From Depth to Safe Depth: Uncertainty-Aware Monocular Perception for Mobile Robot Navigation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "From Depth to Safe Depth".
Paper summary: Rosa: Looking at the title, "From Depth to Safe Depth: Uncertainty-Aware Monocular Perception for Mobile Robot Navigation," it really captures what this work is about >
Dev: It's about taking a potentially dangerous, off-the-shelf monocular depth prediction and making it safe enough for real robot navigation by adding an uncertainty layer >
Taro: What this means in practice is that we can use these powerful foundation models without having to rewrite the entire navigation stack from scratch because we just plug this module in >
Rosa: It’s about using the uncertainty term to decide where obstacles are and whether a pixel can even be considered floor, which addresses those issues where shrinking depth along camera rays lifts the floor into phantom obstacles >
Dev: The authors point out that just correcting the bias isn't enough because there are still systematic errors left over, and only that uncertainty term can cover those remaining issues >
Taro: They also mention a limitation: the experiments were done in simulation with one robot and one camera configuration, and they used ground-truth odometry to isolate perception but this hides localization error >
Rosa: And they acknowledge that because the head was trained on simulated data from the same generator family, verifying it on real images is still something that needs to be done >
Dev: So, in short, this approach nearly doubles the success rate from thirty-seven percent up to seventy percent and brings collisions down from sixty-three percent to just three percent >
Conclusion: Rosa: So, this paper is about taking those depth predictions from monocular foundation models and making them safe enough for real robots to drive with on a regular navigation stack.
Dev: Right, they're proposing this new little plug-in module that takes a predicted depth and adds an uncertainty layer to make it safe.
Rosa: What they did is train this small head on top of the frozen depth model to predict both a correction and an uncertainty value for every single pixel.
Taro: So, instead of just getting one number for depth, you get two things—the systematic error and how much we can trust that prediction.
Dev: Exactly. And they use those two things to calculate a 'safe' depth by taking the predicted depth and subtracting some correction based on that uncertainty and a calibration factor.
Rosa: The training part is interesting because they use a loss function that tries to learn what the typical error size is for each pixel, which helps them separate the systematic errors from the random noise.
Taro: But they also used this beta negative log-likelihood thing to stop the uncertainty term from just blowing up on pixels where there's no real signal.
Dev: That’s a smart move because if you let uncertainty get too high on a hard spot, it can mess up the whole safety calculation.
Rosa: And then they use split conformal calibration to figure out that safety factor, which essentially tells the system how much margin to keep around obstacles based on where you are.
Taro: The results show this uncertainty-aware approach drops collisions from sixty-three percent down to just three percent, which is way better than what raw depth could do alone.
Dev: It’s a big jump, and they even showed that the safe variant wins in paired testing against constant margin approaches.
Rosa: They also talked about how this uncertainty term helps solve those tricky problems where shrinking depth along a camera ray lifts the floor into phantom obstacles.
Taro: It basically lets the system decide if something is an obstacle or not, which is crucial for safety when you're dealing with hard-to-see things.
Dev: It sounds like they’ve managed to make this off-the-shelf depth perception much more reliable without needing a whole new sensor setup.
Rosa: The main thing to remember is that this isn't just about fixing errors; it’s about building a way for these massive AI models to actually be deployed safely on the road.
Taro: It’s not perfect yet, though they did mention limitations like training only in simulation and needing real-world verification.
Dev: Definitely. We need to see how this holds up when the world isn't exactly like their simulated training data anymore, or when we run it at a higher speed on actual hardware.
Rosa: So, for those of you listening who are driving or walking, this means the next time you see a robot using monocular depth, there’s a much better chance it won't suddenly misjudge something important.
Saeid Firouzi Daghigh, Saeed Ayat
Payame Noor University
cs.RO, cs.CV
Submitted: 2026-11-10
Updated: 2026-11-10
Code: https://github.com/saeed5959/depth
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 91/100
The gist: Monocular depth foundation models make it tempting to replace depth sensors on mobile robots with a single RGB camera…
Key concepts
- Safe Depth
- A calculated depth value derived from the predicted monocular depth by applying a safety factor based on learned uncertainty. It ensures that the perceived obstacles are conservatively estimated, preventing collisions by accounting for prediction errors.
- Log-Depth Correction (µ)
- A parameter predicted by a small neural network head that estimates the systematic error in the depth prediction at each pixel. This term learns the consistent bias or offset in how much deeper or shallower an object appears compared to ground truth.
- Uncertainty (σ)
- A parameter predicted by the head indicating the typical error size for each pixel's depth estimate. It quantifies how much confidence the model has in its prediction, which is crucial for determining how conservatively to treat that specific depth value.
Terminology
Summary
Monocular depth foundation models make it tempting to replace depth sensors on mobile robots with a single RGB camera, but using them off the shelf is unsafe because standard navigation stacks driven by Depth Anything V2 collide in 63% of episodes due to systematic over-estimation of nearby obstacles
The Proposed Solution
The authors propose a lightweight plug-in module that transforms predicted depth into safe depth This module consists of a 0.58 M-parameter head on top of the frozen depth model that predicts per pixel log-depth correction µ and uncertainty σ The navigation stack then consumes the one-sided bound dˆsafe = d eµ−kσ, where k is determined by conformal calibration on a held-out world
Training and Correction Mechanism
The training objective for the head is defined by the Gaussian negative log-likelihood using the residual r = log dGT−log ˆd, which is LNLL = 1/2 r squared e-2s + s The minimum of this loss causes σ to learn the typical error size of each pixel while µ learns the systematic part of the error To prevent plain NLL from inflating σ on hard pixels, a β-NLL is used by multiplying the loss by sg(σ 2β) with β = 0.5 Pixels are weighted by w = 1/ max(dGT, 0. p. 1 5, where sg stops the gradient. Pixels are weighted
by w = 1/ max(dGT, 0. Pixels are weighted
by w = 1/ max(dGT, 0. 5 m), since nearby obstacles cause
collisions, and pixels with invalid ground truth or depth outside
[0.Where the paper says it
Where the paper says it
Safety Calibration and Integration
The safe depth is calculated as the one-sided lower bound ˆdsafe = dDA eµ−kσ The safety factor k is set by split-conformal calibration, where k is chosen as the 97.7% quantile of normalized scores z = (log ˆd−log dGT)/σ for pixels closer than 3 m on a validation world This uncertainty-aware safety mechanism reduces depth error from AbsRel 0. p. 1 (3)
Instead of fixing k = 2, which assumes a perfectly calibrated
Gaussian, we choose k by split-conformal calibration [17]: on
a held-out validation world we compute the normalised scores
z = (log ˆd−log dGT)/σ of pixels closer than 3 m and set k to
their 97. 7% quantile, so that dGT ≥ ˆdsafe holds for 97. The correction halves the
depth error (AbsRel 0. 134) and reduces near-range
over-estimation from 59.Where the paper says it
Where the paper says it
Performance Evaluation
The results show that the safe variant raises the success rate from 37% to 70% and lowers collisions from 63% to 3%, which is close to the ground-truth depth performance of 2% The safe variant wins significantly in paired outcomes, winning 45 of 50 discordant episodes against raw depth Furthermore, uncertainty-aware safety wins 52 of 56 discordant paired episodes against a constant margin approach
Discussion on Uncertainty Effects
The paper resolves the issue where shrinking depth along camera rays lifts the floor into phantom obstacles by using an uncertainty-aware ground test The uncertainty term decides both where obstacles are and whether a pixel is an obstacle at all, which is crucial for safety The discussion notes that bias correction alone is insufficient because the remaining errors are not systematic, and only the uncertainty term can cover them Additionally, lowering depth along camera rays has side effects like lifting the floor and pulling side walls inwards
Limitations of the Study
The experiments are conducted in simulation with one robot, one camera configuration, and one depth model (ViTS) The ground-truth odometry used isolates perception but hides localization error The ground test assumes a flat floor The head was trained on simulated data from the same generator family, meaning transfer to real images needs verification Monocular inference at about 1 Hz on a 4 GB GPU limits the speed at which the robot can safely drive
Conclusion
Off-the-shelf monocular depth from a foundation model is not a safe drop-in replacement for a depth sensor in mobile robot navigation A small head on the frozen model, trained to predict correction and uncertainty, and a conformally calibrated lower bound turn this depth into safe depth that an unmodified Nav2 stack can use This approach nearly doubles the success rate and brings collisions close to those of ground-truth depth while a constant margin leaves the robot unable to move Future work will validate the module on real robots and combine aleatoric with epistemic uncertainty for unseen environments
References
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Safe depth module integration: A small head on a frozen depth foundation model predicts a log-depth correction mu and uncertainty sigma, outputting d safe depth as dDA e mu minus k sigma for an unchanged Nav2 stack. This allows standard navigation stacks to utilize uncertainty-aware bounds, improving success rates from 37% to 70% in navigation benchmarks.
-
Uncertainty quantification in cost mapping: The system treats pixels as floor if the floor lies within its own uncertainty band, defined by k sigma, which resolves the issue where shrinking depth along rays lifts the floor into phantom obstacles. This uncertainty-aware ground test decides both where obstacles are and whether a pixel is an obstacle at all.
-
Uncertainty-aware calibration for safety margin: The system sets the constant safety margin c using split-conformal calibration on a held-out world, ensuring that
both baselines are therefore equally safe at the pixel level,
whereas a fixed margin freezes the robot. This allows for calibrated caution where errors actually occur rather than applying uniform conservatism.
Abstract
Monocular depth foundation models make it tempting to replace depth sensors on mobile robots with a single RGB camera. We show that, used off the shelf, this is unsafe: in a simulated indoor navigation benchmark, a standard Nav2 stack driven by Depth Anything V2 (metric) collides in 63% of 120 episodes, because the model systematically over-estimates the distance to nearby obstacles. We propose a lightweight plug-in module that turns predicted depth into safe depth. A 0.58,M-parameter head on top of the frozen depth model predicts, per pixel, a log-depth correction μ and an uncertainty σ, trained with a β-weighted Gaussian negative log-likelihood. The navigation stack consumes the one-sided bound d safe=d,e μ-kσ, where k is set by conformal calibration on a held-out world, and uncertainty also decides which pixels may be treated as floor. The correction halves the depth error (AbsRel 0.262 to0.134) and reduces near-range over-estimation from 59.8% to 12.0%. In navigation, safe depth raises the success rate from 37% to 70% and lowers collisions from 63% to 3%, close to the 2% of ground-truth depth. A constant safety margin calibrated to the same pixel coverage reaches only 30% success because it freezes the robot; uncertainty-aware safety wins 52 of 56 discordant paired episodes against it. Gains hold on glass, texture-less walls and dark regions.
Sources
- MonoNav: MAV Navigation via Monocular Depth Estimation and Reconstruction
- NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration
- An Open-Source LiDAR and Monocular Off-Road Autonomous Navigation Stack
- A Critical Synthesis of Uncertainty Quantification and Foundation Models in Monocular Depth Estimation
- CARE: Enhancing Safety of Visual Navigation through Collision Avoidance via Repulsive Estimation
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation