ERA: Error-Resistant Autonomous Object Navigation with Predicted Depth

TL;DR ERA finds named objects in unseen buildings using depth predicted from RGB instead of a depth sensor. It lets later views erase false obstacles from the map and keeps exploring reachable free space when no frontier is left, reaching 70.1% success on HM3D v2 (MVP-Nav 59.4%) and 11 of 12 runs on a real wheeled humanoid.
What they do
Zero-shot object navigation systems build their maps from a depth sensor. Swapping in depth predicted from RGB by Depth Anything 3 breaks them: VLFM falls from 68.2% to 10.8% success on HM3D v2. The authors trace this to inconsistent predictions of the same surface piling up into thick false obstacles that close passages and remove every frontier, so the agent stops before new views could correct the map. ERA, a training-free system, is built so that movement corrects the map and the corrected map lets the agent keep moving.
How it works
- Revisable occupancy. Each grid cell holds bounded signed evidence. A new obstacle counts at once, but one or two later free-space observations clear it. Cells the robot has driven through stay free, and failed forward moves mark obstacles that vision cannot clear.
- Exploration from reachable viewpoints. Goals come from all reachable known-free space, not only frontiers, scored by visible unknown space, image-text similarity and coverage against route length, clearance and revisits.
- Target verification. The target surface is re-estimated during the approach; the agent stops only when close and an earlier view from a different position agrees.
- All models (depth, C-RADIOv4 with SigLIP2 and SAM 3 heads) stay frozen.
Results
- Best training-free method without a depth sensor on all three benchmarks: 55.2% success on HM3D v1, 70.1% on HM3D v2 and 46.4% on MP3D.
- Added to VLFM under the same predicted depth, the two map and exploration components raise success from 10.8% to 61.7%; L3MVN rises from 29.2% to 59.1%. The gains hold with VGGT depth.
- Without free-space clearing, success drops to 13.9%; with frontier goals only, to 32.6%.
- On an AgiBot Genie G1 with only its head camera and wheel odometry, ERA stopped within 1 m of the target in 11 of 12 runs. From six shared starts, a VLFM-style base succeeded 0 times and ERA 6.
Limitations
ERA relies on the poses the platform provides; with simulated odometry drift it kept 95% of its success, but it does not estimate pose itself. Exploring from mapped space can lengthen the path. On the robot the models run on a cabled workstation at about 0.7 s per step, and the real tests are 12 corridor runs in one building.
Figures



Abstract
Depth and 3D foundation models can predict geometry from RGB images, allowing zero-shot object navigation without a depth sensor. However, when predicted depth is mapped as if it came from a sensor, view-dependent errors accumulate in the map, remove exploration goals, and end the search before the agent can observe and correct them. We propose ERA, a training-free object navigation system that replaces the depth sensor with depth predicted from RGB images and poses. ERA follows a paradigm we call Move with Error: movement provides new views that correct map errors, and the corrected map enables further movement. It realizes this paradigm with revisable occupancy and with exploration from reachable viewpoints, which continues the search when no frontier remains, and verifies the target across views before stopping. On three benchmarks, ERA achieves state-of-the-art performance among training-free methods without depth sensors. Its map and exploration components also raise VLFM's success from 10.8% to 61.7% under the same predicted depth. We further deploy ERA on a real robot.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
