Zhen Fu

Tengye Xu, Yangting Sun, Ziju Shen et al.

Designing reward functions that generalize beyond controlled laboratory settings remains a fundamental challenge in reinforcement learning for robotics. In open-world manipulation problems, a single task can appear in numerous variants through different object instances, positions, and camera viewpoints. Recent vision-based reward models tend to memorize specific pixel distributions and fail to generalize beyond their training conditions. To address this, we propose a framework that learns invariant symbolic reward functions from as few as five demonstrations. The insight is to shift from visual feature-fitting to the discovery of behavioral invariants: task-level properties that remain constant across diverse visual instantiations. The framework has two coupled components: a structural reward formulation that encodes task-level strategies and physical constraints while preserving optimal policy invariance, and a hybrid symbolic-numerical procedure that distills these invariants from demonstrations without online interaction. Experiments on eight Meta-World tasks and three Franka manipulation tasks demonstrate that our method achieves stronger process alignment and policy rollout ranking abilities compared to baselines, accelerating downstream policy learning. Three real-world out-of-distribution experiments further show that the same learned reward generalizes zero-shot to position, viewpoint, and object variations, enabling a single reward representation to be reused across diverse task variants in practice.

Wenqi Ge, Junde Guo, Zhen Fu et al.

Achieving everyday tasks with humanoid robots requires coordinating stable locomotion with versatile manipulation. However, existing whole-body controllers still face significant challenges. Methods trained solely via command sampling, without motion-capture (MoCap) data, often struggle with sparse rewards and require carefully tuned curricula to converge. This is especially problematic for upper-body control, where the resulting motions deviate from human-like statistics and degrade whole-body coordination. Conversely, approaches that imitate full-body MoCap data suffer from dataset imbalance, as many locomotion trajectories are overly aggressive for stable-locomotion scenarios, necessitating extensive data filtering and augmentation. To address this, we present Composite Whole-Body Imitation (CWI), a framework that decouples the use of MoCap data for upper-body manipulation and lower-body locomotion. This decoupling allows us to exploit the full MoCap dataset of diverse manipulation references, while stable, command-conditioned lower-body locomotion is guided by dual discriminators trained on curated expert-quality walking and squatting clips via an Adversarial Motion Prior (AMP). A multi-critic architecture reduces conflicts among locomotion, manipulation, and motion-style objectives, and a teacher--student distillation stage yields a whole-body policy conditioned only on bimanual hand poses and velocity/height commands. We evaluate CWI through simulation experiments and real-world deployment on a full-size LimX Oli humanoid. The results show competitive loco-manipulation performance, robust whole-body coordination, and practical teleoperation without full-body motion-capture equipment. A project page with supplementary material can be found at https://cwi-ral.github.io/CWI-RAL-Webpage.

Kaiwen Jiang, Zhen Fu, Junde Guo et al.

In this paper, we study the whole-body loco-manipulation problem using reinforcement learning (RL). Specifically, we focus on the problem of how to coordinate the floating base and the robotic arm of a wheeled-quadrupedal manipulator robot to achieve direct six-dimensional (6D) end-effector (EE) pose tracking in task space. Different from conventional whole-body loco-manipulation problems that track both floating-base and end-effector commands, the direct EE pose tracking problem requires inherent balance among redundant degrees of freedom in the whole-body motion. We leverage RL to solve this challenging problem. To address the associated difficulties, we develop a novel reward fusion module (RFM) that systematically integrates reward terms corresponding to different tasks in a nonlinear manner. In such a way, the inherent multi-stage and hierarchical feature of the loco-manipulation problem can be carefully accommodated. By combining the proposed RFM with the a teacher-student RL training paradigm, we present a complete RL scheme to achieve 6D EE pose tracking for the wheeled-quadruped manipulator robot. Extensive simulation and hardware experiments demonstrate the significance of the RFM. In particular, we enable smooth and precise tracking performance, achieving state-of-the-art tracking position error of less than 5 cm, and rotation error of less than 0.1 rad. Please refer to https://clearlab-sustech.github.io/RFM_loco_mani/ for more experimental videos.

Shunpeng Yang, Hua Chen, Zhen Fu et al.

This paper studies stabilizer design for position-controlled humanoid robots. Stabilizers are an essential part for position-controlled humanoids, whose primary objective is to adjust the control input sent to the robot to assist the tracking controller to better follow the planned reference trajectory. To achieve this goal, this paper develops a novel force-feedback based whole-body stabilizer that fully exploits the six-dimensional force measurement information and the whole-body dynamics to improve tracking performance. Relying on rigorous analysis of whole-body dynamics of position-controlled humanoids under unknown contact, the developed stabilizer leverages quadratic-programming based technique that allows cooperative consideration of both the center-of-mass tracking and contact force tracking. The effectiveness of the proposed stabilizer is demonstrated on the UBTECH Walker robot in the MuJoCo simulator. Simulation validations show a significant improvement in various scenarios as compared to commonly adopted stabilizers based on the zero-moment-point feedback and the linear inverted pendulum model.

Zhen Fu, Bo Wang, Fei Chen et al.

Hearing-impaired listeners usually have troubles attending target talker in multi-talker scenes, even with hearing aids (HAs). The problem can be solved with eye-gaze steering HAs, which requires listeners eye-gazing on the target. In a situation where head rotates, eye-gaze is subject to both behaviors of saccade and head rotation. However, existing methods of eye-gaze estimation did not work reliably, since the listener's strategy of eye-gaze varies and measurements of the two behaviors were not properly combined. Besides, existing methods were based on hand-craft features, which could overlook some important information. In this paper, a head-fixed and a head-free experiments were conducted. We used horizontal electrooculography (HEOG) and neck electromyography (NEMG), which separately measured saccade and head rotation to commonly estimate eye-gaze. Besides traditional classifier and hand-craft features, deep neural networks (DNN) were introduced to automatically extract features from intact waveforms. Evaluation results showed that when the input was HEOG with inertial measurement unit, the best performance of our proposed DNN classifiers achieved 93.3%; and when HEOG was with NEMG together, the accuracy reached 72.6%, higher than that with HEOG (about 71.0%) or NEMG (about 35.7%) alone. These results indicated the feasibility to estimate eye-gaze with HEOG and NEMG.

Zhen Fu, Bo Wang, Xihong Wu et al.

The auditory attention decoding (AAD) approach was proposed to determine the identity of the attended talker in a multi-talker scenario by analyzing electroencephalography (EEG) data. Although the linear model-based method has been widely used in AAD, the linear assumption was considered oversimplified and the decoding accuracy remained lower for shorter decoding windows. Recently, nonlinear models based on deep neural networks (DNN) have been proposed to solve this problem. However, these models did not fully utilize both the spatial and temporal features of EEG, and the interpretability of DNN models was rarely investigated. In this paper, we proposed novel convolutional recurrent neural network (CRNN) based regression model and classification model, and compared them with both the linear model and the state-of-the-art DNN models. Results showed that, our proposed CRNN-based classification model outperformed others for shorter decoding windows (around 90% for 2 s and 5 s). Although worse than classification models, the decoding accuracy of the proposed CRNN-based regression model was about 5% greater than other regression models. The interpretability of DNN models was also investigated by visualizing layers' weight.