OpenSpace Lab Solution to the IROS 2026 Indoor Exploration Competition

arXiv:2610.01505 · cs.RO · Submitted 2026-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "OpenSpace Lab Solution to the IROS 2026 Indoor Exploration Competition".

Dev: This report presents OpenSpace Lab’s solution to the Competition on Intelligent Information Gathering for Single and Multi-Robot Systems Workshops at IROS 2026,

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, to recap, this paper outlines OpenSpace Lab’s solution to the IROS two thousand twenty-six Indoor Exploration Competition for both single and multi-robot systems. The central thesis is that they tackle exploration by using pre-trained map completion predictions for global planning to prioritize unexplored areas.

Dev: They claim success by introducing a remaining-time-based exploration strategy that cleverly integrates homing constraints directly into the decision-making process to ensure timely return and data delivery regardless of the mission length.

Taro: Why does this matter beyond just winning a competition? I'm thinking about real-world applications; if this system can reliably map an unknown indoor space, what kind of infrastructure could it help build or maintain?

Rosa: It matters because they achieved top rankings in both the single-robot public track and the private tracks for both single and multi-robot systems. This shows a robust framework for intelligent information gathering under strict operational constraints.

Dev: The framework's ability to use utility-driven selection for multi-agent coordination, combined with budget management, is what really sets it apart from simpler pathfinding methods.

Taro: I’m interested in the specific mechanism of how they integrate map completion predictions into the global planner; does that predictive element offer any advantages over purely reactive exploration methods?

Rosa: The pre-trained map completion runs upon initial target selection and updates every eight decision steps, which informs the planning process by helping to prioritize areas that are likely unexplored based on prior knowledge.

Dev: And they use those predictions to generate candidates based on current observation, utility checks, and path feasibility before comparing them against the remaining step budget configuration.

Taro: So it's not just about finding a path; it's about intelligently selecting *where* to go next based on predicted information gain and cost simultaneously?

Rosa: Precisely, and they use those predictions to reconcile map coverage with the limited operation time by tying the exploration goal selection directly to the remaining step constraint.

Dev: The paper claims effectiveness in handling various budget settings—one thousand one thousand five hundred and two thousand steps—showing adaptability across different operational timelines.

Taro: It's impressive that they managed to develop a system that handles these varying time constraints so gracefully while still maintaining high exploration rates.

Rosa: That adaptability extends to the multi-robot extension where they incorporate shared maps and planned paths alongside target adjustments based on teammate information for fleet coordination.

Dev: So, even in a multi-robot setting, they aren't just running independent explorations; they are coordinating their efforts around shared goals and historical trajectories.

Taro: That level of coordination suggests a system capable of handling complex search patterns where redundant effort needs to be actively managed by the agents themselves.

Rosa: The whole point is that this approach provides a structured, time-aware mechanism for exploration that works well across different scales of robotic systems and environmental complexity.

Conclusion: Rosa: Looking at the "OpenSpace Lab Solution to the IROS two thousand twenty-six Indoor Exploration Competition" title, it really captures the essence of what they built: a complete solution addressing a specific competition challenge. The authors are Yuxuan Zhang, Dong Li, Zezhou Sun, Yuxuan Xu, Siyu Teng, Yuchen Li, and Jianjian Yang.

Dev: I think the implication for us is that this framework moves exploration from being purely reactive to being proactive by using predictive models to guide where the robot should go next.

Taro: Proactive planning based on predictions could mean that autonomous systems could navigate unknown environments much more efficiently, potentially leading to quicker deployment in areas like disaster response or infrastructure inspection.

Rosa: Exactly; if this AI can intelligently prioritize unexplored areas based on predicted structure and cost, those systems could cover vast indoor spaces in significantly less time than current methods.

Dev: And the multi-robot coordination aspect suggests a future where fleets of robots can operate as a cohesive unit, sharing knowledge dynamically instead of just following pre-programmed assignments.

Taro: I see that capability translating into complex scenarios where multiple agents need to work together to cover an area efficiently, which is something that's really needed for large-scale mapping projects.

Rosa: The paper shows how to manage the tension between needing thorough exploration and the physical limitations of time and energy in a way that seems quite practical for real deployment.

Dev: It gives us concrete engineering insights into how to design a system where communication management and return constraints are baked into every decision point, which is crucial for reliable autonomous operation.

Taro: The focus on budget awareness means we're looking at systems that can adapt their behavior based on how much time they have left to complete their task.

Rosa: That adaptability, combined with the predictive map completion, suggests that future robots won't just be traversing spaces; they will be actively building a model of the space while simultaneously optimizing their mission execution in real-time.

Yuxuan Zhang, Dong Li, Zezhou Sun, Yuxuan Xu, Siyu Teng, Yuchen Li, Jianjian Yang

China University of Mining and Technology-Beijing

cs.RO

Submitted: 2026-10-01

Updated: 2026-10-01

Comments: https://github.com/OpenSpace-Lab/Indoor-Exploration-IROS2026

Code: https://github.com/OpenSpaceLab/Indoor-Exploration-IROS2026

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: This report presents OpenSpace Lab’s solution to the Competition on Intelligent Information Gathering for Single and Multi-Robot Systems Workshops at IROS 2026, detailing strategies that

Key concepts

Map Completion
This involves using a pre-trained model (LaMa) within MapEx to intelligently fill in missing parts of an indoor map. This is done before planning starts, providing a more complete initial understanding of the environment, which helps robots plan better routes and identify unexplored areas efficiently.
Budget-Aware Engineering Heuristics
These are adaptive rules that change how robots select targets and paths depending on their remaining steps (budgets). For example, short budgets trigger rules to avoid backtracking, while long budgets use features like geodesic distance advantage to prioritize distant targets for better coverage.
Decentralized Dispersion
In multi-robot settings, this strategy tells agents to spread out into different angular sectors around the base station when starting a mission. This prevents robots from overlapping too much in the initial exploration phase, ensuring broader and more efficient coverage across the entire area.
Homing Condition
This is the trigger for returning to base. It happens when a robot's remaining budget drops below an estimate of its return cost plus a safety margin. The cost calculation considers travel time, reaching communication core, and ensuring line-of-sight to the base station is re-established before final upload.

Terminology

Summary

This report presents OpenSpace Lab’s solution to the Competition on Intelligent Information Gathering for Single and Multi-Robot Systems Workshops at IROS 2026, detailing strategies that successfully achieved top rankings in both single-robot and multi-robot exploration tracks. The core contribution lies in developing time-aware exploration frameworks that integrate pre-trained map completion for planning and utility-driven selection for multi-agent coordination, demonstrating effective coverage rates of 61.04% in the single-robot public track.

Single Robot Exploration Strategy

The single-robot method combines learning augmented global task ordering with fine-grained local planning, while dynamically adapting execution heuristics across different step budgets. The global planner utilizes the pre-trained LaMa model within MapEx for neural map completion, which runs upon initial target selection and updates every eight decision steps. Candidate frontier locations are organized into macroregion tasks labeled by openness, shape, and narrow connectivity. To evaluate visiting sequences π = (q0,..., qL−1), a beam search maximizes the sequence score S(π) defined in Equation (1), which accounts for information gain, predicted structure, travel cost, and homing burden, while enforcing budget feasibility by verifying cumulative travel and homing overhead against the remaining step budget.

Budget-Aware Engineering Heuristics

The team incorporates adaptive engineering heuristics into target selection and pathing to maximize efficiency across three step budgets (1,000, 1,500, and 2,000 steps). Specific rules are applied based on the budget: for short budgets (1,000 steps), the policy actively identifies alternative goals when route overlap exceeds 75% (≥ 80 overlapping steps) to minimize back-tracking overhead. Medium-budget runs (1,500 steps) scale this threshold to 16% and defer lateral branch exploration until after the initial 30% phase (450 steps). Long-budget settings (2,000 steps) adjust late-stage target sorting by integrating a distance-advantage feature derived from coarse-grid geodesic distances, scaled by a late-stage weight of 0.18.

Homing Condition and Data Upload

Final homing is triggered when the remaining budget falls below the estimated return cost plus a safety margin (10 steps). The homing cost H is evaluated using Equation (2): H = max (Hpath, max (0, Hbase − ccore)) + 1. This includes evaluating Hpath to the first LOS-communicable grid, Hbase to the base station grid, and ccore offsets for the cached reachable communication core. Communication is restored as soon as line-of-sight to the base station is re-established, allowing immediate map upload before physically reaching the base grid.

Multi-Robot Extension

The multi-robot strategy extends localized exploration to fleet-wide coordination by incorporating shared maps and planned paths, target adjustments based on teammate information, and decentralized dispersion. Candidate viewpoints are evaluated using ray-casting information gain G(q) and dual-map A∗ path costs C(q). To minimize redundancy, candidates are penalized based on their proximity to current and stale teammate intent points Pintent(q) and historical trajectories Ptraj(q), while a directional momentum bonus B(q) discourages frequent oscillations. At mission departure, agents preferentially disperse into angular sectors centered at the base station to reduce initial spatial overlap.

Communication Management in Multi-Robot Systems

To maintain communication under budget constraints, the fleet employs event-triggered data relays with dynamic handoffs. When the number of unreported cells not delegated to a teammate reaches threshold C0, or when the remaining exploration allowance is below a preset threshold, an agent initiates an ordinary relay towards the base station if a connected teammate closer to the base station is available. The multi-robot return estimate H is calculated as H = max(Hcomm, Hbase −20) + δ, where Hcomm denotes steps to the first position along the return path where communication with the base station is predicted to be available. When budget reaches critical limits (R ≤ H +M) or no return path can be found, agents switch to final homing mode.

Experimental Results

In preliminary testing across seven public maps, OpenSpace Lab achieved an average coverage rate of 61.04%, ranking 1st overall and outperforming the second-place team by 3.53 percentage points on large-scale maps (49.81% vs. 46.28%). In the final evaluation across ten unreleased private maps, the framework placed 3rd in both tracks, securing 1st place in the single-robot medium-scale map category with a 42.53% coverage rate. The team reached a "61.

Improvements for AI systems

Here are specific improvements that can be made to existing AI systems, based on the methodologies presented in this research:

) Single-Robot Exploration System Improvements:

  1. Improved Global Planning via Learned Map Completion: Instead of relying solely on traditional geometric representations (point clouds), integrate a pre-trained deep learning model (like LaMa mentioned) for neural map completion during the initial global planning phase.

  2. Dynamic Frontier Prioritization using Utility Function: Implement a utility function that explicitly weights multiple factors simultaneously during goal selection, including:

Narrow connectivity/shape of regions, predicted structural information gain, estimated travel cost, and a penalty based on homing burden (time constraint). This moves beyond simple geometric proximity to prioritize high-value exploration.

  1. Adaptive Budget-Aware Heuristics: Develop a set of context-aware heuristics that automatically adjust pathing and goal selection thresholds based on the remaining step budget (e.g., dynamically changing overlap thresholds for short, medium, and long missions).

  2. Proactive Homing Constraint Integration: Modify the real-time frontier evaluation to continuously estimate the remaining time cost required to return to a base station (using formulas like Equation 2) and incorporate this cost directly into the path planning beam search (Equation 1), ensuring that exploration goals are always balanced against the deadline for data retrieval.

) Multi-Robot Exploration System Improvements:

  1. Coordinated Target Selection with Intent Sharing: Implement a utility-driven target selection strategy where agents leverage shared map data and explicit teammate intent points to actively penalize redundant search areas. This allows the system to coordinate exploration paths that maximize collective information gain while minimizing overlap, significantly boosting coordination efficiency compared to independent exploration.

  2. Decentralized Dispersion and Momentum Control: Integrate a directional momentum bonus (B(q)) into the candidate selection process for multi-robot agents to discourage oscillation and promote smooth, efficient movement across the environment. At mission departure, implement a mechanism that forces agents to disperse into angular sectors centered on the base station to avoid initial spatial overlap.

  3. Dynamic Data Relaying and Handoff Protocol: Implement an event-triggered data relay system where agents dynamically assess teammate proximity and communication availability. If a teammate is closer to the base station, the system should automatically trigger a dynamic handoff, optimizing data transfer routes in real-time to maintain connectivity under strict budget constraints.

  4. Predictive Return Path Management: Enhance the return estimate (H) by incorporating both path-dependent costs and uncertainty (e.g., using different offsets for known free space vs. unknown cells), allowing the system to make more informed decisions about when to switch from active exploration to mandatory homing mode, even under tight budgets.

) Overall System Capabilities:

The improved AI systems can achieve the following capabilities:

  1. Maximize information coverage (up to 61% in preliminary tests) within strict temporal and resource constraints.

  2. Achieve superior performance in complex environments by actively managing trade-offs between exploration gain, travel cost, and communication deadlines (homing).

  3. Ensure high coordination efficiency in multi-agent scenarios by reducing search redundancy through shared knowledge and intent awareness.

  4. Adapt robustly to varying mission scales (small, medium, large maps) by dynamically adjusting algorithmic parameters based on the remaining budget.

Sources

Related papers