Mingjie Zhang

Jiongjiong Gu, Jianfeng Wang, Zidong Han et al.

Modern AI serving increasingly relies on NPUs for conventional inference and large language model serving. However, current NPU deployments commonly expose physical devices directly to applications, which limits runtime control over scheduling and makes it difficult to adapt execution to phase-level workload behavior. This limitation is particularly evident in LLM serving, where the prefill phase is compute-intensive while the decode phase is often constrained by memory bandwidth and KV-cache accesses. Static prefill-decode (PD) disaggregation reduces phase interference, but can introduce resource imbalance and unnecessary data movement. We present FlexNPU, a transparent user-space virtualization layer for Ascend NPUs. FlexNPU interposes on AscendCL APIs and routes NPU operations through per-device daemons, decoupling unmodified from physical NPU devices without modifying model code, AI frameworks, or NPU drivers. This runtime boundary allows FlexNPU to virtualize NPU objects, control operator dispatch, and support phase-aware scheduling for LLM serving. In particular, FlexNPU enables dynamic PD co-location, which adapts scheduling between prefill and decode according to their complementary resource characteristics. We implement FlexNPU on Huawei Ascend NPUs and evaluate it with typical LLM workloads. Compared with direct NPU passthrough, FlexNPU introduces no measurable inference overhead and slightly improves throughput in some scenarios. On a 384-card Ascend 910C deployment of DeepSeek-R1, FlexNPU improves throughput over static PD disaggregation by 5.15% and 26.33%. On Qwen2.5-7B, compared with static PD co-location, FlexNPU maintains comparable throughput while reducing TTFT by over 92% across tested workloads with nearly unchanged TPOT. These results show that transparent NPU virtualization is a practical substrate for efficient and responsive LLM serving.

Guiyong Zheng, Yueting Ban, Mingjie Zhang et al.

Aerial vision-language navigation (AVLN) enables UAVs to follow natural-language instructions in complex 3D environments. However, existing zero-shot AVLN methods often suffer from unstable single-stream Vision-Language Model decision-making, unreliable long-horizon progress monitoring, and a trade-off between safety and efficiency. We propose OnFly, a fully onboard, real-time framework for zero-shot AVLN. OnFly adopts a shared-perception dual-agent architecture that decouples high-frequency target generation from low-frequency progress monitoring, thereby stabilizing decision-making. It further employs a hybrid keyframe-recent-frame memory to preserve global trajectory context while maintaining KV-cache prefix stability, enabling reliable long-horizon monitoring with termination and recovery signals. In addition, a semantic-geometric verifier refines VLM-predicted targets for instruction consistency and geometric safety using VLM features and depth cues, while a receding-horizon planner generates optimized collision-free trajectories under geometric safety constraints, improving both safety and efficiency. In simulation, OnFly improves task success from 26.4% to 67.8%, compared with the strongest state-of-the-art baseline, while fully onboard real-world flights validate its feasibility for real-time deployment. The code will be released at https://github.com/Robotics-STAR-Lab/OnFly

Xinlu Yan, Mingjie Zhang, Yuhao Fang et al.

Efficient multi-UAV exploration under limited communication is severely bottlenecked by inadequate task representation and allocation. Previous task representations either impose heavy communication requirements for coordination or lack the flexibility to handle complex environments, often leading to inefficient traversal. Furthermore, short-horizon allocation strategies neglect spatiotemporal contiguity, causing non-contiguous assignments and frequent cross-region detours. To address this, we propose C$^2$-Explorer, a decentralized framework that constructs a connectivity graph to decompose disconnected unknown components into independent task units. We then introduce a contiguity-driven allocation formulation with a graph-based neighborhood penalty to discourage non-adjacent assignments, promoting more contiguous task sequences over time. Extensive simulation experiments show that C$^2$-Explorer consistently outperforms state-of-the-art (SOTA) baselines, reducing average exploration time by 43.1\% and path length by 33.3\%. Real-world flights further demonstrate the system's feasibility. The code will be released at https://github.com/Robotics-STAR-Lab/C2-Explorer

Chengkai Wu, Ruilin Wang, Yixin Zeng et al.

Humans seamlessly fuse anticipatory planning with immediate feedback to perform successive mobile manipulation tasks without stopping, achieving both high efficiency and reliability. Replicating this fluid and reliable behavior in robots remains fundamentally challenging, not only due to conflicts between long-horizon planning and real-time reactivity, but also because excessively pursuing efficiency undermines reliability in uncertain environments: it impairs stable perception and the potential for compensation, while also increasing the risk of unintended contact. In this work, we present a unified framework that synergizes efficiency and reliability for continuous mobile manipulation. It features a reliability-aware trajectory planner that embeds essential elements for reliable execution into spatiotemporal optimization, generating efficient and reliability-promising global trajectories. It is coupled with a phase-dependent switching controller that seamlessly transitions between global trajectory tracking for efficiency and task-error compensation for reliability. We also investigate a hierarchical initialization that facilitates online replanning despite the complexity of long-horizon planning problems. Real-world evaluations demonstrate that our approach enables efficient and reliable completion of successive tasks under uncertainty (e.g., dynamic disturbances, perception and control errors). Moreover, the framework generalizes to tasks with diverse end-effector constraints. Compared with state-of-the-art baselines, our method consistently achieves the highest efficiency while improving the task success rate by 26.67\%--81.67\%. Comprehensive ablation studies further validate the contribution of each component. The source code will be released.