Three recent arXiv preprints start from the same observation: robots that work only on flat, structured ground are too limited for real-world settings. One paper targets roofing, where humanoid robots must coordinate balance, locomotion, and work-related motions on a pitched surface. Another argues that general-purpose motion trackers trained only on flat ground fail to exploit bipedal mobility across complex terrain. A third points to disaster sites, planetary surfaces, and agricultural fields as environments where the right robot often doesn't yet exist.
Each paper proposes a different fix. The roofing system adapts whole-body control to slopes and notes that directly retargeted human demonstrations may not be enough. GigaBrain-WBC-0.5 adds a behavior world model so a general tracker can maintain balance while interacting with the environment and crossing terrain. GLAMDRING takes a more radical route: it couples gait learning with morphology co-design through reinforcement learning of central pattern generators, effectively asking which robot body should exist in the first place.
The papers agree on the problem but differ on the solution: one is task-specific, one improves a general tracking policy, and one rethinks the robot's body itself. Together they point to a shift away from flat-ground benchmarks and toward deployment in uneven, unstructured environments.