OpenSpace Lab Solution to the IROS 2026 Indoor Exploration Competition
This preprint describes the OpenSpace Lab entry for the IROS 2026 Indoor Exploration Competition. It combines neural map-completion predictions, budget-aware goal selection, and homing constraints for single-robot exploration, with a coordinated multi-robot extension using shared maps and intent penalties. The system placed first in the single-robot public track at 61.04% coverage and third in both private tracks (39.53% single, 39.91% multi).
How can autonomous robots maximize indoor map coverage under strict step budgets and a requirement to return to base before the budget expires, in both single-robot and multi-robot settings?
The competition requires robots to explore unknown indoor maps within fixed step budgets and upload maps only after returning to communication range. Existing exploration methods typically treat time or path length as a passive post-completion metric rather than an active constraint, so agents can run out of time far from base or waste budget on redundant routes.
Conventional autonomous exploration systems often evaluate information gain at frontiers and minimize path length, but do not explicitly enforce remaining-time or homing constraints during real-time target selection. The report notes that time and path length are usually treated as evaluation metrics after mission completion rather than as online decision constraints.
For single-robot exploration, the system uses a pretrained LaMa model within MapEx to predict unseen map completions every eight steps, clusters frontiers into up to seven macro-region tasks, and runs beam search (width 36, max length 6) with a 0.92 discount to order visits. It adds budget-specific heuristics for 1,000/1,500/2,000-step budgets and triggers homing when remaining budget is below estimated return cost plus a 10-step margin. For multi-robot exploration, it extends this with shared maps and intents, penalizes candidate viewpoints near teammate intents and historical trajectories, disperses agents into angular sectors at mission start, and uses event-triggered relays and dynamic handoffs to deliver map data.
On seven public maps, the single-robot entry ranked 1st with 61.04% average coverage, ahead of HeRoLab at 60.12% and TAMU-UADY-Robotics at 55.89%. On large-scale public maps, it averaged 49.81% vs 46.28% for the second-place team, a 3.53-point margin. On ten private maps, it placed 3rd in single-robot at 39.53% (winner MobilityAI 40.26%, second Xplore 39.65%) and 3rd in multi-robot at 39.91% (winner HeRoLab 42.64%, second Xplore 41.12%); it won the single-robot medium-scale private map category at 42.53%. Multi-robot settings averaged teams of 2–5 robots with 500/1,000/1,500-step budgets; single-robot budgets were 1,000/1,500/2,000 steps for small/medium/large maps.
The paper does not report real-robot tests or ablations; evaluation is in a competition simulation with a limited set of public and private maps. Code is not yet available and is promised only after the extended manuscript is accepted. The approach relies on tuned heuristics for specific step budgets and map categories, and the private-track margin to the winner is small (0.73 points single, 2.73 points multi), suggesting limited evidence of robustness outside the public tuning set.
Could inform autonomous mobile robots for indoor mapping, warehouse inspection, and search-and-rescue, especially where battery/time budgets and data return are hard constraints. The multi-robot relay and handoff logic may interest vendors of AMR fleets. Real-world adoption would require validation in physical environments, generalizing beyond competition maps, and releasing robust code. Near-term indoor AMR mapping could be 1–3 years; planetary or safety-critical deployment would be 3+ years.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.