A new indoor-navigation system for robots reported higher success rates and more efficient routes than a closely matched baseline in both computer simulations and trials with a physical humanoid robot. The system, called EgoNav, links learned visual waypoints, which are intermediate destinations inferred from visual guidance, with local control that accounts for the geometry around the robot.
The manuscript is a preprint version identified in its front matter as accepted in August 2026. Its evidence comes from robot-navigation evaluations in Habitat-sim with Matterport3D scenes and on a physical humanoid platform in office environments.
The gap appeared at short and long distances
In Habitat-sim, EgoNav recorded a success rate of 96.7% and a path-efficiency score of 89.1 at distances below 10 metres. The matched PlaceNav+Falco baseline recorded 80.0% and 74.3%, respectively. At 10 to 20 metres, the figures were 76.7% and 64.9% for EgoNav, compared with 56.7% and 50.2% for the baseline. Beyond 20 metres, EgoNav scored 60.0% and 50.8%, versus 46.7% and 41.0%.
The same pattern appeared on the physical humanoid robot in office scenes. At distances below 10 metres, EgoNav's success rate and path-efficiency score were 73.3% and 65.6%, compared with 60.0% and 57.2% for PlaceNav+Falco. At 10 to 20 metres, the scores were 66.7% and 58.1%, versus 50.0% and 44.8%. Beyond 20 metres, EgoNav recorded 46.7% and 34.0%, compared with 33.3% and 28.2%.
Success rate was the share of runs in which the robot reached within 1 metre of its target without colliding. The path-efficiency measure, known as SPL, compares the shortest route with the route actually taken among successful runs, so a higher score indicates a more direct trip.
A layered system connects vision to movement
EgoNav was presented as a systems contribution rather than as a single new learning or planning algorithm. It couples learned waypoint prediction, geometry-informed refinement and adaptive local planning. Candidate waypoints are scored for safety, directional coherence and fidelity to the learned prior, while refinement outcomes modulate the local planner.
PlaceNav+Falco was used as a sensor-matched control. It used the same RGB-D inputs, meaning colour images paired with depth data, and the same Falco planner as EgoNav, but without waypoint refinement, adaptive planner modulation or relocalization.
The tests covered 12 MP3D simulation scenes and seven real-world office scenes spread across two floors. Simulation used random start and goal sampling, with 30 episodes per method at each distance level. Refinement and planning parameters were fixed across settings. Candidate values were swept on five separate held-out simulation scenes, with medium-distance success used to select the empirical settings.
The component comparisons showed a clear sequence of scores
A progressive ablation compared configurations one at a time in a 10 to 20 metre simulation task with 30 runs. The sequence reported success rate and SPL of 36.7% and 34.5% with GNM alone; 46.7% and 44.0% with visual subgoal retrieval; 56.7% and 50.2% with a local path planner; and 70.0% and 61.2% with waypoint refinement. With adaptive planner modulation, the reported success rate was 70.0% and SPL was 62.6%. The full configuration recorded 76.7% success and 64.9% SPL.
A separate criterion comparison used the same distance range and number of runs. Full refinement scored 76.7% for success and 62.4% for SPL. The configuration without directional coherence scored 56.7% and 48.0%, the largest observed success-rate reduction relative to full refinement in that comparison. Without the learned prior, the scores were 63.3% and 54.5%; without safety, they were 76.7% and 59.8%; and without refinement altogether, they were 53.3% and 47.9%.
Moving obstacles were tested separately
In dynamic-obstacle tests, EgoNav recorded higher scores than PlaceNav+Falco in both settings. In simulation, across 30 episodes, EgoNav recorded 63.3% success and 50.4% SPL, compared with 40.0% and 32.8% for the baseline. In the physical tests, 10 newly collected episodes across five scenes produced scores of 50.0% and 36.5% for EgoNav, versus 30.0% and 20.8%.
The moving-obstacle results are limited to the reported simulation and office scenarios. They do not establish reliable avoidance of fast-moving obstacles or dense dynamic traffic.
The system was timed for reactive control
On the humanoid platform, the reported average per-cycle latency was 100 milliseconds for preprocessing, 20 milliseconds for waypoint prediction, 40 milliseconds for refinement and 2.5 milliseconds for planning. Total latency was approximately 160 milliseconds, supporting the authors' stated 5-hertz reactive control loop.
The results remain tied to the tested settings
The authors frame EgoNav as a systems contribution centred on coupling learned guidance with geometry-aware local control. These results support descriptive comparisons within the tested settings, not broad claims about transfer beyond the listed simulation and office scenes.
The sequential tests show how the reported scores differed as components or refinement criteria were added or omitted, but they do not by themselves establish that the same pattern will hold in other environments. No funding source is reported in the document.
Paper data and sources
Original title: EgoNav: Bridging Learned Waypoints and Geometry-Aware Local Control for Robust Indoor Navigation
Authors: Jing Wang, Shiqi Zhao, Hairong Qu, Peng Yin
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text