- The paper presents a novel training-free executive layer that bridges the action consistency gap in zero-shot object navigation.
- It leverages a modular approach combining persistent candidate memory, finite-state control, and stability-aware action control to streamline evidence-to-action transitions.
- Experimental results demonstrate significant improvements in success rate and efficiency across MP3D, HM3Dv1, and HM3Dv2 benchmarks.
Executive Control in Zero-Shot Object Navigation: An Analysis of ConsistNav
Introduction
"ConsistNav: Closing the Action Consistency Gap in Zero-Shot Object Navigation with Semantic Executive Control" (2605.09869) addresses fundamental limitations in current zero-shot object navigation (ZSON) pipelines. The core innovation is a training-free, modular executive control layer that explicitly tackles the issue of inconsistent evidence-to-action translation: agents frequently oscillate between exploration and pursuit, or terminate prematurely, due to inadequate commitment to semantic observations. The ConsistNav framework aims to systematically bridge the so-called "action consistency gap," and demonstrates marked improvements in reliability and sample efficiency across standard benchmarks without altering the underlying detection or planning stack.
Action Consistency Gap in ZSON
Existing ZSON systems leverage open-vocabulary object detectors and vision-LLMs for semantic goal conditioning. Although such approaches facilitate zero-shot generalization across target categories, they rarely encode persistent intent or commitment at the executive level. As a result, agents misinterpret intermittent semantic cues, abandon viable candidates before confirmation, or invoke STOP commands prematurely. This critical "action consistency gap" arises because semantic evidence is repeatedly re-scored or reinterpreted in a reactive, memoryless fashion rather than being integrated into stable, temporal commitments. ConsistNav's executive explicitly encodes these semantics-to-action transformations.
ConsistNav Framework
ConsistNav is structured as a modular, perception-planning-execution stack with an executive control layer comprising three coordinated modules:
- Persistent Candidate Memory (PCM): Aggregates cross-frame detection hypotheses, encoding location, semantic confidence, accumulated positive and negative evidence, consistency score, and historical verification status. This persistent memory enables robust candidate ranking and filtering, suppressing noise from transient or false detections.
- Finite-State Executive Controller (FSEC): Implements a guarded, monotonic progression of semantic commitment states (SEARCH, SUSPECT, APPROACH, VERIFY, FINAL-APPROACH, FAILOVER, SUCCESS). State transitions are gated by candidate viability, rank persistence, distance, and verification statistics, enabling explicit and recoverable navigation commitments.
- Stability-Aware Action Control (SAAC): Constrains planner outputs according to the executive state, deploying anti-spin thresholds, stall detection, bounded recovery budgets, and STOP authorization gates that require semantic-geometric agreement. This module minimizes rotational stagnation, ineffective pursuit, and ensures multi-cue verification prior to task completion.
Notably, ConsistNav is agnostic to the choice of upstream detectors, mappers, or low-level planners, and introduces these modules without any policy retraining.
Empirical Results and Diagnostic Analysis
Experiments span the MP3D, HM3Dv1, and HM3Dv2 datasets using the standardized Habitat ObjectNav protocol. ConsistNav achieves state-of-the-art results on all benchmarks in controlled, same-stack ablations:
| Benchmark |
Baseline SR / SPL |
ConsistNav SR / SPL |
ΔSR |
ΔSPL |
| MP3D |
39.2 / 17.8 |
50.6 / 25.7 |
+11.4 |
+7.9 |
| HM3Dv1 |
59.6 / 33.0 |
63.2 / 34.8 |
+3.6 |
+1.8 |
| HM3Dv2 |
76.2 / 38.0 |
84.2 / 41.2 |
+8.0 |
+3.2 |
These improvements, achieved without modifying the perception or planning modules, directly evidence the efficacy of executive control for action consistency.
Failure cause analysis reveals a significant reduction in unstable commitments, step-limit timeouts, and premature stops. Success rates increase by up to 11.34% (MP3D), while timeouts are reduced by up to 10.21%. Weak candidates are explicitly recognized as search failures rather than producing ambiguous or incorrect task outcomes, reflecting a conservative and interpretable control policy.
Real-world deployment using the AgileX LIMO platform further demonstrates the robustness and transferability of the approach under physical actuation and environmental noise. The mean task duration remains competitive, and verified successes are consistent with simulation outcomes.
Ablation Studies
Ablation analysis confirms complementary gains from each executive module. PCM stabilizes hypotheses across frames but cannot enforce commitment or recovery. FSEC yields the largest jump in reliability by enabling guarded intention escalation, while SAAC further addresses action-level bottlenecks such as rotational stagnation and verification-based termination.
- PCM alone: +0.9% SR, +1.6% SPL
- PCM + FSEC: +3.9% SR, +1.15% SPL (over previous)
- Full (PCM + FSEC + SAAC): +8.0% SR, +3.24% SPL (over baseline)
Each component eliminates a different class of evidence-to-action inconsistency, confirming the benefit of an explicitly partitioned executive layer.
Theoretical and Practical Implications
ConsistNav demonstrates that a lightweight, modular executive can offer strong, generalizable reliability gains in zero-shot object navigation, decoupled from upstream or downstream learning components. This approach recapitulates classical planning abstractions—goal commitment, guarded transitions, and recovery loops—in a modern VLM-driven pipeline, thereby enabling deeper diagnostics, safer deployment, and interpretable failure handling.
Theoretically, ConsistNav suggests that persistent, stateful executive control is essential for closing the gap between open-ended perception and robust execution in modular navigation stacks, especially as embodied agents move toward broader, open-vocabulary semantic goals.
Practically, the method provides a path toward deployment-ready navigation systems that remain robust to perceptual noise and environmental uncertainty, without requiring retraining for new object categories or environments.
Limitations and Future Directions
Current system limitations include dependence on an off-board inference server during real-world deployment and sensitivity to simulator-planner-executor interface details. Future work will explore adaptive state transition thresholds and generalization to multi-target or instruction-conditioned navigation, where commitment reasoning may involve multi-objective trade-offs and dynamic goal hierarchies.
Conclusion
ConsistNav targets and closes the action consistency gap in zero-shot object navigation through a modular, training-free executive architecture that integrates persistent memory, finite-state commitment, and action stabilization. Quantitative and ablation results as well as real-world demonstrations indicate that executive control—rather than improved perceptual or planning modules—constitutes a decisive factor in advancing embodied zero-shot navigation. This explicit, interpretable approach lays the groundwork for more reliable and deployable semantic navigation agents.
