Walking on Roofs: Exploring the Potential of Walking Robots for Construction Work on Roofs

arXiv:2610.12272 · cs.RO · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Walking on Roofs".

Dev: The gist The first systematic experimental evaluation of quadruped locomotion in tiled roof environments establishes a baseline for future research.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So, this paper is essentially taking a commercial robot called the Unitree Go2 and putting it on real roof tiles to see if it can actually do construction work safely without failing, and honestly, the conclusion is pretty clear that right now, it's not ready for that kind of job.

Dev: Yeah, they found that standard ball feet just aren't cutting it; the slippage increases quadratically as the slope gets steeper because those feet are too simple and don’t handle the load distribution well.

Taro: And the failure points weren't just about the overall angle of twenty-five degrees or something; it was about these tiny, abrupt features like tile steps or curves that trigger a slip even on a flat surface <ref:two thousand six hundred ten point one two two seven two#pg10.

Rosa: They also showed that without any changes to how the robot controls its balance, it just doesn't work reliably when things get tricky on those inclines.

Dev: The numbers are telling because they measured theoretical friction coefficients suggesting you *could* climb angles over forty-eight degrees, but the actual robot failed at much lower slopes than that <ref:2610.12272#pg6>.

Taro: What this means for us is that we can’t just rely on measuring static friction coefficients to predict how a robot will actually move dynamically across a surface <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: Exactly, because the way the center of mass shifts during movement creates those transient load changes that slip happens because of, even when the theoretical friction is high.

Dev: That leads directly to their suggestion that we need active and incline-aware center of mass management strategies to keep those forces within the limits of what the feet can handle <ref:two thousand six hundred ten point one two two seven two#pg10.

Taro: And they’re not stopping there; they argue that for real-world deployment, you need perception capabilities so the AI can proactively avoid those known problematic spots, like stepping between tiles <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: So the big takeaway here is that making a quadruped reliable for roof work isn't just about making stronger feet or tweaking basic control policies; it requires a whole new approach to how the robot senses and plans its movement in that specific terrain <ref:two thousand six hundred ten point one two two seven two#pg3.

Dev: It points toward needing hardware changes, like feet with more adaptable contact areas, combined with smart software that learns to manage its weight distribution while seeing what's around it <ref:two thousand six hundred ten point one two two seven two#pg11.

Taro: And I think the paper suggests that reinforcement learning is going to be key here; you need to teach the robot how to actively adjust its COM based on the visual input so it can adapt its gait rather than just following a pre-set, standard policy <ref:two thousand six hundred ten point one two two seven two#pg11.

Rosa: So we’ve established the baseline, and now we have a roadmap for what future research needs to focus on if we want these robots actually doing construction work up there safely <ref:two thousand six hundred ten point one two two seven two#pg3.

Dev: It’s clear that the current system is limited by not being incline-aware in its control loop, which is a major failure mode when you're dealing with changing slopes <ref:two thousand six hundred ten point one two two seven two#pg10.

Taro: We need to move beyond just testing the robot on flat surfaces and start teaching it how to handle those dynamic shifts and specific terrain features that cause trouble <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: That’s what we’re looking at now, figuring out how to integrate these hardware improvements with those advanced AI strategies before you can actually send a robot up there for a real job.

The paper's summary: Taro: So, we’re looking at what the authors suggest to fix these issues—the paper says we need several major changes to make this work reliably outside of a controlled lab setting.

Rosa: Right, they’re saying dynamic locomotion performance can't be predicted just by measuring static friction coefficients; you have to model how the center of mass shifts during movement and how that affects normal forces <ref:two thousand six hundred ten point one two two seven two#pg10.

Dev: They strongly push for active and incline-aware center of mass management strategies so the AI can actively shift weight to keep those tangential forces in check, rather than relying on a fixed policy <ref:two thousand six hundred ten point one two two seven two#pg10.

Rosa: It’s about making the control system smarter, teaching it to anticipate what will happen when the slope changes, which is something their current setup totally misses because their controllers aren't incline-aware <ref:two thousand six hundred ten point one two two seven two#pg10.

Dev: And hardware has to play a huge role too; they say we need feet with larger contact areas or even actuated ankle joints so the robot can better handle those repetitive tile structures without losing grip <ref:two thousand six hundred ten point one two two seven two#pg11.

Taro: I agree, and they mention perception is critical; you need vision capabilities so the AI can look ahead and avoid problematic foot positions, like those steps between tiles that cause slippage <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: So it sounds like they’re calling for an integrated approach where hardware adaptations meet smarter software—using things like reinforcement learning to teach the robot how to actively manage its weight distribution based on what it sees <ref:two thousand six hundred ten point one two two seven two#pg11.

Dev: It’s a lot of work, but if you can get that loop rate and latency right while handling those dynamic load shifts, then you might actually get reliable performance on a roof.

Taro: That’s the big picture; we need to combine tailored locomotion strategies with perception and hardware changes so the robot can handle the complexity of real construction environments <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: So, they’re moving past just testing if it *can* walk up a slope, toward designing robots that can safely *operate* in them.

Dev: It moves from just measuring static friction to modeling the whole dynamic process of how the robot interacts with that surface under load <ref:two thousand six hundred ten point one two two seven two#pg10.

Taro: That’s where the real autonomy challenge lies—teaching it to manage those complex physical constraints on its own while navigating a visually rich, uneven world <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: So, they’re laying out a blueprint for what a roof-capable quadruped needs to look like before we can actually start building something that works in the field.

The paper's improvements: Dev: We established that standard ball feet aren't enough and the current control systems lack the necessary incline awareness to handle those load shifts correctly <ref:two thousand six hundred ten point one two two seven two#pg10.

Taro: The authors laid out a clear path forward, emphasizing that we need to combine better hardware design with sophisticated software techniques like reinforcement learning for gait adaptation and perception modules for avoiding obstacles <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: Essentially, they’re telling us that for this application to work, we can’t just rely on tweaking existing algorithms; we have to fundamentally change how the robot interacts with its environment <ref:two thousand six hundred ten point one two two seven two#pg3.

Dev: It changes things because it highlights that dynamic performance isn't something you can predict from simple static friction measurements alone; you need a model for how the load moves <ref:two thousand six hundred ten point one two two seven two#pg10.

Taro: And what this means for autonomy is that the robot needs to be much more proactive in managing its center of mass and choosing stable footholds based on visual cues, not just reacting to slips <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: It’s a pretty sobering look at the current limitations of commercial quadruped robots when you move them out of a controlled lab setting and into something as unpredictable as a tiled roof <ref:two thousand six hundred ten point one two two seven two#pg3.

Dev: We need to keep pushing for that better loop rate and more complex state estimation if we’re going to start getting these kinds of systems reliably operating on real structures <ref:two thousand six hundred ten point one two two seven two#pg3.

Taro: I think the next big challenge is building those models, like GeniWorld, that can generalize robustly across all those unseen scenarios the way this paper suggests <ref:two thousand six hundred ten point one two two seven two#pg3.

Rosa: Exactly, so we’re looking at a lot of heavy lifting here to bridge that gap between a successful lab demonstration and actual field deployment on a job site <ref:two thousand six hundred ten point one two two seven two#pg3.

Dev: I think the next big thing is figuring out how to make those incline-aware COM management strategies work in real-time, not just in simulation <ref:two thousand six hundred ten point one two two seven two#pg10.

Conclusion: Rosa: So we’ve covered that paper, "Walking on Roofs: Exploring the Potential of Walking Robots for Construction Work on Roofs," and it really shows us just how far away we are from having a robot safely doing construction work up there right now.

Dev: We established that standard ball feet aren't enough and the current control systems lack the necessary incline awareness to handle those load shifts correctly.

Dev: It changes things because it highlights that dynamic performance isn't something you can predict from simple static friction measurements alone; you need a model for how the load moves.

Dev: We need to keep pushing for that better loop rate and more complex state estimation if we’re going to start getting these kinds of systems reliably operating on real structures.

Taro: I think the next big challenge is building those models, like GeniWorld, that can generalize robustly across all those unseen scenarios the way this paper suggests.

Rosa: Exactly, so we’re looking at a lot of heavy lifting here to bridge that gap between a successful lab demonstration and actual field deployment on a job site.

Dev: It’s clear that static friction measurements alone don't tell us much about dynamic performance when you introduce those complex shifts in load.

Taro: And the paper shows that even with good theoretical friction, the lack of incline awareness in the control loop is a big failure mode for deployment on real structures.

Rosa: We’ve covered what they tested and what they found regarding the limitations of current designs for roof work, which should give us a really clear picture of why this area needs more tailored solutions.

Dev: So, moving forward, we need to think about how to build those incline-aware COM management strategies and improve the foot hardware itself.

Taro: I think incorporating reinforcement learning for gait adaptation is the next logical step to teach the robot how to actively manage its weight distribution in response to that changing slope.

Rosa: That sounds like a solid plan for future research, moving beyond just testing and starting to build solutions that actually work in this kind of challenging environment.

Dev: We'll keep an eye on how those hardware improvements and software strategies play out in the next round of testing against these specific failure modes.

Taro: That’s what we need to focus on, making sure the robot doesn't just fail spectacularly when it hits a step or a curve.

Rosa: That brings us to the end of our talk on this paper, "Walking on Roofs: Exploring the Potential of Walking Robots for Construction Work on Roofs." Thanks for joining us as we explored these initial findings.

Bjoern-Felix Dettmar, Arne Roennau

Machine Intelligence and Robotics Lab at Karlsruhe Institute of Technology (KIT)

cs.RO

Submitted: 2026-10-08

Updated: 2026-10-08

Code: https://github.com/legion1581/go2

The gist: The gist The first systematic experimental evaluation of quadruped locomotion in tiled roof environments establishes a baseline for future research.

Key concepts

Quadruped Locomotion: This refers to the movement of a four-legged robot. The paper tested how this type of robot moves on different surfaces, specifically focusing on how it balances and walks when the surface is tilted, like a roof.
Friction Measurement: Friction is a force that resists motion between two surfaces. Researchers measured this by pushing the robot's foot onto a roof tile and measuring the opposing force needed to drag it across. This measurement helps determine how much grip the foot has on the surface.
Incline Walking: This is when a robot walks up or down an inclined surface, such as a roof slope. The study specifically tested this skill. Results showed that slippage increases quadratically with the incline angle, meaning walking becomes much harder and riskier on steeper slopes.
Center of Mass (COM) Management: The COM is the center point where all the weight of the robot is considered to be balanced. The paper found that for inclined walking, robots need active control strategies to manage this point dynamically, shifting it appropriately to maintain stability and prevent slippage.

Terminology

Adaptation — 11× in this paper · explained in 1 episode
Reinforcement Learning — 11× in this paper · explained in 4 episodes
Core — 4× in this paper · explained in 2 episodes
Adaptability — 3× in this paper · explained in 1 episode
Accuracy — 2× in this paper · explained in 1 episode
Failure Modes — 2× in this paper · explained in 3 episodes
Karlsruhe Institute of Technology — 2× in this paper
Center of Mass (CoM) — 1× in this paper · explained in 2 episodes
Energy Efficiency — 1× in this paper · explained in 3 episodes
Physical Constraints — 1× in this paper · explained in 7 episodes
Robustness — 1× in this paper · explained in 5 episodes
Scaffolding — 1× in this paper · explained in 3 episodes

Summary

The gist The first systematic experimental evaluation of quadruped locomotion in tiled roof environments establishes a baseline for future research.

Introduction and Motivation

Construction work on roofs represents a particularly hazardous area where tasks are physically demanding and repetitive, motivating the exploration of robotic assistance Within this domain, work on roofs represents a particularly hazardous area where tasks are physically demanding and repetitive, providing a strong motivation to explore robotic assistance <ref:2610.12272#pg3>. Quadruped robots have become an increasingly active research field over the past decade and have consequently attracted growing interest from the industrial sector <ref:2610.12272#pg3>. For commercial applications, both academia and industry have predominantly proposed sectors such as Inspection & Maintenance, Search & Rescue, and Military & Surveillance <ref:2610.12272#pg3>. To enable quadruped deployment at scale, novel fields of applications with well defined tasks and challenges should be investigated <ref:2610.12272#pg3>.

Evaluation Methodology

The evaluation focused on two primary factors for quadruped locomotion performance in roof environments: friction and stability <ref:2610.12272#pg4>. Friction was measured by estimating slippage on a defined incline angle using the OptiTrack system, which provides measurement accuracy below 1.0 mm <ref:2610.12272#pg4>. The feet of a quadruped robot are the primary, and often only, components that directly interact with the environment <ref:2610.12272#pg4>. Standard ball-shaped feet provide a simple contact modeling approach but result in high stress concentrations and an increased risk of slippage on inclined surfaces because rubber friction is area dependent <ref:2610.12272#pg4>.

Experimental Setup and Results

The test rig was developed using the Frankfurther Pfanne as the surface material, which is the most common type of roof tile in Germany <ref:2610.12272#pg4>. Three core skills were examined for the Unitree Go2: balancing on continuously increasing inclines, get-up/lie-down sequences, and incline walking <ref:2610.12272#pg4>. For incline walking, slippage was quantified using Eq. (2) to define SFLF/RLF as the cumulative incremental marker displacement opposing the walking direction X <ref:2610.12272#pg5>. Regression analysis revealed a highly significant linear relationship between incline angle and traveled distance, a highly significant quadratic relationship between incline angle and slippage, but no explainable correlation between incline angle and trajectory deviation <ref:2610.12272#pg9>.

Key Findings and Limitations

The results indicate that without significant modifications, the Go2 is not suited for operation within roof environments with frequent critical failures during incline walking at 25◦ <ref:2610.12272#pg10>. The measured friction coefficients suggest theoretically achievable incline angles above 48◦, yet locomotion failed at substantially lower slopes <ref:2610.12272#pg10>. Failures during balancing were consistently preceded by slippage of the front limbs because the center of mass shifts toward the rear legs as the incline angle increases, reducing normal forces and thus friction forces at the front feet <ref:2610.12272#pg10>. Furthermore, discrete terrain features dominate slip events, as slippage was frequently triggered by abrupt slip-off events at tile steps and curved sections, even on 0◦ incline <ref:2610.12272#pg10>. Standard quadruped control policies are not incline-aware, as the evaluated controller applied similar push-off forces across all inclines <ref:2610.12272#pg10>.

Design Implications for Future Work

The study suggests several design implications for future roofcapable quadrupeds. First, dynamic locomotion performance cannot be predicted from friction measurements alone, highlighting that dynamic locomotion is dominated by transient load redistribution and controller behavior <ref:2610.12272#pg10>. Second, incline locomotion requires active and incline-aware COM management strategies to reduce tangential forces to enforce actuation within the friction cone limits <ref:2610.12272#pg10>. Third, foot geometry is a critical hardware bottleneck; future designs should employ feet with larger and adaptable contact areas or actuated ankle joints <ref:2610.12272#pg11>. Finally, environment-specific locomotion strategies, such as reinforcement learning, are necessary to teach the robot to actively shift its COM and to avoid problematic foot positions like steps between tiles <ref:2610.12272#pg11>.

Conclusion

Roof environments present a highly challenging new environment for quadrupeds for which a combination of hardware and software adaptations will likely be necessary <ref:2610.12272#pg12>. The experiments underscore that, without appropriate adaptations, quadrupeds are not yet capable of safe and reliable operation in roof environments <ref:2610.12272#pg10>. Future work should adopt an integrated approach combining tailored locomotion strategies and perception with hardware adaptations to meet the specific demands <ref:2610.12272#pg11>.

--- Page 1 ---

Bjoern-Felix Dettmar and Arne Roennau Machine Intelligence and Robotics Lab at Karlsruhe Institute of Technology (KIT), Germany dettmar@kit.edu Preprint of paper which appeared in the Proceedings of CLAWAR 2026, 29th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines, Cranfield University, Cranfield, Beford, United Kingdome, 2026 Abstract. This paper investigates the feasibility of deploying quadruped walking robots for the automation of work in roof environments. While quadruped have demonstrated versatility across various domains, their large-scale deployment remains limited, partly due to lack of applicationspecific designs. Roof environments represent a novel and unexplored use case, combining high safety risks for human workers with repetitive, strenuous tasks that could benefit from robotic assistance. A dedicated test rig of a roof’s surface was designed to evaluate the baseline performance of a commercial quadruped, the Unitree Go2, in this new environment. Experiments revealed that standard ball feet are inherently inadequate for locomotion on sloped roofs: slippage increased quadratically with incline angle, get-up and lie-down sequences were only possible on small inclines, and critical failures already occurred regularly on moderate inclines of 25◦. This work provides the first systematic assessment of quadruped locomotion in roof environments, highlighting both the potential and the current limitations of this application and establishes a baseline for future research. Keywords: walking robots, quadruped, roof environment, construction, locomotion, friction, stability, motion capture Fig. 1. Examples of the Unitree Go2 on the test rig while balancing (left), walking (middle) and overview of the test rig setup (right). arXiv:2610.12272v1 [cs.RO] 8 Oct 2026

--- Page 3 ---

A potential new environment for quadruped robots to operate in is presented by roofs, offering a unique set of challenges and tasks. Quadrupeds could be deployed in roof environments for a multitude of applications, which are presented in Table 1. Residential buildings often feature moderate to steep inclines (25◦ to 60◦), smooth roof tile surfaces, and regularly repeating structural elements such as convex and flat areas or tile row steps, making locomotion inherently difficult [1]. Conversely, the repetitive nature of standardized roof tiles provides an opportunity to tailor quadrupeds feet specifically for this environment. This work provides: – The first systematic experimental evaluation of quadruped locomotion in tiled roof environments. – Quantitative failure thresholds for balanced standing, get-up/lie-down motions, and incline walking establishing a baseline for future research. – A motion-capture-based methodology for objectively observing foot slippage during inclined locomotion.

--- Page 4 ---

3 Experimental Setup As a preliminary experiment the dynamic coefficient of friction f of the ball feet was measured by using the internal force-torque sensor of a UR10e. The foot is pushed onto the roof tile surface with a defined normal force of 60 N, which is approximately the average force applied to a single foot of the Go2 during walking. Next, the foot is dragged over the tile for a distance of 100 mm, while the antagonistic friction force is measured. The experiment is repeated 10 times, both on flat and curved sections of the roof tile. The mean coefficient of friction is calculated through the ratio between measured tangential force FT and applied normal force FN depicted in Eq. (1). To objectively assess the performance of the Go2 (software version 1.4.4) and create a baseline for future comparison with custom feet, an adjustable test rig of a roof environment was developed (see Fig. 1). As surface material the Frankfurther Pfanne was chosen, which is the most common type of roof tile in Germany [2]. With the Go2, three core skills were examined: balancing on continuously increasing inclines, get-up/lie-down sequences, and incline walking.

--- Page 5 ---

4 Baseline Results Overall, the results indicate that, without significant modifications, the Go2 is not suited for operation within roof environments. Manufacturer claims of slopeclimbing capabilities of up to 40◦ were not achieved under the test conditions [18]. The coefficient of friction for a rubber ball foot on a roof tile was measured at 1.144±0.1299 on flat and 1.430±0.1298 on curved sections of the roof <ref:2610.12272#pg6>. Using Eq. (1), this leads to respective maximum material specific incline angles ρ of (48.84 ± 3.22)◦ and (55.03 ± 2.44)◦ <ref:2610.12272#pg6>. For standing on an incline, the critical angle was measured at (27.58±0.72)◦ on average <ref:2610.12272#pg10>. Failures were primarily caused by automatic limb repositioning, which appears to be triggered by slippage of individual limbs or atypical joint angles <ref:2610.12272#pg10>. Get-up and lie-down experiments revealed additional limitations. Full task success was achieved only at low inclines: up to 11.0◦ for head-up and head-down orientations, and up to 7.0◦ for the sideways orientation <ref:2610.12272#pg10>.

Improvements for AI systems

Bold header: Incorporation of environment-specific reinforcement learning for gait adaptation. The paper suggests, "The use of environment specific reinforcement learning is a promising first step to tackle some of the aforementioned issues, by teaching the robot to actively shift its COM to redistribute weight more evenly and reduce additional tangential forces to enforce actuation within the friction cone limits. This will allow the AI system to learn gaits that are environment-specific" rather than relying on standard control policies.
Bold header: Development of perception-driven foothold selection modules. The paper notes, Perception capabilities could be included to avoid problematic foot positions, like the steps between tiles, that promote slipping. This will enable the AI system to use visual data to proactively select stable footholds and avoid terrain features that promote slipping.
Bold header: Implementation of incline-aware center-of-mass (COM) management strategies. Since failures were primarily caused by automatic limb repositioning, which appears to be triggered by slippage of individual limbs or atypical joint angles when standing on inclines, the AI system should incorporate incline-aware COM management strategies to prevent the COM from shifting in a way that reduces normal forces and friction.
Bold header: Integration of hardware-aware foot geometry models. To address the bottleneck, future systems should move beyond ball feet by designing feet with larger and adaptable contact areas, high-friction materials, and the ability to mechanically interlock with repetitive terrain structures. This ensures that foot placement strategies are informed by the physical constraints of the chosen hardware.
Bold header: Dynamic friction modeling based on transient load redistribution. The system should move beyond static friction measurements by incorporating models that account for dynamic effects, as dynamic locomotion performance cannot be predicted from friction measurements alone, specifically addressing how COM shifts reduce normal forces and cause slippage even with high theoretical coefficients.