TL;DR
ERA (Error-Resistant Autonomous navigation) enables robots to find objects in unseen environments using only a single RGB camera, bypassing the need for expensive depth sensors. By adopting a "Move with Error" paradigm, the system treats its own AI-predicted depth maps as fallible, using movement to actively correct geometric errors and "ghost" obstacles that would otherwise stall traditional navigation algorithms.
What Problem Does This Paper Address?
Most modern autonomous navigation systems are "map-based": they use depth sensors (like LiDAR or RGB-D cameras) to build a geometric map of their surroundings. This map tells the robot where it can go and where "frontiers"—the boundaries between explored and unexplored space—lie.
Recently, "Depth-from-RGB" models have become powerful enough to predict 3D geometry from simple images. In theory, we should be able to swap a physical depth sensor for one of these AI models. However, in practice, this fails spectacularly. Predicted depth is rarely perfect; errors are view-dependent and inconsistent. When a robot maps these imperfect predictions as if they were ground truth, errors accumulate. A slightly overestimated wall becomes a thick "ghost" obstacle that blocks a real hallway.
As shown in the paper's analysis, these errors often surround the robot, eliminating all "frontiers." Traditional agents, seeing no way forward and no new places to explore, simply give up and stop prematurely.
Methods
The authors propose ERA, a training-free system built on the Move with Error paradigm. Instead of assuming the map is correct, ERA assumes the map is a work in progress that requires constant validation through movement. It consists of three core components:
- Revisable Occupancy (Move2Error): Unlike standard mappers that "lock in" obstacle detections, ERA uses a bounded evidence approach. If a new viewpoint shows that a previously "blocked" area is actually free space, the system clears that obstacle. This allows the robot to "see through" its previous mistakes.
- Exploration from Reachable Viewpoints (Error2Move): Traditional exploration targets "frontiers." If errors hide all frontiers, the robot stalls. ERA instead selects exploration goals from any reachable known-free space. This encourages the robot to keep moving and looking around, which provides the new perspectives needed to trigger the "Move2Error" correction.
- Target Verification: To prevent false positives (stopping because a blob of pixels looks like a sofa from one bad angle), ERA requires the robot to verify the target from multiple viewpoints and distances before calling "STOP."
The system is "zero-shot" and "training-free," meaning it uses off-the-shelf foundation models (like Depth Anything 3 and SigLIP) without any environment-specific fine-tuning.
Datasets
The researchers evaluated ERA on three major indoor navigation benchmarks within the Habitat simulator:
- HM3D v1 & v2 (Habitat Matterport 3D): Large-scale datasets of high-fidelity 3D scans of real-world residential buildings.
- MP3D (Matterport3D): A diverse dataset of 90 housing environments with 21 object categories.
They also validated the system on a physical AgiBot Genie G1 wheeled humanoid robot in a real university building.
Key Insights and Ideas
- The Circular Dependency of Errors: To fix a map error, you need a new viewpoint. To get to a new viewpoint, you need a map that says the path is clear. If the error is the obstacle blocking the path, the robot is stuck. ERA breaks this loop by allowing the robot to move within "safe" areas to peek around corners and re-evaluate blocked paths.
- Frontiers are Fragile: In RGB-only navigation, the "frontier" is a high-variance signal. Relying solely on it for exploration is a recipe for failure.
- Asymmetric Updates: ERA's occupancy grid updates are asymmetric—it's quick to mark an obstacle but requires consistent "free space" evidence to remove one. This balances safety with the need for error correction.
Key Results
ERA sets a new state-of-the-art for training-free, RGB-only object navigation:
- Massive Recovery of Performance: On the HM3D v2 benchmark, simply replacing the depth sensor in a baseline model (VLFM) dropped the success rate from 68.2% to 10.8%. Adding ERA’s components brought that success rate back up to 61.7%.
- Benchmark Superiority: ERA achieved a 70.1% success rate on HM3D v2, outperforming the previous best RGB-only method (MVP-Nav) by over 10 points.
- Real-World Robustness: In real-world robot trials, ERA successfully located objects like chairs and trash cans in 11 out of 12 runs, even when the target was not initially visible.
Limitations and Future Work
- Pose Estimation: ERA relies on accurate pose estimates (odometry). While the authors showed it can handle some drift, significant odometry failure would still degrade the map.
- Efficiency: The "Exploration from Reachable Viewpoints" strategy is robust but can sometimes lead to longer paths compared to a perfect-depth frontier search, as the robot might backtrack to re-observe an area.
- Future Directions: The authors suggest integrating more sophisticated pose estimation and exploring more efficient ways to select "corrective" viewpoints to further shorten navigation paths.
For more details, you can find the full paper under the title "ERA: Error-Resistant Autonomous Object Navigation with Predicted Depth."
Summary updated