Mingguo Zhao

Humanoid soccer poses a representative challenge for embodied intelligence, requiring robots to coordinate agile locomotion with unreliable visual perception in dynamic environments. However, existing systems typically rely on modular pipelines that separate perception from control or assume ideal sensing, making it difficult to achieve coherent and reactive behavior under real-world perceptual limitations. In this work, we present a unified reinforcement learning–based controller that enables humanoid robots to learn vision-driven reactive soccer skills by directly coupling visual perception with locomotion control. The robot is trained in simulation to acquire soccer behaviors, and adversarial motion priors guide policy learning toward natural motion patterns. To support robust performance under imperfect sensing, we introduce an encoder-decoder architecture together with a virtual perception system that models key characteristics of onboard vision, exposing the policy to perceptual noise and detection failures during training. This design encourages the policy to internalize perceptual uncertainty and continuously adapt its motion in a closed loop. The resulting controller produces coordinated soccer behaviors using only onboard vision, including ball searching, chasing, and multidirectional kicking. It reduces ball position estimation error by 46% and shortens time-to-kick by up to 64% compared with a rule-based baseline, achieving around 90% kicking success in frontfield positions. Experiments across diverse environments and dynamic scenarios, including real RoboCup competitions, further demonstrate the robust performance of the controller. These results highlight the practical effectiveness of integrating perceptual uncertainty directly into policy learning for achieving reliable vision-driven behaviors in humanoid robots operating under real-world conditions.

Dynamic bipedal jumping is highly sensitive to takeoff momentum and landing configuration, making the direct execution of purely offline-optimized trajectories unreliable on full-sized humanoid robots in the presence of modeling inaccuracies and execution uncertainties. A key challenge is the gap between dynamic feasibility predicted offline using simplified models and executability on physical hardware, because such models cannot fully capture full-body dynamics, actuator behavior, contact transitions, and execution uncertainty. This paper proposes an offline-to-online planning and whole-body control framework for robust in-place jumping of full-sized humanoid robots. The framework integrates phase-consistent offline trajectory optimization, lightweight online reference reshaping, constraint-aware whole-body control, and actuator-level command mapping to improve execution robustness without online re-optimization. In the offline stage, centroidal-dynamics-based trajectory optimization generates jumping references subject to kinematic-consistency and contact-feasibility constraints. During execution, these references are adapted online using real-time state estimates to compensate for takeoff deviations and regulate the landing state; a weighted quadratic-programming whole-body controller then tracks the adapted references. Hardware experiments on a 79.5 kg humanoid robot demonstrate repeatable in-place vertical jumps with a height of approximately 30 cm and stable landings. The results show that robust jumping on a full-sized humanoid robot can be achieved by combining offline nominal trajectory generation with online execution adaptation rather than relying on exact reproduction of the offline trajectories.

The integration of AI tasks into Artificial Intelligence of Things (AIoT) devices has gained significant attention for its ability to reduce transmission latency and enhance data privacy compared to traditional cloud-based solutions. However, heterogeneous AIoT platforms face the challenge of balancing stringent latency constraints with energy efficiency under dynamic AI workloads. To address this issue, we propose BOLT-PM, a Bayesian Optimization framework for Latency-sensitive Task and Power Management, which dynamically allocates computing resources and optimizes power consumption in AIoT devices. We design a multi-branch neural network to accurately predict task execution times under varying resource constraints, enabling precise and adaptive performance estimation. We further enhance the Bayesian optimization process by embedding a variational autoencoder (VAE) to construct a smooth latent representation of the search space, thereby accelerating convergence and improving optimization stability. We implement BOLT-PM as a lightweight, platform-agnostic runtime framework that continuously adapts to workload fluctuations, maintaining latency guarantees while minimizing energy consumption. We evaluate BOLT-PM on several commercial AIoT boards, including RK3588, Jetson TX2, and Raspberry Pi, and compare it against classical heuristic methods and state-of-the-art approaches. Experimental results show that BOLT-PM achieves substantial energy savings while ensuring low-latency AI task execution, demonstrating its effectiveness as a robust and energy-efficient solution for power-aware AIoT applications.

Accurate proprioceptive odometry is fundamental for legged robot navigation in GPS-denied and visually degraded environments where conventional visual odometry systems fail. Current approaches face critical limitations: analytical filtering methods suffer from modeling uncertainties and cumulative drift, hybrid learning-filtering approaches remain constrained by their analytical components, while pure learning-based methods struggle with simulation-to-reality transfer and demand extensive real-world data collection. This paper introduces AutoOdom, a novel autoregressive proprioceptive odometry system that overcomes these challenges through an innovative two-stage training paradigm. Stage 1 employs large-scale simulation data to learn complex nonlinear dynamics and rapidly changing contact states inherent in legged locomotion, while Stage 2 introduces an autoregressive enhancement mechanism using limited real-world data to effectively bridge the sim-to-real gap. The key innovation lies in our autoregressive training approach, where the model learns from its own predictions to develop resilience against sensor noise and improve robustness in highly dynamic environments. Comprehensive experimental validation on the Booster T1 humanoid robot demonstrates that AutoOdom significantly outperforms state-of-the-art methods across all evaluation metrics, achieving 57.2% improvement in absolute trajectory error, 59.2% improvement in Umeyama-aligned error, and 36.2% improvement in relative pose error compared to the Legolas baseline. Extensive ablation studies provide critical insights into sensor modality selection and temporal modeling, revealing counterintuitive findings about IMU acceleration data and validating our systematic design choices for robust proprioceptive odometry in challenging locomotion scenarios.

Through long-term evolution, humans have developed extensive experience in adjusting body posture using the torso to reduce physical effort and create a comfortable workspace for arms. Imitation learning can efficiently transfer the human experience to robots. Therefore, we propose a joint-space torso motion planning network (TMPNet) for our humanoid upper-body robot to imitate the motion strategy between the human torso and dual arms, thereby guiding the robot torso motion. Compared to several mainstream inverse kinematics methods, TMPNet can enhance dual-arm flexibility through torso motion, compensate for the head’s degrees of freedom, and avoid sudden large angle changes and high iterative computation costs. To further reduce the robot’s energy consumption, a brain-inspired edge computer equipped with a neuromorphic chip is selected as the hardware computing platform. In addition, as the spiking neural network with its spike emission mechanism can fully leverage the neuromorphic chip’s advantages, it is employed to implement TMPNet.

With the rapid growth of IoT technology and the rise of smart devices, edge computing, particularly federated learning (FL), has gained importance for preserving user data privacy. However, FL faces challenges like non-independent identically distributed data and device heterogeneity, leading to model disparities and reduced precision. Our research proposes a novel adaptive FL framework specifically engineered to synchronize computational capacities within heterogeneous edge computing landscapes. Building upon the proof of convergence boundaries for local aggregation model, this algorithm adapts the number of iterations for local updates by considering the resource consumption relationship between local aggregation model and the local updated model by various clients. This method exhibit adaptability within an environment where disparities in edge device computational capacities exist, effectively balancing computational prowess among diverse devices and enhancing the output performance of federated learning Experiments on MNIST and PlantVillage datasets show that in heterogeneous environments, our algorithm outperforms existing methods, improving the loss function by at least 16.87% and the convergence speed by at least 2 times, in various environments (MobileNet, AlexNet).

This article presents an innovative approach to robotic kinematic control through a state-disturbance observer framework. Inspired by the dynamics of the human arm, this study introduces, for the first time, a virtual flexible joint dynamics primitive (VFJDP) and integrates it into the kinematic control of a humanoid upper body robot to generate smooth and precise movements. Two jounce-level control schemes leveraging the VFJDP are developed, along with a novel quasi-sliding mode disturbance observer for state feedback control. The VFJDP-based schemes enable precise trajectory tracking within joint angle and velocity constraints, significantly improving the robot’s manipulability and producing higher order joint commands. Theoretical analysis proves the asymptotic convergence of the observer and control algorithm. Comparative simulations show that the proposed observer improves state and disturbance estimation accuracy by over 80% compared to state-of-the-art methods. Simulations under noisy conditions further verify the robustness of the proposed approach. Furthermore, experiments involving visual servoing tasks validate the VFJDP-based schemes, achieving a 16% improvement in trajectory tracking accuracy and a 20% increase in manipulability compared to existing methods, including the improved clamping weighted least-norm (ICWLN) and zeroing neural network (ZNN) methods. These results confirm the proposed framework’s effectiveness in tackling kinematic control challenges in humanoid robotics.