeled nonlinearly: $e$6 so that as $e$7 increases toward 1, urgency grows sharply (Wang et al., 15 Sep 2025). Scope is defined as

$e$8

favoring local questions because they can often be answered with limited exploration (Wang et al., 15 Sep 2025). Reward is based on how many other questions refer to the same or nearby locations or objects, encouraging exploration of information-rich places that may help multiple questions (Wang et al., 15 Sep 2025). Dependency is defined as

$e$9

and dependencies are maintained in a DAG (Wang et al., 15 Sep 2025).

This machinery formalizes a major conceptual shift. Classical EQA asks where to navigate next for the current question; EQsA asks which question should control the next navigation decision. A plausible implication is that urgency-aware scheduling becomes as fundamental as target selection, because poor question ordering directly degrades both efficiency and responsiveness.

The exploration strategy remains grounded in the EQA tradition. ParaEQsA uses targeted exploration aligned with Explore-EQA style frontier-based navigation (Wang et al., 15 Sep 2025). Explore-EQA itself builds a semantic map of the scene from depth information and visual prompting of a VLM, then uses conformal prediction to calibrate the VLM’s question answering confidence for stopping (Ren et al., 2024). GraphEQA performs hierarchical planning over object and frontier nodes in a real-time 3D semantic scene graph (Saxena et al., 2024). FAST-EQA combines question-conditioned target localization, global and local relevance-guided navigation, bounded visual memory, and LLM-based reasoning (Zhang et al., 17 Feb 2026). These methods focus on efficient evidence acquisition for one question; EQsA reuses such exploration policies but places them under a higher-level scheduler.

5. Benchmarks and evaluation

The benchmark introduced for EQsA is PAEQs, the Parallel Asynchronous Embodied Questions benchmark (Wang et al., 15 Sep 2025). It is built on HM3D indoor 3D scenes and contains 40 high-quality residential scenes, each with 5 questions, for 200 total questions (Wang et al., 15 Sep 2025). Each scenario has 3 initial questions and 2 follow-up questions, with follow-up questions introduced 120 seconds after the previous question (Wang et al., 15 Sep 2025). All 200 questions were manually written and reviewed, and ground-truth answers were verified carefully (Wang et al., 15 Sep 2025).

The benchmark covers five semantic types: existence, counting, state, identification, and location (Wang et al., 15 Sep 2025). Each question also has an urgency score $T$0, generated using gpt-oss-20b with few-shot prompting and temperature 0 (Wang et al., 15 Sep 2025). Urgency is grouped into low, medium, and high categories (Wang et al., 15 Sep 2025). The agent does not know these labels at test time; they are only for benchmarking and evaluation (Wang et al., 15 Sep 2025).

EQsA requires metrics beyond answer accuracy. ParaEQsA defines: $T$1 for accuracy (Wang et al., 15 Sep 2025). It then introduces Direct Answer Rate (DAR),

$T$2

where $T$3 is the subset of questions answered directly from memory with zero exploration steps (Wang et al., 15 Sep 2025). It also defines normalized steps,

$T$4

to measure exploration efficiency (Wang et al., 15 Sep 2025). The key responsiveness metric is Normalized Urgency-Weighted Latency (NUWL), intended to penalize questions that wait longer before being addressed, with more urgent questions contributing more heavily (Wang et al., 15 Sep 2025).

These metrics are distinctive because they disentangle three aspects of performance: correctness, memory reuse, and urgency-sensitive delay. A plausible implication is that DAR operationalizes a form of embodied amortization: a good EQsA system converts past exploration into immediate future answers.

6. Relation to neighboring EQA problem settings

EQsA emerges against a rapidly broadening EQA landscape. Several nearby task variants clarify what EQsA includes and what it does not.

Long-term Active Embodied Question Answering (LA-EQA) requires an agent to combine persistent long-term memory with active exploration in a changing environment (Ginting et al., 17 Jul 2025). It is defined as the tuple

$T$5

where $T$6 is the question, $T$7 is a set of episodic memories, $T$8 is the current environment, $T$9 is the initial pose, and $g^0$0 is the ground-truth answer (Ginting et al., 17 Jul 2025). Its action space is explicitly three-way: retrieve a past observation from memory, explore the current environment, or answer and terminate (Ginting et al., 17 Jul 2025). LA-EQA studies recall–exploration trade-offs over long timescales, whereas EQsA studies scheduling–reuse trade-offs across simultaneous or asynchronous questions.

NoisyEQA benchmarks whether EQA agents can remain reliable when humans ask noisy questions containing mistaken assumptions, memory errors, perceptual confusions, or semantic substitutions (Wu et al., 2024). Its four noise types are Latent Hallucination Noise, Memory Noise, Perception Noise, and Semantic Noise (Wu et al., 2024). It adds a Self-Correction mechanism with NAP and NACoT prompting, and evaluates detection and correction through DR and CR metrics (Wu et al., 2024). This makes question reliability the central issue, rather than multi-question scheduling.

AbstainEQA asks whether an embodied agent should answer at all. It introduces five abstention categories: actionability limitation, referential underspecification, preference dependence, information unavailability, and false presupposition (Wu et al., 4 Dec 2025). The paper argues that abstention is a prerequisite for reliable embodied interaction and a necessary basis for effective clarification (Wu et al., 4 Dec 2025). EQsA does not center abstention, but in realistic deployment the two concerns intersect: urgency-aware scheduling is valuable only if the agent can also identify when a question should not be answered.

Other extensions widen the reasoning target. S-EQA introduces situational queries such as “Is the house ready for sleeptime?”, where the answer depends on consensus over multiple object-states (Dorbala et al., 2024). StreamEQA moves EQA into streaming video understanding over egocentric video streams with embodied and temporal dimensions (Wang et al., 4 Dec 2025). EQA-Decision extends embodied QA from perception to decision-making with four reasoning dimensions—static scene construction, spatial understanding, task dynamics reasoning, and instant decision (Gong et al., 25 May 2026). These works broaden embodied QA along semantics, time, and action; EQsA broadens it along concurrency and responsiveness.

7. Empirical findings, significance, and open issues

ParaEQsA is evaluated against two sequential baselines adapted from recent EQA systems: Explore-EQA and Memory-EQA (Wang et al., 15 Sep 2025). The reported results are:

Method	Acc	DAR	NS	NUWL
ParaEQsA	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

These results show slightly better accuracy, substantially reduced normalized steps, greatly lower NUWL, and nonzero direct answer rate only for ParaEQsA (Wang et al., 15 Sep 2025). The paper reports about a 57% reduction in NUWL versus Memory-EQA and 63% versus Explore-EQA (Wang et al., 15 Sep 2025). Ablations show that removing priority altogether causes the biggest degradation, urgency modeling is especially important for lowering NUWL, and reward and dependency both help efficiency and timeliness (Wang et al., 15 Sep 2025).

The central insight is that embodied agents in realistic environments should not solve questions one by one in arrival order (Wang et al., 15 Sep 2025). Instead, they should exploit shared information across questions, prioritize urgent and ready questions, exploit local or high-reward regions first, and delay low-urgency or blocked questions when appropriate (Wang et al., 15 Sep 2025). This leads to less redundant exploration, more direct answers from memory, faster responses to critical questions, and better overall workload handling (Wang et al., 15 Sep 2025).

At the same time, EQsA inherits unresolved limitations from the wider EQA literature. Memory quality remains structurally important: MEIA’s ablations show that removing multimodal memory lowers embodied QA accuracy from 70.0% to 56.7% (Liu et al., 2024), while GraphEQA’s ablations show that scene-graph-only or image-only memory underperforms the combined representation (Saxena et al., 2024). Robustness to noisy or ambiguous questions remains weak in current models (Wu et al., 2024, Wu et al., 4 Dec 2025). Efficient stopping remains an open issue, addressed differently by conformal prediction in Explore-EQA (Ren et al., 2024), outlier-based stopping in EfficientEQA (Cheng et al., 2024), GPT-4o-based stopping in FAST-EQA (Zhang et al., 17 Feb 2026), and value-of-information-based stopping in LA-EQA (Ginting et al., 17 Jul 2025). This suggests that a mature EQsA system will likely need to combine asynchronous scheduling with calibrated stopping, memory retrieval, abstention, and noise handling.

In summary, EQsA designates a multi-question, asynchronous, urgency-aware extension of embodied question answering in which shared memory and question scheduling are first-class components (Wang et al., 15 Sep 2025). Its significance lies in recasting embodied QA from a single-query benchmark into a persistent interactive workload, where the agent must decide not only where to look and when to stop, but also which question deserves attention next.