There is a persistent gap between what vendors promise for unmanned ground vehicles and what those machines actually deliver in the field. Platforms that perform well on pavement or a polished lab floor routinely struggle when faced with the messy realities of off-road terrain: loose soil, stairs, rubble, mud, steep inclines, overhangs and confined passages. The DARPA Subterranean Challenge made that gulf painfully visible when teams deployed dozens of heterogeneous robots into unknown tunnels and caverns and watched many of them fail to traverse or even survive short runs.
Mobility is not a single technical problem. It is a compound of mechanical design, sensing, autonomy and logistics. Choosing wheels, tracks, legs or hybrids forces tradeoffs. Wheels are simple and energy efficient on firm ground but lose out on loose or highly uneven terrain. Tracks improve traction at the cost of weight and mechanical complexity. Legged systems promise obstacle negotiation and gap crossing but introduce huge control, sensing and power burdens. Hybrid wheeled-legged designs try to pick the best of both worlds but add complexity that often reduces reliability in contested or dirty environments. Recent engineering reviews and field reports show this trade space is real, measurable and often underestimated in procurement decisions.
Perception and traversability estimation are the next weak links. Sensors that work well in clear daylight fail in dust, smoke, dripping water or narrow shafts. LiDAR returns degrade on reflective or particulate-filled air and cameras are easy to blind with low light or strong backlighting. Building robust, real-time traversability maps from noisy, partial data remains an unsolved systems problem for many teams, and the Subterranean efforts leaned heavily on sensor redundancy and heuristic recovery behaviours precisely because single-sensor solutions could not be trusted. Robust autonomy needs not only better models but explicit fallbacks and recovery actions when the perception stack is uncertain.
Energy and endurance are often treated as afterthoughts until a mission fails because batteries ran out. Legged locomotion in particular pays a steep energy tax. Even ambitious research prototypes show that wheeled locomotion is generally far more energy efficient on flat surfaces, and hybrid solutions aim to exploit wheels for transit and legs for obstacle negotiation. But adding actuators and mechanical subsystems to support hybrid locomotion increases mass and points of failure, which in turn shortens mission time and complicates maintenance in the field. Real deployments require honest accounting of cost of transport, recharge strategies, and the logistics tail needed to keep robots operational.
Communication and operator dependence are additional operational constraints. Many so-called autonomous UGVs are effectively tele-operated under the hood, with an operator intervening when the autonomy cannot resolve a local decision. That mode amplifies bandwidth requirements and exposes the system to degraded links in cluttered terrain or used electromagnetic environments. Practical experiments and program roadmaps show that current systems must be designed to degrade gracefully to tele-operation but also to continue useful behaviour when communications fail. That design requirement changes everything from interface choices to how much local compute the robot carries.
Field failures also spotlight maintainability, sealing and robustness. Dirt, water ingress and shock from falls are leading causes of mission-ending faults. Military environments demand dust- and water-resistant systems, field-replaceable modules and predictable failure modes. The more complex a mobility solution, the more spare parts, tools and expertise it requires to keep running. In practice this drives many field users back to simpler, proven platforms even as research labs chase legged miracles.
There are positive takeaways. The Subterranean Challenge produced practical techniques that matter: heterogeneous teams, sensor heterogeneity, multi-fidelity traversability estimates, and explicit recovery behaviours when things go wrong. Those are not glamorous but they are the difference between a robot that collects data and a robot that dies in a fissure. Teams that embraced conservative engineering and layered redundancy tended to score better than those that relied on single flashy sensors or unproven locomotion modes.
What should program managers and engineers stop doing and start doing? Stop equating impressive demonstrations on groomed ranges with operational capability. Start funding honest integration tests in representative environments, including extended missions that stress energy, communications and maintenance. Stop assuming a single locomotion modality will solve every problem. Start adopting modular bodies and plug-and-play mobility modules that let a single control stack manage wheeled, tracked or legged gaits depending on terrain. Finally, stop hiding tele-operation behind the word autonomous. Be explicit about operator load, bandwidth requirements and the human procedures that will be used when autonomy fails.
The bottom line is practical and unglamorous. Mobility remains the bottleneck for most fieldable UGV concepts. Progress is steady but incremental. Fixing mobility will require hard systems engineering, not more slogan-driven demos. If you want usable ground robots, invest in sensor redundancy, energy-aware locomotion choices, recovery behaviours and real environmental testing before you sign a procurement contract or trust a machine to do a soldier’s job in a complex battlefield.