motion primitive design

Autonomous robots face challenges in decoupling trajectory planning from control and ensuring formal verification when executing Signal Temporal Logic (STL) tasks in complex environments. Method: This paper proposes a model-free planning–control co-design framework. It offline constructs a spatiotemporal motion primitive library via reinforcement learning and establishes a verifiable mapping from primitives to STL spatiotemporal semantics. Online, it integrates sampling-based STL-satisfying planning to synthesize safe motion sequences adhering to diverse temporal constraints. Contribution/Results: The framework requires no system dynamics model and enables end-to-end behavioral verification. Experiments on differential-drive and quadrupedal robot platforms demonstrate its effectiveness and generalization capability in dynamic obstacle avoidance and multi-objective temporal tasks.

Behavioral modeling in robotics lacks systematic empirical understanding of the practical differences and commonalities between Behavior Trees (BTs) and State Machines (SMs). Method: We conduct the first large-scale empirical comparison across 1,200+ open-source ROS projects, leveraging domain-specific language (DSL) parsing, code mining, and conceptual mapping to analyze BT and SM usage across language design, structural abstraction, reuse patterns, and engineering practice. Contribution/Results: We find a significant upward trend in BT DSL adoption; uncover deep isomorphisms between BTs and SMs in control-flow abstraction granularity and modular reuse mechanisms; and release RoboBT-SM-Bench—the first cross-DSL, fully annotated benchmark dataset of robotic behavioral models. This work establishes an empirical foundation and infrastructure support for unifying theoretical frameworks and designing reusable architectures for behavioral modeling languages.

Existing theoretical frameworks struggle to fully account for the empirical success of action chunking in behavioral cloning. Through both simulated and real-world robotic experiments, this work demonstrates that its core advantages stem from non-Markovian representational capacity, mitigation of error accumulation, and an implicit ensemble effect. We find that action chunking effectively learns diverse temporal dependencies, thereby achieving performance akin to model ensembling without explicit architectural modifications. Building on this insight, we propose a novel approach that reproduces the benefits of action chunking without requiring chunked action outputs, and further introduce an explicit ensemble strategy that significantly outperforms conventional action chunking across multiple tasks.

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.

Traditional hierarchical robotic planning simplifies task-level actions into open-loop kinematic skills, hindering seamless integration of pre-trained closed-loop motor controllers. Method: We propose Composable Interaction Primitives (CIPs), a framework enabling plug-and-play composition of heterogeneous, non-composable pre-trained skills within task-and-motion planning. Building upon CIPs, we introduce Task-and-Skill Planning (TASP), a unified architecture that jointly models symbolic task planning, geometric motion planning, and learned closed-loop control. Contribution/Results: TASP transcends reliance on motion-centric skills by elevating task semantics to the perception–action closed-loop level. Evaluated on a real mobile manipulator, it achieves end-to-end autonomous execution of multi-step complex tasks—including dynamic collaborative transport and tool manipulation—demonstrating significantly improved skill reusability and environmental adaptability.

This work addresses the limitations of robot learning imposed by the high cost of data collection and the limited diversity of human demonstrations, which often fail to cover the state space required for complex manipulation tasks. The authors propose a task-agnostic, black-box simulation framework that integrates RRT-style exploration, sampling-based model predictive control (MPC), and a novel sampling mechanism grounded in stable-state manifolds. This approach enhances exploration diversity while preserving action stability, enabling the direct construction of long-horizon policy trees in simulation. The method successfully generates a wide range of manipulation strategies—including pushing, grasping, rotating, throwing, and tool use—across diverse robot morphologies, thereby overcoming the constraints of conventional local trajectory optimization techniques.

This work addresses the challenge of reusing skills across tasks in offline hierarchical reinforcement learning by proposing a reusable skill abstraction mechanism grounded in local dynamics similarity. Building upon the HIQL framework, the method employs contrastive action representation learning (CARL) to automatically cluster semantically meaningful low-level skills and aligns action sequences across tasks to identify their applicable contexts. By capturing fine-grained dynamic similarities, the approach substantially enhances the high-level policy’s ability to effectively schedule low-level skills. Empirical evaluations demonstrate significant performance gains on downstream tasks in complex humanoid environments and the OGBench benchmark, highlighting the method’s capacity to improve cross-task skill transfer in offline settings.

This work addresses the limited exploration capability and difficulty in converging to global optima inherent in sampling-based controllers for path planning. To overcome these challenges, the authors propose integrating motion primitives into the Model Predictive Path Integral (MPPI) framework. By fusing motion primitive-guided structured sampling with perturbed control sequences within the real-time optimization loop, the method substantially enhances exploration efficiency and global optimality in the control space while preserving MPPI’s intrinsic fast response characteristics. Evaluations on quadrotor obstacle navigation tasks demonstrate that the proposed algorithm significantly improves both exploratory behavior and real-time performance, thereby validating its superiority over conventional approaches.

This work addresses the fragility and lack of reactivity inherent in action chunking strategies when executed open-loop under stochastic dynamics, actuation errors, and partial observability. To mitigate these limitations, the authors propose a test-time augmentation method that integrates a lightweight latent world model into the action chunking framework. Without requiring policy fine-tuning, the approach performs predictive rollouts over multiple candidate action chunks and selects the sequence whose predicted states best align with actual observations. This enables dynamic adaptation to perturbations without additional training, substantially enhancing robustness and responsiveness in long-horizon tasks. Experiments on the Kinetix benchmark and four real-world robotic manipulation tasks demonstrate significant improvements in success rates over baseline methods, particularly when demonstration data includes corrective behaviors.

This work addresses the limited generalizability and transferability of current foundation models in robotics, which often require custom integration of perception, actuation, and safety mechanisms. To overcome this, the authors propose ROSClaw—a model-agnostic execution layer that enables plug-and-play deployment of arbitrary foundation models on any ROS 2 robot by integrating the OpenClaw agent runtime with ROS 2. Key innovations include standardized capability discovery, multimodal observation normalization, action validation within configurable safety bounds, and structured audit logging. Experiments across three robotic platforms and four foundation models demonstrate up to a 4.8× difference in non-policy action proposal rates and show that the proposed execution layer significantly improves task success rates and safety across diverse frameworks.