Occupancy Anticipation for Efficient Exploration and Navigation

Santhosh K. Ramakrishnan; Ziad Al-Halah; Kristen Grauman

Source-linked AI summary

Occupancy Anticipation for Efficient Exploration and Navigation

Santhosh K. Ramakrishnan, Ziad Al-Halah, Kristen Grauman

arXiv:2008.09285v2Aug 21, 2020cs.CV

TL;DR

Existing navigation maps are limited to occupancy directly observed by the agent, motivating a way to infer unseen regions from egocentric observations. The paper introduces occupancy anticipation and integrates it with sequential exploration and navigation policies. On Gibson and Matterport3D, the approach improves mapping and navigation over strong baselines, including a 6.3 SPL-point margin in the 2020 Habitat PointNav Challenge.

Problem
State-of-the-art navigation maps encode only observed regions, so agents must directly see evidence to classify space as free or occupied.
Method
The framework predicts occupancy beyond the visible field from egocentric RGB(D) context, aggregates predictions over time, and incorporates anticipation into exploration and navigation policies.
Results
The approach improves exploration mapping and navigation over strong baselines on Gibson and Matterport3D, winning the 2020 Habitat PointNav Challenge by 6.3 SPL points.
Takeaways & Limitations
Anticipating navigable areas beyond the field of view produces more accurate maps more efficiently in novel environments and improves sequential exploration and navigation.
Takeaways & Limitations
Evaluation uses simulated actuation and odometry noise based on approximate values provided by prior work.

Abstract

from arXiv · show

State-of-the-art navigation methods leverage a spatial memory to generalize to new environments, but their occupancy maps are limited to capturing the geometric structures directly observed by the agent. We propose occupancy anticipation, where the agent uses its egocentric RGB-D observations to infer the occupancy state beyond the visible regions. In doing so, the agent builds its spatial awareness more rapidly, which facilitates efficient exploration and navigation in 3D environments. By exploiting context in both the egocentric views and top-down maps our model successfully anticipates a broader map of the environment, with performance significantly better than strong baselines. Furthermore, when deployed for the sequential decision-making tasks of exploration and navigation, our model outperforms state-of-the-art methods on the Gibson and Matterport3D datasets. Our approach is the winning entry in the 2020 Habitat PointNav Challenge. Project page: http://vision.cs.utexas.edu/projects/occupancy_anticipation/

1 Introduction

Existing navigation maps encode only observed regions, forcing agents to see occupancy evidence directly. This paper introduces occupancy anticipation from egocentric RGB(D) context and integrates it with exploration and navigation policies to improve mapping and movement in 3D environments.

Problem: State-of-the-art navigation maps capture geometry and semantics but encode only regions directly observed by the agent.These maps exploit prior knowledge in training scenes yet remain limited to visible evidence in novel environments.
Approach: Occupancy anticipation infers whether unseen regions are free or occupied from visual context in egocentric views.The model can extrapolate structures such as walls, corridors, and nearby free space beyond the current observation.
Approach: The proposed framework aggregates anticipated occupancy maps over time and incorporates the anticipation objective into the exploration policy.The policy encourages movements that yield broader and more accurate inferred maps.
Evaluation: The approach is evaluated on Gibson and Matterport3D, datasets covering more than 170 real-world spaces with varied obstacles and floor plans.Experiments assess exploration mapping and navigation using RGB(D) inputs.
Results: Occupancy anticipation improves exploration-map speed and accuracy and enables more efficient navigation than strong baselines.The reported results include faster, more accurate maps and improved navigation under noisy actuation and sensing.
Relation to prior work: The work addresses prior scene-completion and occupancy-extrapolation limits by inferring occluded regions with standard RGB(D) cameras and semantic-geometric context.Earlier approaches either did not extrapolate unseen regions, could not infer occluded areas, or assumed wide-FoV LIDAR and simple 2D layouts.

3 Approach

The approach anticipates occupancy beyond directly visible regions from egocentric RGB-D observations, then integrates predictions over time for exploration and mapping. It combines RGB-D feature processing with a training reward based on correctly anticipated global-map occupancy and integrates the model into ANS.

3.1 Occupancy anticipation model: Occupancy anticipation infers unseen free and occupied regions from egocentric RGB-D context, including occluded or out-of-view areas.Predictions cover a fixed local region in front of the agent and are integrated over time as it moves.
3.1 Occupancy anticipation model: The model represents egocentric occupancy as a two-channel top-down map whose channels encode occupancy and exploration probabilities.Each map cell represents a 25mm × 25mm region; training targets come from indoor-environment 3D meshes, including occluded and out-of-view regions.
3.1 Occupancy anticipation model: RGB-D feature extraction combines semantic and textural RGB features with a depth-derived visible occupancy map before UNet-based encoding, merging, and decoding.Depth points within approximately 3m are processed into initial occupied, free, and unknown space; the decoder outputs a 2 × V × V probability tensor.
3.2 Anticipation reward for exploration policy learning: The anticipation-based exploration reward increases when predicted global-map occupancy becomes more accurate, regardless of whether the locations were directly observed.Global predictions are registered using pose estimates, and the reward is the increase in map accuracy from time t−1 to t.
3.2 Anticipation reward for exploration policy learning: The reward encourages agents to avoid spending time directly observing correctly anticipated regions and supports learning under noisy pose and map predictions.Correctly anticipating free space behind a table can redirect exploration toward regions that are harder to anticipate; the reward also accounts for mapping errors from noisy estimates.
3.3 Exploration and navigation with occupancy anticipation: The system adapts Active Neural SLAM by replacing its mapper projection unit with occupancy anticipation and replacing area-coverage reward with the anticipation-based reward.ANS retains its mapper, planner, local policy, and global policy structure; reward signals are used during training.

4 Experiments

Experiments evaluate occupancy anticipation for local mapping, exploration, and navigation across Gibson and Matterport3D. OccAnt broadens map coverage beyond visible regions and generally improves mapping speed, map quality, and navigation performance over fixed-backbone baselines.

Experimental setup: OccAnt predicts occupancy beyond directly observed regions, using depth, RGB, or full RGB-D inputs within a common ANS-based architecture.The evaluated variants are OccAnt(depth), OccAnt(rgb), and OccAnt(rgbd), with optional anticipation-reward ablations.
4.2 Occupancy anticipation results: OccAnt models substantially improve occupancy-map quality and extent over all baselines on the Gibson validation set.OccAnt(depth) outperforms OccAnt(rgb) across metrics, while RGB-D is not better than depth-only in this evaluation.
4.3 Exploration results: OccAnt generally reaches higher map quality more rapidly than baselines, both as episode duration increases and for a fixed area seen.OccAnt(depth) outperforms ANS(depth) in all reported exploration cases, and OccAnt(rgb) converges faster than ANS(rgb) on a same-modality comparison.
4.3 Exploration results: The anticipation reward generally benefits noisy-setting map accuracy and alleviates declines caused by accumulated pose-estimation errors during map registration.In noise-free conditions, adding the anticipation reward does not improve performance in the same way.
4.3 Exploration results: At T=500, OccAnt obtains significant map-accuracy improvements, while perfect anticipation yields +10-20% IoU gains in noise-free exploration.The comparison supports occupancy anticipation as a source of faster exploration and mapping, with the learned model moving toward the perfect-anticipation ceiling.
4.4 Navigation results: OccAnt outperforms navigation baselines on Gibson, enabling faster movement to specified goals by modeling navigable spaces.Ground-truth anticipation produces additional navigation gains, and the learned methods bridge part of the gap between prior methods and perfect anticipation.
4.4 Navigation results: OccAnt achieves 0.8 SPL in 2M frames, whereas DD-PPO requires more than 50× the experience to reach the same performance.DD-PPO reports 0.96 SPL with 2.5 billion frames, but fails under noisy conditions; the paper attributes OccAnt’s sample efficiency to explicit mapping and occupancy anticipation.
4.4 Navigation results: OccAnt wins the 2020 Habitat PointNav Challenge under noisy RGB-D sensors and actuators without an odometer, while DD-PPO obtains 0 SPL in that setting.The challenge setup is more difficult than evaluations assuming perfect odometry and noise-free sensing and actuation.

5 Conclusion

Occupancy anticipation enables agents to predict navigable areas beyond their field of view, producing more accurate maps efficiently across exploration and navigation tasks.

Occupancy anticipation predicts navigable areas beyond the agent’s actual field of view.
The approach supports both individual local maps and sequential models for exploration and navigation.
The experiments report advantages across multiple datasets, including improvements over state-of-the-art embodied AI exploration and navigation models.

Occupancy Anticipation for Efficient Exploration and Navigation Supplementary Materials

The supplementary materials provide experimental details and supporting qualitative and quantitative results for occupancy anticipation, exploration, and navigation.

The supplement contains additional information about experimental settings and supporting qualitative and quantitative results.
Its sections cover noise-free exploration, ablations, qualitative examples, ground-truth generation, noise models, implementation differences, and architecture.
The supplement also includes view extrapolation, model-capacity comparisons, and Habitat Challenge 2020 materials.

S1 Noise-free exploration results

Under noise-free conditions, OccAnt rapidly achieves higher map accuracy than baselines, while anticipation reward has limited additional effect because noise correction is unnecessary.

Anticipation reward largely retains original performance in noise-free conditions rather than improving it.Its main benefit appears under noisy conditions, where it improves pose-estimate noise correction and map registration.
Because noise correction is unnecessary without noise, anticipation reward has limited impact on final performance.
OccAnt rapidly achieves higher map accuracy than baselines as episode duration or area seen increases.This pattern is reported for Gibson small and large splits and Matterport3D under noise-free conditions.

S2 Occupancy anticipation ablation study

Ablations separate the effects of occupancy anticipation and anticipation reward: anticipation improves mapping and navigation broadly, while reward benefits are strongest for noisy exploration.

Study design: The ablations group results by RGB, depth, and RGB-D modality and by use of occupancy anticipation or anticipation reward.
Per-frame maps: Adding occupancy anticipation significantly improves per-frame map IoU and F1 scores over the base model.
Exploration: Occupancy anticipation generally yields better exploration map quality than ANS across modalities and testing conditions.
Exploration: Anticipation reward significantly improves noisy exploration map quality for depth and RGB-D modalities, while RGB slightly underperforms.The improvement is attributed to better mapper training and map registration.
Navigation: Adding occupancy anticipation significantly improves all three navigation metrics.The effect of anticipation reward is limited because PointNav evaluation assumes noise-free test conditions.

S3 Occupancy anticipation qualitative examples

Qualitative examples show that occupancy anticipation improves exploration and mapping efficiency, while prediction errors and ambiguous layouts remain important failure modes. Anticipation reward further improves map registration under noisy conditions.

Adding occupancy anticipation to ANS improves performance, while anticipation reward improves mapping in noisy cases.
Occupancy anticipation avoids local navigation difficulties and achieves higher map quality with lower area coverage in some episodes.
Baseline methods cannot account for occluded or out-of-range regions, whereas view extrapolation often produces noisy predictions and omits occlusions.
OccAnt(rgbd) incorrectly marks narrow corridors as occupied and struggles when multiple scene layouts are plausible.
Incorrect visible-occupancy estimates can cause ANS(depth) to become stuck in local regions during exploration.
False negatives in anticipated occupancy classify explored obstacles as free space, reducing map quality without reducing area seen.
Anticipation reward helps the model account for map-registration noise compared with a vanilla anticipation model without that reward.

S5 Generating ground-truth for occupancy anticipation

The ground-truth generation pipeline projects visible depth into occupancy, then selectively grows regions around visible areas using the 3D environment layout. Mapping robustness is additionally addressed through temporal averaging and entropy filtering of predictions.

Ground-truth local occupancy is obtained from 3D meshes over a V × V region that can include occluded or out-of-view areas.
Visible occupancy is projected geometrically from depth, then a mask is grown through hole filling and morphological dilation around visible regions.
Training and evaluation simulate realistic actuation and odometry using Habitat’s PyRobot noise model with truncated Gaussian rotation and translation noise.
The implementation replaces channelwise max-pooling with a moving average to combine local maps over time and reduce prediction and registration errors.
EntropyFilter masks low-confidence occupancy predictions before temporal aggregation because difficult or ambiguous regions can admit multiple solutions.

S8 Implementation details

The implementation details specify the policy and mapper configuration and describe the RGB feature-extraction component used within the occupancy anticipation architecture.

Policy and mapper hyperparameters are specified in Table S4.
The supplementary architecture descriptions follow PyTorch conventions and represent tensor shapes as (C, H, W).
ConvBR combines convolution, batch normalization, and ReLU layers, while MaxPool, Conv, AvgPool, and 2x Upsample denote standard corresponding operations.
RGB images are processed with ResNet18 blocks, then transformed into compatible top-down RGB features.

S10 ANS projection unit architecture

The ANS projection baseline predicts visible occupancy from egocentric features, whereas the proposed architecture encodes and merges RGB and occupancy features before decoding anticipated occupancy. View extrapolation instead predicts depth for offset cameras to expand field of view.

Occupancy anticipation architecture: The proposed feature encoder processes RGB features and visible occupancy through independent UNet encoding layers.
Occupancy anticipation architecture: MERGE combines RGB and visible-occupancy features layerwise, using visible-occupancy features alone where RGB features are unavailable.
Occupancy anticipation architecture: The anticipation decoder uses a UNet decoder with residual connections to produce the anticipated occupancy map.
ANS projection unit: The ANS(rgb) projection unit extracts ResNet-18 features, applies fully connected layers, and decodes visible occupancy estimates.
View extrapolation baseline: View extrapolation predicts depth for left and right cameras offset by 45° to achieve effective 180° coverage from a 90° egocentric view.
View extrapolation baseline: The view-extrapolation model independently encodes RGB and depth, concatenates the features, and uses a UNet to regress the offset-camera depth predictions.

S12 Comparing the model capacities of different methods

For the 2020 Habitat Challenge, the system addressed missing GPS+Compass, noisy sensing, and noisy actuation through pose estimation, noise-aware occupancy anticipation, and safer navigation. RGB-D pose estimation was computationally efficient, while OccAnt(depth) performed best under noisy sensing.

Challenge changes: The 2020 PointNav track increased difficulty through the lack of GPS+Compass and the introduction of noisy sensing and actuation.The challenge simulated RGB-D sensing noise and actuation noise based on sensor and LoCoBot noise models.
System adaptations: An online RGB-D pose estimator tracks the agent’s pose by estimating relative pose changes between consecutive observations and summing them over time.Compared with the original ANS model, RGB-D inputs produced slightly better estimates and were more computationally efficient than top-down maps.
System adaptations: End-to-end training on noisy inputs produced accurate occupancy predictions, with OccAnt(depth) giving the best performance.Adding RGB information to occupancy anticipation did not lead to significant performance changes.
System adaptations: To mitigate actuation-induced pose drift, the system used safe planning that maintained distance from obstacles while planning shortest paths.The authors report that reducing collisions led to faster navigation, and also incorporated collision-based occupancy updates and an analytical local policy.