motion retargeting

This work addresses the “execution gap” between high-level semantic tasks and executable robot motions by introducing Motion Statecharts—a symbolic, executable motion representation that supports concurrency and hierarchical nesting. Coupled with a unified differentiable kinematic world model, this framework enables end-to-end mapping from semantic task specifications to low-level motion control. Smooth and dynamically feasible trajectories are generated through a linear model predictive control (lMPC)-driven task-function approach incorporating snap (jerk derivative) constraints. The proposed system has been successfully deployed across eight heterogeneous robotic platforms, demonstrating strong cross-platform generalization and real-world efficacy. The accompanying software framework, Giskard, has been publicly released.

In teleoperation of complex assembly tasks—such as sequential screw-driving—the tight coupling between translational and rotational motions leads to prolonged task completion time, large alignment errors, and low success rates. To address this, we propose a task-specific, decoupled motion retargeting framework. Our approach formally defines a dynamic decoupling mechanism between translation and rotation coordinate systems, enabling real-time coordinate alignment and overcoming limitations of conventional one-time calibration or discrete mode-switching strategies. The method integrates optimal control-based trajectory planning, VR controller input parsing, shared control architecture, and real-time kinematic mapping onto a UR5e manipulator. Experimental evaluation demonstrates significant improvements: task completion time is substantially reduced; rotational alignment error decreases by 62%; and the success rate for sequential screw-driving reaches 94%.

This work addresses the challenge of transferring motion trajectories across 3D scenes with significant structural differences while preserving semantic consistency, spatial coherence, and avoiding collisions or geometric distort日消息. To this end, the authors propose a training-free trajectory analogy transfer method that first clusters scenes in an object-centric manner, leverages 3D foundation models to extract features, and constructs a cross-scene hierarchical smooth mapping. The method then integrates mappings from individual clusters through a spatial decoupling and combinatorial optimization strategy. Requiring only approximately 0.6 seconds per transfer, the approach substantially outperforms baseline methods based on large language models, vision-language models, and scene graph matching, and demonstrates successful applications in virtual co-presence, multi-trajectory transfer, camera path migration, and human-robot motion transfer.

In cross-morphological character motion retargeting, contact semantics—especially contact points—are often lost, leading to motion distortion. To address this, we propose the first contact-aware real-time motion retargeting framework. Our method automatically identifies binding key vertices to construct a low-dimensional shape-pose joint embedding; it introduces trajectory-based motion descriptors and a dynamic feature weighting optimization scheme that adaptively preserves source-contact constraints within a constrained optimization formulation. The framework supports multi-character coordinated retargeting and adaptation to non-planar terrain. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods in contact accuracy and motion smoothness, while maintaining robustness and real-time performance across diverse morphologies, complex interactions, and highly uneven terrains.

This work addresses the challenge of enabling quadrupedal robots to dynamically imitate animal locomotion across morphologically distinct bodies. We propose a Spatio-Temporal Motion Retargeting (STMR) framework that decouples Spatial Motion Retargeting (SMR) and Temporal Motion Retargeting (TMR), respectively ensuring kinematic feasibility and dynamic executability—overcoming the failure of conventional methods on highly dynamic aerial phases. The method integrates keypoint trajectory mapping, full-body inverse kinematics, temporal-domain optimization, and end-to-end imitation learning, supporting monocular video input and real-hardware deployment. Evaluated in simulation and on two physically distinct quadruped platforms, STMR successfully reproduces six complex animal gaits with high-fidelity trajectory tracking. Quantitative results demonstrate significant performance gains over baseline approaches in both accuracy and dynamic consistency.

Existing text-driven 3D human motion generation methods struggle to simultaneously preserve linguistic descriptions of gait, style, and intent while accurately satisfying geometric constraints such as root or multi-joint trajectories. This work proposes KV-Control—a lightweight attention-side control interface that injects trajectory constraints as key/value memory into the self-attention layers of a frozen masked text-to-motion Transformer, thereby avoiding interference with the pre-trained query stream and text cross-attention. By integrating PartVQ for decoupled motion substructures, a T-Concat frame-part token mechanism, and a shared trajectory encoder, the method introduces only a minimal number of trainable parameters. Under a refinement inheritance protocol, it achieves sub-centimeter trajectory tracking accuracy while maintaining the naturalness and semantic consistency of generated motions.

This work addresses the common decoupling of task scheduling and motion planning in human-robot collaboration, which often stems from the difficulty of jointly modeling users’ individualized spatiotemporal behaviors, thereby limiting both efficiency and safety. To bridge this gap, the authors propose RAPIDDS—a novel framework that unifies task-level scheduling and motion-level planning for the first time. RAPIDDS adaptively learns personalized user spatiotemporal preferences through iterative interactions and leverages a diffusion model to jointly optimize robotic trajectories. This approach enables coherent modeling and co-optimization of human spatiotemporal characteristics. Comprehensive evaluations—including simulations, real-world experiments on a 7-DoF robotic arm, and a user study with 32 participants—demonstrate that RAPIDDS significantly outperforms non-adaptive baselines across multiple objective and subjective metrics, including collaboration efficiency, proximity, fluency, and user preference.

This work proposes an online trajectory generation method based on piecewise quintic/quartic splines to address the challenge of converting arbitrary geometric paths into kinematically feasible and collision-free trajectories in dynamic environments. The approach explicitly enforces jerk constraints and supports real-time replanning under high-frequency goal updates. By integrating dynamic environment perception and a responsive adaptation mechanism, it guarantees collision avoidance within finite time while permitting bounded deviations from the original path. Both simulation and real-world experiments demonstrate that the method outperforms existing approaches in trajectory smoothness, computational efficiency, and real-time performance, achieving stable operation in human-in-the-loop dynamic scenarios with target update rates up to 1 kHz.

This work addresses the challenge of preserving interaction semantics—such as self-contacts and proximal distances—in cross-body motion retargeting. The authors propose a geometry-aware motion retargeting framework that employs spatially adaptive anchors to achieve proximity-aware matching. These anchors are dynamically adjusted via differentiable soft projection combined with a Transformer-based optimization strategy to align with the target character’s reachable regions. To maintain the spatial structure of the source motion, a graph autoencoder is integrated into the pipeline. An alternating training scheme jointly optimizes anchor adaptation and motion generation. Experiments demonstrate that the proposed method significantly outperforms existing approaches across diverse body geometries, achieving markedly improved fidelity in interaction preservation.