PaperlayerView exact revision on arXiv

Source-linked AI summary

PONI: Potential Functions for ObjectGoal Navigation with Interaction-free Learning

Santhosh Kumar Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, Kristen Grauman

arXiv:2201.10029v2cs.CVcs.AI

TL;DR

ObjectNav methods based on reinforcement learning require substantial computation and may generalize poorly, motivating a more efficient way to decide where to search. PONI learns potential functions from passive semantic maps and integrates them into modular navigation, achieving state-of-the-art performance with up to 1,600× lower training cost.

  • Problem

    End-to-end reinforcement learning for ObjectNav is computationally expensive, sample-inefficient, and generalizes poorly because navigation skills are learned from scratch.

  • Method

    PONI uses an encoder-decoder potential function network trained with supervised learning on passive semantic maps to guide where to look within a modular ObjectNav pipeline.

  • Results

    PONI achieves state-of-the-art ObjectNav performance on Gibson and Matterport3D, with 7× and 1,600× lower training costs than prior state-of-the-art methods, respectively.

  • Takeaways & Limitations

    Treating where to look as a perception problem enables competitive ObjectNav performance without interactive learning and with substantially lower training cost.

  • Takeaways & Limitations

    PONI relies on semantic-map segmentation for deciding when to stop and lacks an inbuilt mechanism to handle segmentation failures.

Abstract

from arXiv · show

State-of-the-art approaches to ObjectGoal navigation rely on reinforcement learning and typically require significant computational resources and time for learning. We propose Potential functions for ObjectGoal Navigation with Interaction-free learning (PONI), a modular approach that disentangles the skills of `where to look?' for an object and `how to navigate to (x, y)?'. Our key insight is that `where to look?' can be treated purely as a perception problem, and learned without environment interactions. To address this, we propose a network that predicts two complementary potential functions conditioned on a semantic map and uses them to decide where to look for an unseen object. We train the potential function network using supervised learning on a passive dataset of top-down semantic maps, and integrate it into a modular framework to perform ObjectGoal navigation. Experiments on Gibson and Matterport3D demonstrate that our method achieves the state-of-the-art for ObjectGoal navigation while incurring up to 1,600x less computational cost for training. Code and pre-trained models are available: https://vision.cs.utexas.edu/projects/poni/

1. Introduction

ObjectNav asks an agent to find a specified object in a novel 3D scene, but prevailing reinforcement-learning approaches are costly and often generalize poorly. PONI treats “where to look?” as perception, learns potential functions from passive semantic maps, and achieves state-of-the-art results with substantially lower training cost.

  • ObjectNav setting: ObjectNav requires semantic reasoning to navigate efficiently to a target object and stop near it within a time budget.The task is performed in novel, unmapped 3D scenes.
  • Limitations of prior work: End-to-end reinforcement learning has high computational cost, poor sample efficiency, and weak generalization because multiple navigation skills are learned from scratch.These skills include collision-free movement, exploration, and stopping near the object.
  • PONI approach: PONI disentangles deciding where to look for an object from navigating to an (x, y) location within a modular ObjectNav framework.This follows the modular-navigation strategy of separating semantic exploration from low-level navigation.
  • PONI approach: PONI defines a 0-1-valued potential function on frontiers of a 2D top-down semantic map and selects the maximum-potential location as the next place to look.Frontiers are edges between explored and unexplored regions; the semantic map contains an object category per map location.
  • PONI approach: A convolutional encoder-decoder predicts potential functions from partially filled semantic maps and is trained non-interactively on passive map data.The method combines area and object potential predictions when choosing where to look for a goal object.
  • Results: PONI outperforms prior methods on Gibson and Matterport3D while reducing training cost by 7× versus SemExp on Gibson and 1,600× versus THDA on MP3D.It also sets the state of the art on the Habitat ObjectNav challenge leaderboard among previously published methods.

3. Approach

PONI separates ObjectNav into semantic mapping, long-term goal selection, and local navigation. Its potential-function network predicts complementary exploration and object-directed potentials from partial semantic maps, trained without environment interactions.

  • Architecture: PONI uses a semantic mapper, a potential function network, and a local policy to navigate toward object goals.The mapper builds an allocentric semantic map, the network samples a long-term goal, and analytical path-planning handles local navigation.
  • Potential functions: Potential functions are defined on map frontiers because every path to an unexplored location must pass through a frontier.The method reports that frontier-only prediction is more effective than prediction at all map locations.
  • Potential functions: The area potential values frontiers leading to more unexplored free space, providing an exploration-oriented signal.Training targets are computed from connected components of unexplored free-space cells associated with each frontier.
  • Potential functions: The object potential assigns higher value to frontiers geodesically closer to the nearest instance of the goal category.Its distance-based value decays to zero beyond a validation-selected maximum distance and resembles an A* heuristic.
  • Long-term goal sampling: The network combines area and object potentials, filters explored locations, and samples the maximum remaining location as the long-term goal.The combination uses U_t = αU_t^a + (1 − α)U_t^o with α = 0.5, while retaining predictions around noisy frontier boundaries.
  • Interaction-free training: PONI treats where to look as perception and trains the potential network with supervised learning on pre-computed semantic maps, then transfers it directly into navigation.The network predicts area and object potentials from partial maps using an encoder and two decoders, trained with pixel-wise mean-squared errors.

4. Experimental Setup

The experiments evaluate PONI on Gibson and Matterport3D using Habitat ObjectNav settings, standard metrics, and non-interactive, end-to-end RL, and modular baselines. PONI achieves state-of-the-art validation performance with substantially lower training cost.

  • Evaluation: ObjectNav performance is measured using Success, SPL, DTS, and SoftSPL.SPL measures success-weighted path efficiency, DTS measures final distance from the success threshold, and SoftSPL measures progress-based efficiency.
  • Baselines: The evaluation compares non-interactive, end-to-end RL, and modular baselines.These include behavior cloning and Predict-* methods, distributed and augmented RL methods, and modular methods such as SemExp and frontier exploration.
  • Results and cost: PONI achieves state-of-the-art validation performance on Gibson and MP3D while having one of the lowest training costs.Figure 4 plots ObjectNav SPL against effective GPU hours, and the MP3D cost axis is logarithmic.

5. Results

Across Gibson and Matterport3D, PONI outperforms prior non-interactive and modular baselines while matching or exceeding end-to-end RL performance at substantially lower training cost. Ablations and qualitative results support combining object and area potential functions, while segmentation failures remain a major error source.

  • Baseline comparisons: PONI achieves higher validation performance than the next-best non-interactive method on both Gibson and MP3D.On Gibson, it gains 3.7% success and 5.2% SPL while reducing DTS by 0.19m; on MP3D, it gains 2.4% success and 1.4% SPL while reducing DTS by 0.4m.
  • Baseline comparisons: PONI achieves 1−4% higher SPL than advanced end-to-end RL baselines on MP3D while maintaining competitive success rates.It also significantly outperforms vanilla DD-PPO on both Gibson and MP3D validation splits.
  • Baseline comparisons: PONI is the best modular method, outperforming goal-agnostic FBE and ANS across metrics and datasets and surpassing SemExp on Gibson.These comparisons support learning where to look as a perception problem without environment interactions.
  • Computational cost: 7× lower training cost than SemExp on Gibson and 1,600× lower training cost than THDA on MP3D accompany state-of-the-art performance.Training cost is measured using effective GPU hours.
  • Qualitative behavior: The area potential guides early exploration, while the object potential later directs the agent toward the bedroom and bed.In the qualitative Gibson episode, the area potential dominates from T=1 to 65, object potential rises near the bedroom entrance at T=72, and the bed is found at T=84.
  • Ablations: Removing potential-function components reduces performance, and ground-truth segmentation improves results, indicating that segmentation failures are a major source of PONI errors.The ablation studies examine object and area potentials, frontier-only prediction, and privileged ground-truth segmentation.

6. Conclusion

PONI treats where to look for an unseen object as a perception problem learned without interactions, using two potential functions within a modular ObjectNav system. Experiments show state-of-the-art performance on Gibson and Matterport3D with substantially lower training cost.

  • Conclusion: PONI predicts complementary object and area potential functions from semantic maps to choose where to look for a goal object.The potential-function network is trained supervised on semantic maps obtained from 3D semantic annotations rather than through RL reward design.
  • Conclusion: PONI achieves state-of-the-art ObjectNav performance on Gibson and Matterport3D with significantly lower training cost.The paper positions this as a compute-efficient approach for embodied navigation.

Supplementary Materials

The supplementary materials provide additional experimental details, implementation information, map-processing analyses, potential-function examples, and visualizations of ObjectNav episodes.

  • Supplementary contents: The supplement includes additional experimental details, non-interactive baseline implementations, and semantic-map masking strategies.
  • Supplementary contents: It documents action costs for long-term goal sampling and analyzes how object potential influences decisions over time.
  • Supplementary contents: Additional sections show semantic maps, potential functions, and complete ObjectNav episode visualizations, including animated trajectories on the project website.

S1. Limitations

PONI’s main limitations are dependence on semantic maps and human-annotated semantic information. Segmentation errors reduce success, while self-supervised alternatives may offer greater scalability.

  • PONI relies on the semantic map as its only source for deciding when an object is found and when to execute STOP.
  • 14.9% lower Gibson success and 45.4% lower MP3D success occur without ground-truth segmentation.The SPL-to-success ratio remains relatively stable, indicating that search efficiency is less affected than success rate.
  • Unlike end-to-end RL methods, PONI has no built-in mechanism to handle segmentation failures.The authors suggest improved segmentation, higher-quality scenes, or real-world testing as possible ways to address this limitation.
  • PONI relies on human-annotated semantic information in 3D scenes.Alternative strategies can learn ObjectNav without ground-truth semantic annotations, and such self-supervised approaches may be more scalable.

S2. Additional experimental details

The experiments use Gibson and Matterport3D environments in Habitat, with specified scene splits and validation episodes for evaluation.

  • The Gibson ObjectNav dataset covers six object categories: chair, couch, potted plant, bed, toilet, and tv.Its validation split contains 1,000 episodes from five validation scenes.

S3. Non-interactive baseline details

The non-interactive baselines use passive semantic-map supervision or behavior cloning to predict directions, locations, actions, or expert actions for ObjectNav.

  • BC: Behavior cloning trains a recurrent RGB-D policy to classify expert actions sampled by Habitat’s greedy shortest-path sampler.The policy uses a ResNet-50, pose and goal-category inputs, a 2-layer LSTM, and cross-entropy loss.
  • Predict-θ: Predict-θ uses a partial semantic map to predict eight discretized directions to the nearest object for each category.It outputs an N × 8 array of direction probabilities and uses per-category cross-entropy training.
  • Predict-xy: Predict-xy regresses normalized 0-to-1 map coordinates for the nearest object in each category from a partial semantic map.The predicted location is sampled as the long-term navigation target, using per-category mean-squared error.
  • Predict-A: Predict-A classifies the low-level action needed to reach the nearest object along the shortest path for each category.It is trained with per-category cross-entropy loss on the semantic-maps dataset.

S4. Masking strategy for semantic maps

PONI normally reveals square patches around random shortest paths in semantic maps, while an alternative viewing-cone mask approximates visibility in 3D space.

  • The default masking strategy reveals a 3m × 3m square patch around each location on randomly sampled shortest paths.
  • The alternative strategy reveals a viewing cone in front of the agent to mimic visibility in 3D space.
  • The viewing-cone strategy performs comparably with the default square strategy.The square strategy is retained as the default option.

S5. Action costs for long-term goal sampling

PONI augments long-term goal sampling with a distance potential to account for navigation cost, but this modification does not improve performance. The original strategy works because the agent typically explores one frontier sufficiently before switching.

  • Method: PONI samples long-term goals by selecting maxima of the overall potential function.The alternative incorporates the cost of navigating from the agent’s location to each candidate map location.
  • Method: The distance potential is 1.0 at the agent’s location and decreases linearly with distance.It is added to the area and object potentials as a weighted component.
  • Result: PONI does not benefit from adding action costs, based on a grid-search over MP3D (val).The comparison uses the best action-cost configuration against the current method in Table S3.
  • Interpretation: The agent typically continues exploring a single frontier before moving to another frontier.Consequently, prioritizing the best frontier regardless of distance works well in practice.
  • Analysis: The object potential increasingly influences selected goal locations during later episode stages.Figure S1 measures episodes where the selected goal differs from the area-potential maximum by at least 1 m.

S7. Examples of semantic maps

The supplementary figures show semantic maps and potential functions from Gibson and MP3D, along with qualitative ObjectNav episodes using those representations. The examples illustrate how partial maps support frontier-based area and object potentials during search.

  • Semantic maps: Gibson semantic maps contain up to 15 object categories, including 6 goal categories.The categories match the ObjectNav setup referenced in the paper.
  • Semantic maps: MP3D semantic maps contain up to 21 object categories, all treated as goal categories.The category setup matches the Habitat challenge.
  • Semantic maps: The maps are orthographic projections of 3D point-cloud annotations.This construction is used for the semantic-map datasets supporting potential-function training.
  • Potential functions: Figures S4 and S5 display complete maps, partial maps, area potentials, and object potentials for unseen objects.The potential functions are computed at map frontiers and range from 0.0 to 1.0, with red intensity indicating strength.
  • Navigation examples: Figure S6 visualizes Gibson ObjectNav episodes using egocentric views, predicted semantic maps, and area and object potential functions.Blue crosses mark the maximum location used for navigation at each step.
Loading 2201.10029v2…