Can Quadruped Robots Actually Navigate the Real World? Two New Papers Suggest the Hardware Matters More Than the Algorithm

That last finding deserves unpacking. The conventional assumption in visual-inertial odometry is that adding an IMU helps. It does, under normal conditions. But quadrupeds introduce what the paper calls "embodiment-induced sensory challenges": foot-impact shocks when the leg strikes the ground, high-frequency mechanical vibrations propagating through the chassis, and rapid angular rotations during gait transitions. These are not edge cases. They are the baseline operating condition of a walking robot. Under these conditions, a low-quality IMU can actively mislead the estimator, producing worse results than vision alone.

To be precise, the paper does not claim that IMUs are bad for quadruped SLAM. It claims that the quality tier of the inertial sensor matters substantially, and that blindly fusing standard IMU data with visual features under harsh legged dynamics can introduce more noise than signal. That is a meaningfully different claim, and it has direct implications for system design.

The computational resource analysis is also useful. The paper quantifies trade-offs across localization accuracy, algorithmic robustness, and compute utilization, which gives practitioners actual numbers to work with when specifying hardware for a deployment. I would want to see the full breakdown of those numbers in the final published version, since the abstract does not include specific figures, but the framing is exactly right.

This is genuinely useful work. I would not call it entirely new science, since the broader point that sensor quality affects SLAM performance is well established in the wheeled and aerial literature, but the systematic, embodiment-specific treatment for quadrupeds is incremental in the best sense: it fills a real gap with rigorous methodology rather than just asserting a result.

The failure mode nobody talks about

The second paper, arXiv:2606.19675, takes a different approach entirely. "ForEnt: A Multi-Modal Dataset for Characterizing Quadruped Robot Entrapments in Forest Environments" introduces a dataset collected with the Unitree Go2, a substantially lower-cost platform than the ANYmal D, across eight forest sites in the Southampton Common Woodlands in the UK.

The phenomenon the paper targets is entrapment: situations where a robot's legs become ensnared in vines, root networks, or other low-lying vegetation, causing loss of stability and, often, toppling. Over approximately 1.7 kilometres of traversals across 11 sequences, the researchers recorded 69 entrapment events. The dataset includes time-synchronized RGB-D images, LiDAR scans, proprioceptive data, and third-person video, with labeled sensor streams for each event.

It is worth noting that this is a dataset paper, not an algorithm paper. The contribution is the resource itself, not a solution to the problem. That is a legitimate and important type of contribution. The field cannot develop entrapment detection methods without data, and to my knowledge (I should caveat that I only found these two sources directly addressing forest-specific quadruped failure modes in recent literature) there was no dedicated dataset for this specific failure mode before ForEnt.

The choice of the Unitree Go2 is interesting. It is a consumer-grade platform, considerably cheaper than research-grade quadrupeds. This matters because it lowers the barrier to reproducing the work and because it signals that the researchers are thinking about deployments where cost constraints are real. Ecological surveying in forest environments is not a use case with unlimited hardware budgets. If entrapment detection can be made to work on a Go2, it is more likely to actually get deployed.

The eight forest sites in Southampton Common Woodlands represent a specific geographic and ecological context, and it remains unclear how well the entrapment patterns captured there generalize to denser tropical forests, to conifer stands with different ground cover, or to environments with seasonal variation in vegetation density. The paper is explicit that this is a starting point, not a comprehensive survey, which is the right framing.

So what

Taken together, these two papers are pointing at the same underlying problem from different directions. Legged robot perception is not just a software challenge. It is a hardware-software co-design challenge, and the hardware constraints are more constraining than the recent wave of impressive locomotion and manipulation results might suggest.

The SLAM paper shows that sensor selection at the hardware level has substantial downstream effects on localization reliability, and that the dynamics of legged locomotion specifically (not just environmental complexity) are a source of degradation. The ForEnt paper shows that even when a robot is navigating successfully in terms of gross locomotion, it can encounter failure modes that are essentially invisible to standard perception pipelines because they involve physical entanglement rather than terrain geometry.

Actually, the research shows something slightly uncomfortable for the field's current trajectory: a lot of recent work on quadruped autonomy focuses on the algorithm layer, on better planners, better learned policies, better scene representations, while the hardware and data infrastructure layers are catching up more slowly. The SLAM paper's recommendation to use global shutter cameras and higher-grade IMUs is sensible, but it also means higher cost and weight, which cascades into power budgets and platform design. The ForEnt dataset enables entrapment detection research, but 69 events across 1.7 kilometres is a small sample on which to train robust detectors. These are not criticisms of the papers. They are honest characterisations of where the field is.

I know I am being picky here, but the framing of legged robots as nearly ready for autonomous ecological deployment, which appears in grant proposals and press releases with some regularity, runs ahead of what the evidence supports. These two papers are useful partly because they are honest about the gap.

What I would want to see next

For the SLAM evaluation work, the obvious next step is replication across platforms. The GrandTour dataset on ANYmal D is a single robot with a specific gait, mass distribution, and vibration profile. Whether the stereo-plus-global-shutter recommendation holds equally for lighter platforms like the Go2, or for heavier platforms used in industrial inspection, is not yet established. The paper's design guidelines are a good starting point, but they should be treated as hypotheses to test on other systems, not universal prescriptions.

For ForEnt, the immediate need is more data. Sixty-nine entrapment events across eight sites is enough to establish the dataset and demonstrate the problem's structure, but it is not enough to train a detector that will generalise. The methodology of the collection, time-synchronized multimodal streams with labeled events, is well designed. Expanding it to more sites, more vegetation types, and more seasons would substantially increase its value. It is also worth asking whether the proprioceptive signal alone (joint torques, contact forces) contains enough information to detect entrapment onset early, before toppling occurs. That seems like the most practically useful detection target, and the dataset is structured to support that analysis.

More broadly, the field would benefit from a benchmark that integrates both failure modes together: a robot that needs to localize accurately under its own locomotion dynamics while also detecting physical entanglement from vegetation. Those two problems interact. A robot that loses localization during an entrapment event, or that triggers high-frequency vibrations while struggling against vines, will stress both perception systems simultaneously. That combined evaluation does not appear to exist yet.

The two papers described here are preprints. Neither has gone through peer review as of the time of writing, and the methodology details in the full papers may reveal additional limitations not visible from the abstracts. That caveat applies with particular force to the quantitative claims in the SLAM evaluation, where the specific numbers matter a great deal for anyone trying to use these results to specify hardware.

Mark Kowalski · Yesterday · 6 min