The hard part of legged robots is no longer walking

The next hard problem for legged robots is choosing the right gait. For a decade the bottleneck slid sideways: from balance, to climbing, to push recovery. Each collapsed when a new architecture arrived, and the next question took its place. The next open question appears to be selection — a bounded interpretation of HOUND's results suggesting the bottleneck has shifted from locomotion to selection, though the broader field-wide trend is inferred from one system's mechanism delta, not established by comparative study.

KAIST's HOUND quadruped offers a concrete answer. The 100-pound (45 kg) robot crossed roughly 0.7 miles of campus and 0.2 miles of forest, switching on its own between a steady trot and a faster bound. Most readers will read that as a better robot. The more accurate read is a better selector: one APT-RL policy that picks which of two pre-learned skills to use, replacing the hand-tuned per-gait controllers that stumbled on transitions.

The mechanism is repeatable. Generate many short movement examples in simulation, train one policy to switch among them, ship the result. The pretraining speed — 180,000 sequences in roughly eight minutes — suggests that labs with cheap skill synthesis may have a structural advantage over labs relying on hand-tuned controllers, for this system.

The gain goes to teams building generalist controllers over skill libraries. The loss goes to the assumption that one end-to-end policy will solve locomotion unaided — at least for this class of multi-gait systems, selection just became the gate.