What Matters in RL-Based Methods for Object-Goal Navigation?

An Empirical Study and A Unified Framework

ECCV 2026
1Computer Vision and Geometry Group, ETH Zürich, 2Robotics and Perception Group, University of Zurich, 3Robotic Systems Lab, ETH Zürich, 4Princeton University, 5Microsoft
Present address: École Polytechnique Fédérale de Lausanne (EPFL). All work done at the Computer Vision and Geometry Group, ETH Zürich.
Overview of the empirical study and unified framework.

We dissect modular RL-based Object-Goal Navigation into perception, policy, and test-time enhancement, measure what each contributes, and turn the findings into a state-of-the-art system.

Abstract

Object-Goal Navigation (ObjectNav) is a key capability for deploying mobile robots in everyday environments such as homes, schools, and workplaces. In this task, an agent must locate an instance of a target object category in previously unseen environments using only onboard perception, requiring the integration of semantic understanding, spatial reasoning, and long-horizon planning.

Reinforcement learning (RL) has become a dominant paradigm for ObjectNav, yet modern systems involve numerous design choices across perception modules, policy architectures, and inference-time strategies. The relative impact of these components, however, remains poorly understood. In this work, we present a large-scale empirical study of modular RL-based ObjectNav systems. We decompose the navigation pipeline into three key components — perception, policy, and test-time enhancement — and conduct extensive controlled experiments to analyze their individual contributions.

Our results suggest that improvements in perception quality and test-time strategies often yield larger performance gains than policy improvements alone. Motivated by these findings, we introduce a unified framework for systematically studying modular ObjectNav systems and build an enhanced system that achieves state-of-the-art performance on the Gibson benchmark, improving SPL by 6.6% and success rate by 2.7% over prior methods. We also introduce a human expert baseline, achieving 98% success, highlighting the significant gap between current RL agents and human-level navigation. Finally, we provide practical insights and design recommendations for each module to help guide future research.

A Unified Framework

We follow the natural pipeline of robot navigation and decompose a typical RL-based approach into three core modules, then study each under controlled settings.

Unified framework: perception, policy and test-time enhancement.

A. Perception: RGB-D and pose are fused into a top-down semantic map. B. Policy: the map and auxiliary inputs guide action prediction. C. Test-Time Enhancement: plug-and-play strategies applied at evaluation to boost performance without retraining.

Perception

An object detector extracts semantics from RGB images, fused with depth into a top-down semantic map. We study the object detector, map augmentation, and map size.

Policy

An RL policy infers a long-term goal from the map. We study the observation space, action space, network architecture, and reward design.

Test-Time Enhancement

Training-free strategies that correct common failures: an untrapping helper, dynamic goal selection, and a remapping mask for multi-floor overlap.

Key Findings

Perception dominates

The capability of the perception module has a substantial impact on overall navigation performance — a finetuned detector alone can lift Success Rate by nearly 7%.

Test-time strategies are underrated

Often dismissed as engineering details, plug-and-play test-time strategies provide significant improvements at deployment — with no extra training.

Fair policy comparison

With a unified configuration for perception and test-time strategies, differences between policy designs can finally be measured reliably and fairly.

Empirical Study

We isolate each design choice under controlled settings and report standard (SR, SPL, DTS) and dynamic-timestep (D-SR, D-SPL, D-DTS) metrics. Full tables are in the paper.

Pol.DetectorSizeAug. SR %SPL %DTS m D-SR %D-SPL %D-DTS m
CMRCNN48077.740.91.04761.638.61.641
CRedNet48080.943.90.85165.541.61.488
CRF-DETR-Seg48077.644.20.99164.142.41.512
CRF-DETR+SAM248077.144.71.11263.842.61.669
CYOLO11-XL48076.042.61.14762.840.81.668
CFT-MRCNN48083.847.60.80069.545.31.425
CFT-MRCNN24078.745.21.13564.843.11.658
CFT-MRCNN48083.344.20.77163.541.21.727
FMRCNN48071.238.51.42849.133.52.086
FRedNet48075.040.91.11453.135.71.876
FFT-MRCNN48079.945.80.89257.239.71.714
FRF-DETR-Seg48072.940.31.18353.636.61.871
FRF-DETR+SAM248072.843.71.41653.639.22.006
FYOLO11-XL48069.539.01.51449.834.62.106
FFT-MRCNN24079.643.61.03361.040.11.742
FFT-MRCNN48078.042.51.00053.436.31.791

Takeaway: a Gibson-finetuned detector (FT-MRCNN) gives a clear, consistent boost across all metrics — perception quality propagates directly to navigation. C = Corner Goal Policy, F = Frontier-Based Policy.

Obs. SpaceAction SpaceNetworkReward SR %SPL %DTS m D-SR %D-SPL %D-DTS m
StandardDiscreteTransformerType 283.443.80.71766.341.01.501
CompressedDiscreteTransformerType 283.843.30.68565.840.31.572
CompressedContinuousTransformerType 284.448.20.74169.645.81.415
CompressedDiscreteCNNType 283.743.30.70665.840.51.548
CompressedDiscreteTransformerType 185.245.80.66467.843.01.346

Takeaway: under a unified perception + test-time configuration, policy choices show smaller margins. A continuous action space helps SPL, a compressed observation matches the standard one at >6× smaller size, and CNN vs. Transformer backbones perform comparably.

PolicyUntrappingDynamic GoalRemapping SR %SPL %DTS m D-SR %D-SPL %D-DTS m
RL (Continuous)81.347.30.74169.145.31.509
RL (Continuous)82.648.00.81770.145.91.384
RL (Continuous)84.448.20.74169.645.81.415
RL (Discrete)83.843.30.68565.840.31.572
RL (Discrete)85.347.50.63269.844.91.397

Takeaway: stacking training-free test-time strategies yields clear gains — e.g., dynamic goal selection lifts the discrete policy from 83.8% to 85.3% SR — without any retraining.

Test-Time Strategies in Action

State of the Art & the Human Gap

Guided by our analysis, the best-configured system outperforms prior methods on the Gibson benchmark. A human-expert baseline — researchers under the same setup and observations — exposes a large remaining gap.

MethodSR % ↑SPL % ↑DTS m ↓
SemExp65.733.91.47
SemExp*71.739.61.39
PONI73.641.01.25
3D-aware74.542.11.16
FSE71.536.01.35
SGM78.044.01.11
L3MVN76.838.81.01
NaviFormer82.640.90.76
T-Diff79.644.91.00
HOZ++78.244.91.13
Ours (continuous)84.448.20.74
Human Experts98.053.30.26

Comparisons on Gibson (continuous-action policy shown). Our best configuration outperforms all prior methods; the discrete-action variant reaches 85.3% SR / 47.5% SPL. Human Experts reach 98% success, underscoring the algorithmic gap toward human-level navigation.

BibTeX

@inproceedings{wang2026whatmatters,
  title     = {What Matters in RL-Based Methods for Object-Goal Navigation?
               An Empirical Study and A Unified Framework},
  author    = {Wang, Hongze and Sun, Boyang and Xing, Jiaxu and Yang, Fan and
               Hutter, Marco and Shah, Dhruv and Scaramuzza, Davide and Pollefeys, Marc},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
}