Yongbin Jin

Robotic piano playing is a challenging benchmark for biomimetic dexterous manipulation, requiring precise timing, coordinated multi-finger motion, and stable key contact. This study proposes a robotic piano-playing framework based on a two-stage reinforcement learning curriculum. Musical Instrument Digital Interface (MIDI) data are converted into target-key and fingering grids to provide future musical goals for policy learning in a parallel MJLab simulation environment. A Soft Actor–Critic (SAC) agent takes a 2106-dimensional observation vector, including joint states, previous actions, musical phase, future key targets, fingering assignments, and piano-key states, and outputs a 21-dimensional continuous action vector for wrist, finger, and global hand-positioning control. Stage 1 weakens physical regularization to facilitate key-pressing acquisition, whereas Stage 2 strengthens power, velocity, acceleration, collision, posture, and finger-speed constraints to improve the regularity of policy outputs and readiness for real-world deployment. Simulation experiments on 30 s right-hand excerpts from Für Elise, Canon, and Beethoven’s Symphony No. 5 achieve frame-wise key-state F1 scores above 0.99 on the first two excerpts and approximately 0.945 on Beethoven. Real-world deployment uses open-loop playback of policy-generated high-level trajectories with low-level joint-position feedback and achieves F1 scores of 0.95, 0.91, and 0.83, respectively, while reproducing representative piano techniques such as chords, octaves, mixed black-and-white-key patterns, overlapping finger actions, and rapid sequential movements. These physical results demonstrate the feasibility of the proposed sim-to-real pipeline for complete 30 s executions; they are not intended as a statistical repeatability study. The results further show that biomimetic robotic hands can learn complex piano-playing skills from MIDI-based task objectives without relying on human motion demonstration trajectories.

Existing quadruped robots typically place their center of mass (CoM) near the geometric center of the body to achieve structural symmetry and simplify control design. In contrast, many quadrupedal animals capable of agile running exhibit a pronounced anteriorly biased mass distribution, with the CoM located closer to the front of the body. This biological characteristic motivates a re-examination of whether a geometrically centered CoM necessarily corresponds to dynamically balanced loading between the fore- and hindlimbs during high-speed locomotion. To address this question, this study investigates the influence of longitudinal CoM position on load distribution during high-speed straight-line locomotion of quadruped robots. A unified analytical framework is established by combining whole-body force and pitch moment equilibrium, sagittal-plane kinematics, and Jacobian-based force-to-torque mapping, thereby linking longitudinal CoM position, foot-end support forces, and joint loads. Simulation validation is conducted on the Black Panther 2 quadruped robot using four central-body CoM configurations, denoted as ×0, ×5, ×10, and ×15. In the primary evaluation at 5 m/s, shifting the CoM forward from ×0 to ×15 reduces the absolute median fore–hindlimb differences in support force and joint torque by approximately 86.6% and 93.4%, respectively, indicating a transition from hindlimb-dominated loading toward cooperative load sharing between the fore and hindlimbs. Independent training runs with multiple random seeds further confirm the robustness of this load-redistribution trend to reinforcement learning variability. Consistent behavior is also observed at 8 m/s, while no evident degradation in turning response or locomotion stability is found under the tested turning and randomly generated rough-terrain conditions. These results demonstrate that a moderate forward shift of the longitudinal CoM can alleviate hindlimb load concentration and promote a more balanced fore–hindlimb load distribution, providing a theoretical basis for the morphological design and control optimization of high-speed quadruped robots.

The inherent trade-off between structural compactness and functional versatility poses a fundamental challenge in biomimetic robotic hand design. This paper presents a multi-objective optimization framework for determining the optimal phalanx length allocation of a biomimetic finger under a fixed total-length constraint. Three performance criteria are formulated: grasp capability, measured by the area of a novel shared workspace (Region II) between power grasping and precision manipulation; kinematic dexterity, evaluated as the global average of the reciprocal condition number of the Jacobian matrix; and key-press range, defined as the maximum static fingertip span under perpendicularity and slope constraints. A full grid search reveals distinct optimal configurations for each objective. Pareto analysis of 117 non-dominated solutions shows that the key-press range is most sensitive to dimensional variations, with a 24.7% performance spread. A hierarchical selection strategy that prioritizes the key-press range while balancing the other two objectives yields a recommended compromise design. Experimental validation of the key-press reachability for the recommended compromise design achieves a 98.8% keystroke success rate over 1000 cross-row strikes without wrist movement. These results confirm its practical feasibility for fine manipulation tasks, while experimental characterization of the grasp capability and kinematic dexterity objectives remains as future work.

Quantifying motion similarity, despite its inherently subjective nature, is a foundational problem for humanoid motion transfer and control across different embodiments. Drawing from cognitive studies revealing that human motion similarity judgments are strongly influenced by spatial relationships among body parts and proximal contacts, we develop a graph-based representation that effectively encapsulates both features. This representation enables the definition of a robust similarity metric through graph distance computations. The proposed metric emphasizes spatial, especially proximal, relationships between body parts, facilitating motion retargeting that preserves these perceptually motivated relational cues across humanoid embodiments with shared semantic body parts and varying Degree of Freedom (DoF) configurations and body proportions. For quantitative retargeting evaluation, we introduce an order-preserving spatial similarity metric that measures how consistently inter-joint distance rankings are preserved between source and target motions.

The design of dexterous robotic hands has long been constrained by empirical paradigms that struggle to balance anthropomorphic fidelity with dynamic performance. This study aims to establish a systematic methodology that bridges this gap through data-driven inverse design. We construct a quantitative association map between design variables and performance metrics using a comprehensive dataset of existing dexterous hands, then apply this map to translate explicit high-frequency dynamic targets into an optimized hardware configuration. The analysis reveals that the dominant principles for high-speed performance—tendon-driven transmission, proximal actuation, and lightweight rigid structures—closely mirror the biomechanical architecture of the human hand. Guided by this convergence, we develop the Beyond Hand, a 20-degree-of-freedom (DoF) anthropomorphic hand that preserves human-scale dimensions. Standardized frequency-response tests across all 15 joints show magnitude attenuation below 3 dB at 14 Hz and cutoff frequencies clustered around 10 Hz. In rhythm-game and Tetris-style manipulation tasks, the hand maintains over 90% accuracy at actuation frequencies up to 12 Hz. These results demonstrate that a performance-driven pathway can systematically elevate the dynamic capabilities of humanoid dexterous hands, offering a scalable framework for biomimetic robotic design.

Background: Expressive piano performance poses extreme challenges for robotic manipulation, necessitating high-speed repetitive impacts, substantial force output, and coordinated multi-joint control under stringent dynamic constraints. However, existing robotic systems exhibit significant limitations in replicating human-level dexterity, as well as achievable motion speed and force output. This work presents a data-driven, bio-inspired dexterous robotic hand designed specifically for high-fidelity piano performance. Methods: We first extract kinematic primitives and stable inter-joint coupling patterns from large-scale motion capture data of professional pianists. These human motion priors are directly embedded into the mechanical architecture through morphological coupling and actuator allocation. Actuator selection is further guided by empirically measured human peak velocities and force profiles from biomechanics literature, ensuring sufficient bandwidth for high-speed repetitive motion and adequate force transmission. Results: Experimental results demonstrate that the proposed hand replicates human-like joint coordination, achieves peak joint velocities of 53.88 rad/s, and provides sufficient fingertip force for authentic piano interaction. As a demonstration of its capabilities, the hand successfully performs a Grade 7 piano piece, Croatian Rhapsody, illustrating its potential for expressive musical performance. Conclusion: This research establishes a principled pathway from human motion statistics to embodied robotic intelligence, providing a high-performance hardware foundation for autonomous musical performance.

Robotic racket sports provide exceptional benchmarks for evaluating dynamic motion control capabilities in robots. Due to the highly non-linear dynamics of the shuttlecock, the stringent demands on robots' dynamic responses, and the convergence difficulties caused by sparse rewards in reinforcement learning, badminton strikes remain a formidable challenge for robot systems. To address these issues, this study proposes DTG-IRRL, a novel learning framework for badminton strikes that integrates imitation-relaxation reinforcement learning with dynamic trajectory generation. The framework demonstrates significantly improved training efficiency and performance, achieving faster convergence and twice the landing accuracy. Analysis of the reward function within a specific parameter space hyperplane intuitively reveals the convergence difficulties arising from the inherent sparsity of rewards in racket sports and demonstrates the framework's effectiveness in mitigating local and slow convergence. Implemented on hardware with zero-shot transfer, the framework achieves a 90% hitting rate and a 70% landing accuracy, enabling sustained humanrobot rallies. Cross-platform validation using the UR5 robot demonstrates the framework's generalizability while highlighting the requirement for high dynamic performance of robotic arms in racket sports.

In industrial applications, precision and load capacity are critical performance metrics for robotic arms. These requirements are typically addressed using high-gear-ratio designs. However, robotic arms are increasingly tasked with high-dynamic operations, such as smashes in sports and object throwing, where traditional design paradigms and evaluation metrics are insufficient. Drawing inspiration from the dynamics of ball sports, we establish an optimization design paradigm for high-dynamic robotic arms by introducing angular momentum as performance metric. Based on this paradigm, we developed a 7-DOF robotic arm, KirinArm, and evaluated its performance in a badminton swinging task. KirinArm achieves a maximum end-effector speed of 10.1 m/s and a peak acceleration of 234.7 m/s$^{2}$, outperforming known commercial collaborative robotic arms. Notably, its racket tip reaches a speed of 31.3 m/s, equivalent to the level of an amateur badminton player and 48.45% of the performance of professional athletes under similar conditions. Comparative analysis with five other robotic arms, based on maximum angular momentum and acceleration metrics, highlights KirinArm's superior performance, even under increased dynamic loads.