Embodied Intelligence Networks (EIN)

Vision-Language Navigation (VLN) aims to enable embodied AI agents to follow natural language instructions and navigate in real indoor environments, but training strong VLN policies typically requires large-scale trajectory-instruction data collected in private spaces, including homes and offices, raising substantial privacy concerns. Federated Learning (FL) provides a natural remedy by keeping data on-device and only exchanging model updates; however, vanilla FL struggles in VLN due to extreme cross-client heterogeneity in environment layouts, navigation graphs, and personalized instruction styles, making a single global model suboptimal for many clients. In this paper, we propose pFedNavi, a structure-aware and dynamically adaptive personalized federated learning framework tailored for VLN. Our key idea is to personalize where it matters: pFedNavi (i) adaptively identifies client-specific layers via layer-wise mixing coefficients, and (ii) performs fine-grained parameter fusion on the selected components (e.g., the encoder-decoder projection and environment-sensitive decoder layers) to balance global knowledge sharing with local specialization. We evaluate pFedNavi on two standard VLN benchmarks, R2R and RxR, under both ResNet and CLIP visual representations. Across metrics, pFedNavi consistently outperforms the baseline-FedAvg-based VLN, achieving improvement in navigation success (up to 7.5% in success rate) and navigation trajectory quality (up to 7.8% in normalized dynamic time warping), while converging 1.38× faster under non-IID conditions.

The integration of Large Language Models (LLMs) fosters Embodied AI (EAI) agents to perceive and interact with the physical world through natural language instructions. However, existing EAI agents typically operate within fixed agentic workflows and predefined design spaces, struggling to handle rapidly evolving real-world EAI scenarios (e.g., diverse task pipelines, dynamic environments). Unlike prior systems that treat LLMs mainly as text-generators following rigid workflows, we fully exploit their reasoning capabilities to allow agents to determine their own workflows on the fly. To this end, we propose FSEAI that encourages EAI agents to self-explore and self-evolve via federated collaboration across heterogeneous environments. Inspired by human learning processes, we distill EAI tasks into three atomic operations (observe, reason, act), and empower agents to explore workflows by dynamically and adaptively selecting an appropriate next operation based on the current state. Our evaluations show that, with federated collaboration, our FSEAI agents can achieve up to a 42.6% higher task success rate and a 41.6K token cost reduction than state-of-the-art (SOTA) baselines, while maintaining adaptability to unforeseen EAI scenarios. This highlights the potential of reasoning-driven adaptive agentic workflow towards cognitive EAI.

Replanning is a core and frequent process in embodied AI systems, where agents continuously adapt their plans based on partial observations and dynamic environments. With the increasing adoption of large language models and vision-language models as high-level planners, each replanning cycle requires processing long context histories that grow over time and with the number of agents, making replanning inference cost a critical system bottleneck. However, existing work largely treats this cost as unavoidable, lacking system-level optimization methods specifically targeting the replanning stage. We observe that replanning is a high-level reasoning process that can tolerate information approximation, and not all historical tokens are equally important for current replanning decisions. We propose E-RECAP, a cost-aware token pruning method for embodied replanning that dynamically removes low-importance tokens from planner context while preserving replanning-critical information. E-RECAP is a system-level, drop-in optimization module that operates only during replanning, without modifying task definitions, environments, or control policies. Experimental results demonstrate significant acceleration: up to 2.64 times speedup in single-GPU embodied evaluations and up to 39.7 times speedup in long-context synthetic replanning under multi-GPU inference, while maintaining task success rates and replanning behavior. E-RECAP provides a plug-and-play solution for scalable embodied AI systems, particularly beneficial for long-horizon tasks and multi-agent scenarios where context accumulation is inevitable.

Indoor fire disasters pose severe challenges to autonomous search and rescue due to dense smoke, high temperatures, and dynamically evolving indoor environments. In such time-critical scenarios, multi-agent cooperative navigation is particularly useful, as it enables faster and broader exploration than single-agent approaches. However, existing multi-agent navigation systems are primarily vision-based and designed for benign indoor settings, leading to significant performance degradation under fire-driven dynamic conditions. In this paper, we present VULCAN, a multi-agent cooperative navigation framework based on multi-modal perception and vision-language models (VLMs), tailored for indoor fire disaster response. We extend the Habitat-Matterport3D benchmark by simulating physically realistic fire scenarios, including smoke diffusion, thermal hazards, and sensor degradation. We evaluate representative multi-agent cooperative navigation baselines under both normal and fire-driven environments. Our results reveal critical failure modes of existing methods in fire scenarios and underscore the necessity of robust perception and hazard-aware planning for reliable multi-agent search and rescue.

Embodied Artificial Intelligence (Embodied AI) depends on tightly coupled perception-action loops that require continuous sensing, low-latency communication, distributed reasoning, and coordinated actuation, making it inherently dependent on large-scale Internet of Things (IoT) infrastructures. Despite this natural alignment, existing IoT systems were largely designed for data acquisition and connectivity rather than for closed-loop embodied intelligence, leading to fundamental architectural mismatches that constrain the scalability, responsiveness, and robustness of embodied AI in real-world deployments. This paper presents a challenge-oriented technical perspective on the role of IoT in embodied AI. Using a four-layer embodied AI stack (i.e., sensing, cognition, action, and networking) as a unifying analytical framework, we systematically examine how current IoT technologies support, shape, and limit embodied AI systems. Our analysis exposes key tensions spanning semantic sensing, latency-efficiency trade-offs in communication, fragmentation of distributed computation, and weak coordination of networked actuation. Based on these findings, we argue for new abstractions and cross-layer co-design principles to align future IoT infrastructures with the requirements of closed-loop embodied intelligence.

Edge devices for embodied AI operate under tight computation and energy budgets, motivating edge-cloud collaboration for high-quality closed-loop control. However, conventional text-to-text (T2T) collaboration introduces substantial end-to-end (E2E) latency due to token-by-token decoding, and suffers from output-format drift that undermines dependable actuation. We propose KV-SC, a key-value (KV) cache-based semantic collaboration framework that replaces intermediate text messages with KV-cache transmission. The cloud performs prefill-only inference to produce KV caches as structured semantic states, which are then injected into the edge controller to decode stable, machine-parsable actions. To further reduce communication overhead, we introduce a two-stage KV-cache compression method and show that aggressive compression preserves control quality while significantly reducing downlink payload. Experiments in a multivehicle VMAS road-traffic environment demonstrate that KV-SC improves closed-loop driving performance and consistently reduces E2E latency compared to T2T baselines.

Consumer robotics demands consolidation of safety-critical control, perception pipelines, and user applications on shared multicore platforms. While static partitioning hypervisors provide hardware-enforced isolation, directly transplanting automotive architectures encounters an expertise asymmetry problem in which end-users modifying robot behavior lack the systems knowledge that platform developers possess. We present an architecture addressing this challenge through three integrated components. A Safe IO Cell provides hardware-level override capability. A Parameter Synchronization Service encapsulates cross-domain complexity. A Safety Communication Layer implements IEC 61508-compliant verification. Our empirical evaluation on an ARM Cortex-A55 platform demonstrates that partition isolation reduces cycle-period jitter by 84.5% and cuts tail timing error by nearly an order of magnitude (p99 (|jitter|) from 69.0 (μs) to 7.8 (μs)), eliminating all >50 (μs) excursions.

Resource-constrained entities in an embodied intelligence network can be trained collaboratively by federated learning (FL). However, the data heterogeneity of these physical entities poses a significant difficulty in training convergence, particularly for the non-independent and identically distributed (non-IID) problem. Although hierarchical FL (HFL) uses clustering approaches to group similar data through two-phase model aggregation, it still suffers from the data heterogeneity of two inter-clusters. To solve the problem, this paper presents a fairness aggregation method based on the Minimum Enclosing Circle (MEC) and named it MEC-based HFL. The proposed MEC-based HFL calculates the center of the smallest enclosing circle that encompasses all data points from physical entities at the edge. It provides a fairer aggregation result by being closer to underrepresented clusters and balancing contributions from all edge devices. Extensive experiments demonstrate that the MEC-based HFL offers superior learning accuracy, faster model convergence, and better training efficiency compared to other classic FL algorithms.

Performance and cost optimization are always the top demand of OS. Traditional OS architectures struggle with optimization due to growing system complexity and inadequate heuristic methods. By leveraging advancements in generative AI and large models, this positioning paper defines a new OS paradigm for self-optimization, self-adaptation, and self-evolution. Furthermore, achieving this paradigm is viable due to system knowledge and tunability embedding. Our existing case studies and deployment practices demonstrate that integrating accumulated empirical knowledge with generative AI can effectively tune system to realize 3.89% to 53.2% performance improvement for various realistic applications.

Physical AI systems, particularly humanoid and collaborative robots, are increasingly deployed in safety-critical applications. However, existing authentication mechanisms do not check whether the claimed computational identity is actually tied to the specific actuator hardware that executes the motion, leaving a gap between cyber and physical domains. This paper presents a multimodal behavioral Physical Unclonable Function (PUF) architecture that combines electromechanical actuation characteristics with computational load signatures for device authentication. Our system generates challenge-response pairs by requesting robots to perform simultaneous physical motions and cryptographic operations, extracting unique signatures from motor impedance variations and processor power consumption patterns. A synchronized data acquisition module captures temporal correlations between cyber and physical domains at microsecond precision, while a fuzzy extractor based on Bose-Chaudhuri-Hocquenghem (BCH) codes ensures stable key generation despite environmental variations. Our theoretical analysis indicates that the proposed PUF can provide more than 128 bits of effective min-entropy, with an expected inter-device uniqueness of 49.2% and intra-device reliability of 94.1% under nominal conditions. In addition, the requirement to execute coupled motion and cryptographic tasks makes signal injection, pure software emulation, and straightforward replay attacks significantly harder, because an adversary must reproduce both domains and their timing relationship. The proposed method targets embodied AI platforms in which robots can physically manipulate their environment, and aims to provide an authentication mechanism that explicitly accounts for this actuation capability as part of the security boundary.