Achieving general agile whole-body control on humanoid robots remains a major challenge due to diverse motion demands and data conflicts. While existing frameworks excel in training single motion-specific policies, they struggle to generalize across highly varied behaviors due to conflicting control requirements and mismatched data distributions. In this work, we propose BumbleBee (BB), an expert-generalist learning framework that combines motion clustering and sim-to-real adaptation to overcome these challenges. BB first leverages an autoencoder-based clustering method to group behaviorally similar motions using motion features and motion descriptions. Expert policies are then trained within each cluster and refined with real-world data through iterative delta action modeling to bridge the sim-to-real gap. Finally, these experts are distilled into a unified generalist controller that preserves agility and robustness across all motion types. Experiments on two simulations and a real humanoid robot demonstrate that BB achieves state-of-the-art general whole-body control, setting a new benchmark for agile, robust, and generalizable humanoid performance in the real world.

The partial observability and stochasticity in multi-agent settings can be mitigated by accessing more information about others via communication. However, the coordination problem still exists since agents cannot communicate actual actions with each other at the same time due to the circular dependencies. In this paper, we propose a novel multi-level communication scheme, Sequential Communication (SeqComm). SeqComm treats agents asynchronously (the upper-level agents make decisions before the lower-level ones) and has two communication phases. In the negotiation phase, agents determine the priority of decision-making by communicating hidden states of observations and comparing the value of intention, which is obtained by modeling the environment dynamics. In the launching phase, the upper-level agents take the lead in making decisions and then communicate their actions with the lower-level agents. Theoretically, we prove the policies learned by SeqComm are guaranteed to improve monotonically and converge. Empirically, we show that SeqComm outperforms existing methods in a variety of cooperative multi-agent tasks.

Pulmonary vessel CT segmentation is important to clinical diagnosis of lung diseases. But it is still a challenge due to limited CT quality and complicated vascular structures. This paper proposes new enhanced U-transformer networks that combine transformers, a contrast enhancement block with a reverse attention block to perform end-to-end vessel segmentation. Specifically, the contrast enhancement block directly augments edge or structural information while the reverse attention block conducts the network paying more attention to blurred boundaries and uncertain regions of vessels, leading to improving the accuracy and smoothness of pulmonary vessel segmentation. We validated our proposed method on 50 CT volumes selected from LIDC-IDRI, with the experimental results demonstrating that it works more effectively and stably than currently available approaches. Particularly, the average dice similarity coefficient and recall were improved from (85.23%, 85.37%) to (86.07%, 86.67%), respectively.

This paper proposes a hierarchical wireless system for real-time tracking of moving assets. There are multiple layers of wirelessly connected assets that need to be tracked. The small items, such as bottles, packed in a box are monitored by a sensor module contained within each box. All boxes shipped in a cargo truck form an ad hoc network that adapts to the battery level and link quality. A small subset of boxes are dynamically selected to report sensor information to a backend server. A wireless networking protocol is proposed to minimize the power consumption based on available energy and wireless link conditions in the distributed hierarchical system. The system is validated by simulation and prototype implementations.

Traditional routing protocols designed for multi-hop wireless networks select a route with the smallest hop count. The end-to-end throughput in the route may not be the maximum when network has multiple link data rates, as a small hop number implies a large geographic distance (link distance) for each hop. Due to the radio signal attenuation, this results in a low SNR at the receiving node and consequently a low available link data rate. Reducing the link distance by using routes with more hop counts can increase the link data rate. Yet this may not necessarily improve the end-to-end throughput because more nodes have to be included in the routes. In this work we investigate and analyze the impact of link distance on end-to-end throughput in multi-rate, multi-hop wireless networks. Analysis results show that changing link distance affects network throughput. To achieve a high network throughput, a proper link distance requirement has to be set for each hop, depending on different parameters such as load density.

This paper studies efficient and simple data broadcast in IEEE 802.15.4-based ad hoc networks (e.g., ZigBee). Since finding the minimum number of rebroadcast nodes in general ad hoc networks is NP-hard, current broadcast protocols either employ heuristic algorithms or assume extra knowledge such as position or two-hop neighbor table. However, the ZigBee network is characterized as low data rate and low cost. It cannot provide position or two-hop neighbor information, but it still requires an efficient broadcast algorithm that can reduce the number of rebroadcast nodes with limited computation complexity and storage space. To this end, this paper proposes self-pruning and forward node selection algorithms that exploit the hierarchical address space in ZigBee networks. Only one-hop neighbor information is needed; a partial list of two-hop neighbors is derived without exchanging messages between neighboring nodes. The ZigBee forward node selection algorithm finds the minimum rebroadcast nodes set with polynomial computation time and memory space. Using the proposed localized algorithms, it is proven that the entire network is covered. Simulations are conducted to evaluate the performance improvement in terms of the number of rebroadcast nodes, number of duplicated receivings, coverage time, and communication overhead.

Due to scarce resources, such as transmission power, storage space and communication bandwidth, current broadcast approaches for general ad hoc networks can not be applied to IEEE 802.15.4 based ad hoc networks (e.g., ZigBee networks). This paper proposes a forward node selection algorithm that significantly reduces broadcast redundancy. The algorithm exploits the hierarchical address space in ZigBee networks. Only one-hop neighbor information is needed: a partial list of two-hop neighbors is derived at a node without exchanging messages between neighboring nodes. The complexity of the proposed algorithm is polynomial in terms of both computation time and memory space. The localized algorithm provides an optimal and feasible solution of selecting the minimum number of rebroadcast nodes in ZigBee networks, which is an NP-hard problem for general ad hoc networks. The proposed algorithm is extended to deal with packet loss during data transmission. A ZigBee rebroadcast algorithm is also proposed to further reduce the number of rebroadcast nodes and cover the whole network faster by assigning a non-random rebroadcast timer determined by the number of neighbors to be covered, distance and link quality. Simulations are conducted to evaluate the broadcast redundancy, coverage time, and coverage ratio.