Humanoid robots have recently demonstrated promising capabilities in real-world ball sports. However, achieving professional motion styles while maintaining strong task performance remains challenging. In this work, we propose AdaPT, an Adaptive Motion Planning and Tracking framework that learns professional tennis serving and rally styles directly from broadcast videos. This hierarchical design is motivated by the key insight that the planner generates stylistic kinematic motions, while the tracker executes them with minimal interference with planning. Despite its effectiveness in simulation, a substantial sim-to-real gap emerges: tracking performance inevitably degrades on real robots, and this degradation is partially overlooked by autoregressive planning and further compounded by noisy perception. To address these issues, our adaptation mechanism improves tracking robustness by learning to track randomized execution speeds, while conditioning the planner on a learned motion-speed adapter to mitigate compounding errors. Real-world experiments on the Unitree G1 demonstrate the effectiveness of our adaptation mechanism in bridging the sim-to-real gap. We further deploy AdaPT policies on the full-size Dobot Atom humanoid robot (1.7m) and demonstrate in-the-wild serving without motion capture. Beyond these results, our real-world experiments reveal both algorithmic and engineering insights for future humanoid ball-sports systems. Videos and code are available on our \href{https://humanoidtennis.github.io/AdaPT/}{project website}.
Tao Huang, Ruofei Liu, Xuchen Tang et al.· 0 citations
Upper-limb rehabilitation exoskeleton systems are characterized by strong coupling, high nonlinearity, parametric uncertainties, and unknown disturbances. Furthermore, conventional prescribed-performance methods usually impose fixed and strict error constraints during the convergence process, which may limit the flexibility of transient response. To address these issues, the core innovation of this paper lies in the introduction of an adaptive-boundary prescribed performance mechanism, which enables the constraint boundaries to be dynamically adjusted online according to tracking errors, thereby simultaneously improving both transient flexibility and steady-state convergence accuracy. Specifically, a model-free system representation is first established by combining an ultra-local model with time-delay estimation. Subsequently, a gain-adaptive super-twisting sliding mode observer is developed to estimate and compensate for time-delay estimation errors and lumped uncertainties in real time. On this basis, by introducing a fixed-time nonsingular terminal sliding mode surface and a novel hyperbolic-cosine barrier Lyapunov function, a prescribed-performance fixed-time sliding mode controller is constructed to ensure that the system states achieve fixed-time convergence while strictly satisfying the prescribed performance constraints. Finally, numerical simulations comparing different methods demonstrate that the proposed approach exhibits superior comprehensive performance in tracking accuracy, convergence speed, and robustness. Subsequent visual simulations further verify the effectiveness and practical application potential of the proposed method. Finally, experiments are implemented in the wear-able exoskeleton experimental platform, experiment results demonstrate the effectiveness of the proposed scheme. Note to Practitioners—This work is motivated by the need for safer and more flexible assistance in upper-limb rehabilitation exoskeletons. In clinical training, patients may show different movement abilities, muscle stiffness, fatigue levels, or involuntary motions. Therefore, a fixed tracking boundary may be too strict for some patients at the beginning of training, while a loose boundary may reduce rehabilitation accuracy. The proposed method allows the error boundary to change online according to the tracking error, so that the exoskeleton can tolerate larger transient deviations during difficult movements and gradually provide stricter tracking assistance as the motion becomes stable. For practical use, the initial boundary should be selected according to the patient’s initial motion error and safety range, and the steady-state boundary should be chosen according to the required rehabilitation accuracy. The controller does not require an accurate dynamic model of the exoskeleton, which may reduce the modeling burden for engineers. However, before clinical application, further extensive hardware tests and multi-subject evaluations should be conducted.
Jianjun Sun, Ruofei Liu, Xue Li et al.· IEEE Transactions on Automat...· 0 citations
StableMimic is presented, a unified tracker trained beyond the nominal tracking distribution that achieves the lowest errors on all four tracking metrics among five methods and attains the lowest values on six of seven post-fall motion and load measures, supporting improved interaction safety under this protocol.
Weihao Wu, Mingzhe Huang, Ruofei Liu et al.· 0 citations