KAIST researchers unveil Humanoid v0.7, capable of running at 12 km/h and performing complex balance tasks like the
What Humanoid v0.7 Is Claiming
KAIST researchers have unveiled Humanoid v0.7 with two capabilities that matter in the same sentence: sustained running at 12 km/h and complex balance work under disruption. For bipedal machines, those goals pull in opposite directions. Speed rewards long strides, aggressive push-off, and short ground contact. Balance under disturbance rewards quick recovery steps, wide support margins, and the willingness to abort a planned motion when the center of mass drifts. A platform that can do both has to treat locomotion as a continuous control problem rather than a fixed gait cycle.
Running at 12 km/h is not a parlor trick for demos that last a few meters on a clean floor. It implies the machine can keep a flight phase, absorb impact on heel or midfoot contact, and re-plan foot placement before the next touchdown. Complex balance tasks—recovering from a shove, holding posture on uneven ground, or switching from walk to run without a hard stop—require the same stack to stay stable when the model of the world is briefly wrong.
Why Speed and Balance Fight Each Other
Humanoid control sits on a chain: perception or state estimation, a planner that chooses footsteps and torso targets, a whole-body controller that maps those targets to joint torques, and actuators that can deliver force fast enough to matter. At higher speeds, estimation lag and model mismatch grow. Contact forces spike. Joint torque limits and thermal limits become binding. If the controller is tuned only for graceful walking, it will clip or fall when the robot tries to run. If it is tuned only for sprinting, small slips or delayed foot placement will cascade into a fall.
Useful systems make those tradeoffs explicit. They keep a stable walking mode as a fallback, use contact-aware planning so the next step can be shortened or widened mid-cycle, and reserve actuator headroom for recovery rather than spending every newton on peak speed. Agility is less about a single peak velocity number and more about how quickly the machine can change direction, reject disturbance, and return to a planned trajectory without freezing or thrashing.
- State estimation: joint sensors, IMUs, and contact detection must agree under impact and slip.
- Planning horizon: short enough to react, long enough to keep the next few steps feasible.
- Actuation: high peak torque and bandwidth matter as much as continuous power for recovery steps.
- Fail-safe behavior: when balance is lost, a controlled crouch or multi-step catch beats a rigid stance.
What Practitioners Should Take From the Demo
When you evaluate a humanoid run like Humanoid v0.7, separate the headline speed from the conditions. Ask whether the run is open-loop on a track, closed-loop with vision, or teleoperated assist. Note surface type, payload, and whether balance tasks are separate set pieces or interleaved with locomotion. A robot that runs 12 km/h on a lab floor and then, in another trial, recovers from pushes while standing, is interesting. A robot that can interrupt a run, rebalance, and continue is the harder product-relevant bar.
For teams building or integrating bipeds, the practical checklist is concrete: instrument contact and fall events; stress-test estimation under vibration; keep thermal and torque budgets visible in the control loop; and design operators so a failed step triggers a safe mode instead of a full-body freeze. Records in speed and agility matter when they force better sensing, better contact models, and controllers that treat balance as a continuous duty—not only when the robot is standing still for a photo.