PRA Closed-Loop Architectures

• Perception–Reasoning–Action closed-loop architectures are defined systems that continuously integrate sensory perception, reasoning, and action to enable rapid adaptation.
• They are employed in various domains including robotics, vision-language models, and autonomous agents, enhancing robustness in nonstationary and partially observable settings.
• Implementations leverage dynamic inference, diffusion methods, and hybrid symbolic-neural models to minimize latency and optimize task performance in benchmarks.

Perception–Reasoning–Action Closed-Loop Architectures

Perception–Reasoning–Action (PRA) closed-loop architectures define a system paradigm in which the cycles of sensory data acquisition, high-level reasoning, and physical (or symbolic) action are tightly and continuously coupled, with each step's outputs providing direct feedback to the next. This closed-loop principle underlies contemporary advances in dynamic robotics, embodied intelligence, large-scale vision-language-action models, and adaptive autonomous agents, with a defining feature of rapid, ongoing adaptation to exogenous changes in the environment and to the consequences of an agent’s own prior actions. Strict adherence to closed-loop execution is critical for robust performance in nonstationary, partially observable, and interactive settings.

1. Fundamental Principles of Closed-Loop PRA Architectures

Closed-loop PRA systems are characterized by the continuous, recursive coupling between sensing, inference, and decision/execution. The architectural flow can be canonicalized as:

1. Perception Module: Ingests sensory data (e.g., images, proprioception, depth, or events) and produces structured internal representations (embeddings, object graphs, world models).
2. Reasoning Module: Integrates perceptual representations, context, and goals to perform abstract or numeric reasoning, often encompassing future prediction, planning, or control barrier computation.
3. Action/Execution Module: Generates and deploys actions, possibly as a chunked sequence, and orchestrates their temporal alignment relative to perception–reasoning cycles.
4. Feedback/State Update: The environment—and possibly the agent’s internal state—is altered; fresh observations are acquired, closing the loop.

Persistent feedback distinguishes closed-loop PRA architectures from open-loop counterparts, which perform batch reasoning over static input and act without subsequent verification or re-planning.

A generic high-level workflow (extracted from diverse instantiations) is summarized in the table below:

Phase	Core Operation	Output/Feedback to Next
Perception	Sensory data processing; feature/graph/token extraction	Structured state
Reasoning	Multimodal integration; prediction, planning, constraint	Action or plan
Action	Execution of (possibly temporally aligned/chunked) commands	System/environment
Feedback	Observation of new state	Loop to Perception

2. Canonical Model Families and Instantiations

Several leading PRA closed-loop architectures exemplify the field’s diversity:

DynamicVLA for Dynamic Manipulation

DynamicVLA operationalizes closed-loop control through three modules: (i) a high-throughput FastViT convolutional vision encoder yielding patchwise tokens, (ii) a compact multimodal transformer core with continuous inference and overlapping execution, and (iii) a diffusion-based flow-matching decoder with latent-aware action streaming. Asynchronous inference cycles interleave reasoning and execution to minimize end-to-end latency, with action chunks temporally realigned at execution via explicit discard/overwrite steps. This approach is validated in the Dynamic Object Manipulation (DOM) benchmark, exhibiting dominant success rates and robustness metrics (Xie et al., 29 Jan 2026).

World Models in Embodied Agents

Architectures such as those in "World-in-World" employ a unified generative world model (e.g., latent diffusion or video transformer) to predict counterfactual outcomes for candidate action sequences. Each closed-loop timestep comprises online proposal of multiple plans, fast-forward simulation, and rollout scoring; the best plan's first action is executed before observation begins anew. Standardized APIs enable cross-model benchmarking for embodied control, navigation, and manipulation. Data scaling and compute allocation laws quantify performance improvements, showing that embodied utility—rather than visual open-loop fidelity—dominates task success (Zhang et al., 20 Oct 2025).

Video Diffusion-Based Closed Loop (Vidarc)

Vidarc integrates a VAE-encoded latent representation of video frames with an autoregressive video diffusion transformer and a masked inverse-dynamics action head. Crucially, its closed-loop stack maintains a key-value cache of real observations, regularly overwriting predicted latents with true sensory feedback (“re-prefill”), thus suppressing error drift and reducing control latency by >90%. Grounding action via mask-weighted losses ensures robust control even in dynamic, sparse-data domains (Feng et al., 19 Dec 2025).

Symbolic–Neural and Ontology-Guided Loops

Hybrid systems leverage neural perception (e.g., segmentation, optical flow) piped into a symbolic reasoning substrate (e.g., OWL ontologies, logic rules, or spatial graphs) to formulate queries for perception and guide higher-order planning. The feedback cycle includes attention-focusing (perception-querying) and post-action knowledge-base updates, enabling the system to learn new concepts and functional parts dynamically (e.g., discovering handles for mug–hook support relations) (Pomarlan et al., 28 Jul 2025).

3. Feedback Coupling, Temporal Alignment, and Latency Modeling

A common technical challenge in PRA closed loops is bridging the perception–action gap imposed by nonzero reasoning and execution latencies. Two prominent solutions are:

• Continuous Inference and Overlap: In DynamicVLA, continuous inference allows action decoding to begin before the previous inference finishes. By launching asynchronous inferences at fixed intervals, and overlapping reasoning with partial execution, the architecture achieves

$$\tau_{\mathrm{loop}} = \max(\tau_{\mathrm{perc}}, \tau_{\mathrm{reason}}) + \tau_{\mathrm{action}} - \tau_{\mathrm{overlap}},$$

minimizing action latency relative to environmental change (Xie et al., 29 Jan 2026).

• Latent-Aware Action Streaming (LAAS): Chunks of predicted actions, available only after inference latency $m$, are aligned to current real time by discarding outdated indices; streaming actions are always drawn from the freshest available chunk, overriding stale commands and realigning execution to the environment.

Further, systems like Vidarc employ “KV-cache re-prefill”—processing only new frames instead of re-encoding all history—to avoid drift and slash computational burden (Feng et al., 19 Dec 2025).

4. Representation and Reasoning: Multimodal, Symbolic, and Predictive Models

Closed-loop PRA architectures span a continuum from pure neural approaches to explicit symbolic/graphical models:

• Multimodal Transformer World Models: Systems such as Doe-1 unify perception, description, prediction, and planning as autoregressive token generation. Each timestep interleaves perception (textual Q&A/caption), planning (action tokens), and prediction (video/BEV tokens), wrapping all into a transformer for recurrent closed-loop forecasting and re-planning (Zheng et al., 2024).
• Symbolic Reasoning Engines: Hybrid frameworks (e.g., ExploreVLM, "Hanging Around") employ object-centric spatial-relation graphs or ontologies for grounded reasoning and manipulation planning, incorporating self-reflective planners and execution validators to prompt re-planning on detection of subgoal failure (Lou et al., 16 Aug 2025, Pomarlan et al., 28 Jul 2025).
• Hierarchical Dual-Systems: Architectures such as ORCA for goal-driven video avatars separate strategic reasoning (System 2) from caption-grounded action execution (System 1), with an embedded OTAR (Observe-Think-Act-Reflect) loop ensuring continuous Bayesian belief updating and subgoal verification (He et al., 23 Dec 2025).

5. Empirical Performance and Benchmarking Methodologies

Benchmarking of PRA closed-loop architectures extends across both synthetic and real-world evaluation environments. Examples include:

• DOM (Dynamic Manipulation): 200k synthetic and 2k real-world episodes, multi-axis metrics (success rate, path length, completion time), and ablations revealing the critical role of continuous inference and LAAS (Xie et al., 29 Jan 2026).
• RLBench/COLOSSEUM/GemBench: Used in ActiveVLA to demonstrate state-of-the-art results in 3D manipulation with active perception, measuring success rates and average ranking across 18–14 task types (Liu et al., 13 Jan 2026).
• Autonomous Vehicle Datasets: nuScenes, Bench2Drive, and OpenDrive provide rigorous closed-loop control and prediction evaluation, with metrics on L2 trajectory error, collision rate, driving score, and comfort/efficiency (Zheng et al., 2024, Wang et al., 27 Nov 2025).
• Real-Robot Exploration/Manipulation: ExploreVLM and CLIER report robust sim-to-real transfer and substantial gains over passive or open-loop baselines, explicitly quantifying the contribution of structural graph representations and reflective planners (Lou et al., 16 Aug 2025, Nazarczuk et al., 2024).

Closed-loop benchmarks consistently report significant improvements in both end-task success and robustness to environment change/perturbation over open-loop architectures.

6. Robustness, Generalization, and Theoretical Guarantees

A defining advantage of closed-loop PRA architectures is their adaptability to unmodeled disturbances and partially observable or dynamic environments.

• Generalization: Explicit feedback-driven re-planning and iterative perception prevent overfitting to static or training-phase distributional artifacts. This effect is pronounced in “dual-loop” frameworks where reasoning and perception models regularize each other (e.g., P–D–R navigation agent (Zhang et al., 2022)).
• Robustness to Delays and Latency: By overlapping modules (continuous inference, LAAS, cache re-prefill), architectures maintain task success even as computational or communication latencies fluctuate.
• Theoretical Analysis: Active inference-based nonmodular closed loops, unlike classical LQG modularity, provide integral control that rejects unknown disturbances without requiring explicit efference copy or full plant identification. This underpins their superior disturbance rejection properties (Baltieri et al., 2019).
• Data and Compute Scaling Laws: As demonstrated in World-in-World, increased data and inference-time compute both yield monotonic, sublinear gains in success rate; visual fidelity alone is a weak predictor of embodied utility (Zhang et al., 20 Oct 2025).

7. Limitations, Open Challenges, and Frontier Directions

Despite broad empirical success, several challenges remain:

• Real-Time Constraints: Complex reasoning modules (e.g., LLMs, world model rollouts) can introduce prohibitively high inference latencies for high-frequency control tasks. Solutions include chunked asynchronous inference, lightweight planning heads, and cache engineering.
• Unified Representation: Integrating symbolic and high-dimensional neural perception for seamless reasoning and explainability in heterogeneous, multi-agent, or human-centric environments remains an open problem.
• Learning New Concepts: Closed-loop symbolic–neural hybrids have begun to demonstrate the capacity for self-supervised discovery of novel affordances and parts (e.g., functional handle discovery), but consistent ontological updating—and safe integration—at scale is a challenge (Pomarlan et al., 28 Jul 2025).
• Benchmark Diversity: While synthetic and simulated domains are maturing, systematic real-world, long-horizon, and open-ended manipulations remain under-explored; more comprehensive datasets such as DOM, RLBench, and challenging mobile scenarios are being developed (Xie et al., 29 Jan 2026, Liu et al., 13 Jan 2026).

In summary, perception–reasoning–action closed-loop architectures constitute a foundational design paradigm for adaptive, robust, and generalizable embodied intelligence. The field is defined by technical solutions for latency minimization, feedback-based re-planning, structured and unified representations, and integrative benchmarking, marking a decisive shift away from open-loop or batch inference methodologies (Xie et al., 29 Jan 2026, Feng et al., 19 Dec 2025, Lou et al., 16 Aug 2025, Zhang et al., 20 Oct 2025, Pomarlan et al., 28 Jul 2025, Baltieri et al., 2019).