Source-linked AI summary
Rapid Exploration for Open-World Navigation with Latent Goal Models
Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, Sergey Levine
TL;DR
RECON addresses visual goal-directed navigation in previously unseen, open-world environments using prior experience. It combines an information-bottleneck latent goal model with incrementally built topological memory to explore and navigate from images. In experiments, it discovered user-specified goals up to 80m away in under 20 minutes, while the authors note that it does not explicitly account for the value of information.
Problem
Robustly navigating to user-specified goals in previously unseen environments requires using prior experience while handling distractors, new obstacles, and exploration.
Method
RECON combines an information-bottleneck representation of visual goals with a non-parametric topological memory built incrementally during image-based exploration.
Results
RECON discovered user-specified goals up to 80m away in under 20 minutes across 8 open-world environments, outperforming the evaluated baselines.
Takeaways & Limitations
The learned goal model and topological graph let RECON use prior experience to rapidly explore new environments and later navigate to arbitrary goals.
Takeaways & Limitations
RECON does not explicitly account for the value of information when selecting states, which the authors identify as a limitation.
Abstract
from arXiv · showhide
We describe a robotic learning system for autonomous exploration and navigation in diverse, open-world environments. At the core of our method is a learned latent variable model of distances and actions, along with a non-parametric topological memory of images. We use an information bottleneck to regularize the learned policy, giving us (i) a compact visual representation of goals, (ii) improved generalization capabilities, and (iii) a mechanism for sampling feasible goals for exploration. Trained on a large offline dataset of prior experience, the model acquires a representation of visual goals that is robust to task-irrelevant distractors. We demonstrate our method on a mobile ground robot in open-world exploration scenarios. Given an image of a goal that is up to 80 meters away, our method leverages its representation to explore and discover the goal in under 20 minutes, even amidst previously-unseen obstacles and weather conditions. Please check out the project website for videos of our experiments and information about the real-world dataset used at https://sites.google.com/view/recon-robot.
1 Introduction
RECON addresses robust goal-directed navigation in unseen environments by combining prior experience, compact visual goal representations, and topological memory. It discovers distant user-specified goals through image-based exploration in real-world settings.
- Robust navigation must generalize across training and real-world environments, ignore visual distractors, and actively explore for traversability.
- RECON uses a large cross-environment experience dataset, an information bottleneck, and non-parametric memory to learn navigation in new environments.
- Its compressed goal representation supports robustness to appearance changes, such as navigating to a goal image captured at another time of day.
- RECON was evaluated against 6 competitive baselines across over 100 hours of experiments in 8 open-world environments.
- 80m-away user-specified goals were discovered after 20 minutes of interaction in a new environment.
2 Related Work
Prior navigation and exploration methods use geometric maps, local or global exploration, novelty signals, and transfer learning, but RECON targets visual goal navigation in unseen environments using prior experience across environments.
- Geometric mapping methods can require reliable depth and may omit semantic traversability distinctions such as grass versus wire fences.
- Topological navigation methods learn distances and local policies, but prior work does not explain navigation in new, unseen environments.
- Reinforcement-learning exploration methods use novelty, information gain, counts, densities, or distances from visited states to guide exploration.
- RECON uses an information bottleneck both to support generalization and to measure task-relevant uncertainty for proposing functionally relevant goals.
- Unlike prior goal-conditioned imitation methods, RECON chooses goals in a new environment while using experience from different environments to accelerate learning.
3 Problem Statement and System Overview
The task is visual goal-directed exploration in a novel environment, organized around offline learning, topological map construction, and navigation to goals. RECON operates primarily from onboard images while using other sensors only for safety during data collection.
- The robot must navigate to a goal location given an image observed at that location, across offline learning, map building, and goal navigation stages.
- RECON runs on a Clearpath Jackal with an RGB camera, while GPS and lidar serve only as safety controllers during data collection.
- Its offline self-supervised trajectory dataset spans diverse environments, 18 months, and substantial seasonal and lighting variation.
4 RECON : A Method for Goal-Directed Exploration
RECON combines a context-conditioned latent goal model with an incrementally built topological memory to explore unseen environments and reach user-specified image goals. It alternates feasible-goal navigation, frontier exploration, graph expansion, and model fine-tuning, then plans through the graph for goal reaching.
- Core components: RECON learns an uncertainty-aware, context-conditioned goal representation and a topological map whose nodes are observations and edges are predicted distances.The representation adapts to novel scenes, while the map is constructed incrementally during exploration.
- Learning goal representations: The latent goal model compresses goal images while retaining information about relative goal location needed to predict actions and distances.Its encoder and decoder are conditioned on the current observation, supporting feasible-goal representations rather than arbitrary image reconstructions.
- Building memory: The topological memory grows as the robot executes latent subgoals, collects trajectories, expands the graph, and fine-tunes the model in the new environment.Efficient exploration is reduced to selecting subgoals for data collection.
- Goal reaching: After exploration, RECON uses the topological graph to plan a path of image-based subgoals toward the user-specified goal.The goal-reaching procedure associates current and goal observations with graph vertices and follows a shortest graph path.
- Exploration policy: During exploration, RECON navigates directly to feasible goals, samples latent subgoals at frontiers, or moves toward the least-explored neighboring node.Frontier status is estimated from predicted distance to the least-explored neighbor, while visitation counts identify underexplored nodes.
5 Experimental Evaluation
RECON is evaluated against competitive exploration and navigation baselines across unseen outdoor environments, with tests covering goal discovery, goal recall, robustness to distractors, exploration strategy, and component ablations.
- Goal-directed exploration: RECON reaches goals 50% faster than the best baseline, ECR, across 8 real-world environments.ANS takes up to 2x longer, while NTS fails to find the goal in every environment.
- Goal reaching: On subsequent traversals, RECON navigates to discovered goals 20–85% faster than other baselines and achieves over 30% higher weighted success.Its compact topological map enables recall of feasible paths to previously seen states.
- Goal reaching: Only RECON and ECR reach the goal successfully in the illustrated environment, with RECON taking the shorter route.The comparison visualizes trajectories and learned policies after exploration.
- Goal-directed exploration: RECON discovers goals up to 80m away in under 20 minutes, including cases where no other baseline succeeds.RECON+ECR and ViNG work only in easier environments, while RND, InfoBot, and ANS fail beyond 25m.
- Robustness: RECON navigates successfully under previously unseen obstacles and lighting conditions without additional exploration.The learned representation discards task-irrelevant visual changes such as lighting differences.
- Exploration strategy: Rollouts toward sampled goals produce 5x faster exploration than random action sequences in novel environments.The comparison measures trajectory coverage during 5 minutes of undirected exploration.
- Component analysis: Ablations indicate that compressed goal representations and non-parametric memory are important for RECON’s performance.Removing the bottleneck harms distant-goal discovery, while removing memory severely degrades recall of discovered goals.
6 Discussion
The system uses a learned goal-conditioned distance model with latent visual goals to accelerate goal-directed exploration in new open-world environments. A current limitation is that it does not explicitly account for the value of information when selecting states.
- Discussion: The method combines a learned goal-conditioned distance model with a latent variable representation of visual goals for rapid exploration.Prior experience helps guide learning in new environments where deployment-time data collection is costly.
- Discussion: Using past experience to accelerate learning in new environments reflects real-world robotics settings where collecting deployment data is costly.
- Limitations: The current method does not explicitly account for the value of information, which could improve goal-reaching policies.
A Dataset
The dataset contains over 5000 self-supervised trajectories from 9 diverse real-world environments and is publicly available for download. Figure 7 illustrates example trajectories collected across those environments.
- Dataset: The dataset contains over 5000 self-supervised trajectories collected across 9 distinct real-world environments.The trajectories are intended to provide supervision from prior experience, including interactions collected in diverse environments.
- Dataset: The dataset and helper scripts for loading and visualizing trajectories are publicly available online.
- Dataset: Figure 7 shows example trajectories collected in 9 diverse environments.
A.1 Self-Supervised Data Collection and Labeling
The data collection pipeline uses off-policy learning and a time-correlated random walk to gather diverse, realistic trajectories with minimal human intervention. Collision and stuck detectors support automated recovery and self-supervised event labeling.
- Self-Supervised Data Collection and Labeling: Off-policy learning allows data gathered with any control policy to be reused for training, reducing the need for costly real-world data collection.
- Self-Supervised Data Collection and Labeling: Figure 8 compares the time required to discover goals and reach them again after a map is constructed.Lower values are better for both search time and second-time goal-reaching time; hashed bars indicate failure during search.
- Self-Supervised Data Collection and Labeling: A time-correlated random walk improves environment coverage and produces realistic action sequences compared with naive uniform random control.
- Self-Supervised Data Collection and Labeling: Collision and stuck detectors trigger an automated backup maneuver to help the robot recover and continue data collection.
- Self-Supervised Data Collection and Labeling: The collision detectors provide weak supervision by generating event labels for collisions, bumpiness, and position.
A.2 Environments
The evaluation uses trajectories from 9 diverse open-world environments, with performance compared across environments of increasing complexity. As complexity increases, most methods explore less efficiently to discover goals.
- Environments: The evaluation curates over 40 hours of trajectories from 9 diverse open-world environments with varying complexity.
- Environments: As environment complexity increases, most methods become unable to explore efficiently enough to discover the goal.
B Reproducibility
RECON navigates toward latent subgoals with fixed-horizon policy rollouts, while a non-parametric graph tracks visitation and expands as trajectories end. Frontier progress is encouraged by selecting locally least-explored neighbors.
- SubgoalNavigate rolls out the learned policy for a fixed horizon H toward a desired subgoal latent.
- Each rollout endpoint updates graph visitation counts through AssociateToVertex, associating the endpoint observation with its nearest graph vertex.
- LeastExploredNeighbor selects the neighboring graph vertex with the smallest visitation count to nudge the robot toward unexplored local areas.
- At the end of each trajectory, ExpandGraph adds a new observation node and predicted-distance edges to update the graph’s node and edge sets.
- Algorithm 3 specifies the SubgoalNavigate, AssociateToVertex, LeastExploredNeighbor, and ExpandGraph subroutines used by the exploration algorithm.
B.2 Implementation Details
The implementation encodes current-goal image pairs into stochastic, context-conditioned goal latents and decodes them with current observations to predict actions. Training uses MobileNet-based components, Adam optimization, and documented architectural and hyperparameter settings.
- The encoder receives current and goal observations as two stacked RGB camera images at 160 × 120 pixels.
- Tables 3 and 4 document RECON’s experimental hyperparameters and architectural details, including RGB image inputs ot ∈ [0, 1]^(3×160×120).
- A MobileNet encoder and fully connected projection produce a stochastic, context-conditioned goal representation with 64-dimensional mean and diagonal-covariance parameters.
- The decoder processes the current observation with another MobileNet and combines it with the sampled goal latent to learn optimal actions.
- During pretraining, the model uses batch size 128 and Adam optimization with learning rate λ = 10^-4 until convergence.