Quang Huy Dao

Aggressive standing-up for full-size humanoid robots on grass, debris, and other uneven surfaces. The method optimizes contact timing and a whole-body trajectory, then tracks it with a learned controller so the robot can get up from supine and prone postures, including martial-arts-style maneuvers. The study compares this approach with trajectory optimization paired with linear and nonlinear model predictive control.

A behavioral foundation model for extreme humanoid control. World-model conditioning and decoupled recovery priors support teleoperation, dynamic skills such as flips, and getting back up, including when the robot is disturbed during recovery.

Ukemi-inspired falling and recovery for a humanoid, so the robot can hit the ground and stand back up while limiting impact. Accepted at Humanoids.

The widely studied H∞ zero-sum game problem guarantees the integration of external disturbance into the optimal control problem. In this article, two model-free Q-learning algorithms based on H∞ tracking control are proposed for perturbed discrete-time systems in the presence of external disturbance. Moreover, modification of the output optimal control problem is also made. For the optimal tracking control problem, the existence of a discount factor is necessary to guarantee the final value of the cost function, and the Ricatti equation is modified. With the aid of the deviation between Q functions at two consecutive times and the original principle of Off/On policy, the consideration of H∞ zero-sum game problem, two On/Off Q-learning algorithms based on H∞ tracking control are proposed. Then, by computing the Q function, the influence of probing noise on the Q function is considered. The analysis of solution equivalence proves that convergence and tracking are guaranteed in the proposed algorithm. Eventually, simulation studies are carried out on F-16 aircraft to assess the validity of the presented control schemes.