Daily Papers Arch&EAI

Scaling laws assume larger models trained on more data consistently outperform smaller ones -- an assumption that drives model selection in computer vision but remains untested in resource-constrained Earth observation (EO). We conduct a systematic efficiency analysis across three scaling dimensions: model size, dataset size, and input resolution, on rooftop PV detection in Madagascar. Optimizing for model efficiency (mAP$_{50}$ per unit of model size), we find a consistent efficiency inversion: YOLO11N achieves both the highest efficiency ($24\times$ higher than YOLO11X) and the highest absolute mAP$_{50}$ (0.617). Resolution is the dominant resource allocation lever ($+$120% efficiency gain), while additional data yields negligible returns at low resolution. These findings are robust to the deployment objective: small high-resolution configurations are Pareto-dominant across all 44 setups in the joint accuracy-throughput space, leaving no tradeoff to resolve. In data-scarce EO, bigger is not just unnecessary: it can be worse.

Scaling laws assume larger models trained on more data consistently outperform smaller ones -- an assumption that drives model selection in computer vision but remains untested in resource-constrained Earth observation (EO).

The adoption of fisheye cameras in robotic manipulation, driven by their exceptionally wide Field of View (FoV), is rapidly outpacing a systematic understanding of their downstream effects on policy learning. This paper presents the first comprehensive empirical study to bridge this gap, rigorously analyzing the properties of wrist-mounted fisheye cameras for imitation learning. Through extensive experiments in both simulation and the real world, we investigate three critical research questions: spatial localization, scene generalization, and hardware generalization. Our investigation reveals that: (1) The wide FoV significantly enhances spatial localization, but this benefit is critically contingent on the visual complexity of the environment. (2) Fisheye-trained policies, while prone to overfitting in simple scenes, unlock superior scene generalization when trained with sufficient environmental diversity. (3) While naive cross-camera transfer leads to failures, we identify the root cause as scale overfitting and demonstrate that hardware generalization performance can be improved with a simple Random Scale Augmentation (RSA) strategy. Collectively, our findings provide concrete, actionable guidance for the large-scale collection and effective use of fisheye datasets in robotic learning. More results and videos are available on https://robo-fisheye.github.io/

The adoption of fisheye cameras in robotic manipulation, driven by their exceptionally wide Field of View (FoV), is rapidly outpacing a systematic understanding of their downstream effects on policy learning.

Goal-conditioned reinforcement learning has shown considerable potential in robotic manipulation; however, existing approaches remain limited by their reliance on prioritizing collected experience, resulting in suboptimal performance across diverse tasks. Inspired by human learning behaviors, we propose a more comprehensive learning paradigm, ACDC, which integrates multidimensional Adaptive Curriculum (AC) Planning with Dynamic Contrastive (DC) Control to guide the agent along a well-designed learning trajectory. More specifically, at the planning level, the AC component schedules the learning curriculum by dynamically balancing diversity-driven exploration and quality-driven exploitation based on the agent's success rate and training progress. At the control level, the DC component implements the curriculum plan through norm-constrained contrastive learning, enabling magnitude-guided experience selection aligned with the current curriculum focus. Extensive experiments on challenging robotic manipulation tasks demonstrate that ACDC consistently outperforms the state-of-the-art baselines in both sample efficiency and final task success rate.

Goal-conditioned reinforcement learning has shown considerable potential in robotic manipulation; however, existing approaches remain limited by their reliance on prioritizing collected experience, resulting in suboptimal performance across diverse tasks.

Flow-based vision-language-action (VLA) models excel in embodied control but suffer from intractable likelihoods during multi-step sampling, hindering online reinforcement learning. We propose \textbf{\textit{$\boldsymbolπ$-StepNFT}} (Step-wise Negative-aware Fine-Tuning), a critic-and-likelihood-free framework that requires only a single forward pass per optimization step and eliminates auxiliary value networks. We identify that wider exploration spaces necessitate finer-grained, step-wise guidance for alignment. Empirically, $π$-StepNFT unlocks latent potential on LIBERO with competitive few-shot robustness. Moreover, it achieves superior generalization on ManiSkill, outperforming value-based baselines in OOD scenarios by preventing overfitting to multimodal features. This property offers a scalable solution promising for complex real-world applications.

Flow-based vision-language-action (VLA) models excel in embodied control but suffer from intractable likelihoods during multi-step sampling, hindering online reinforcement learning.

Effective exploration in reinforcement learning requires not only tracking where an agent has been, but also understanding how the agent perceives and represents the world. To learn powerful representations, an agent should actively explore states that contribute to its knowledge of the environment. Temporal representations can capture the information necessary to solve a wide range of potential tasks while avoiding the computational cost associated with full state reconstruction. In this paper, we propose an exploration method that leverages temporal contrastive representations to guide exploration, prioritizing states with unpredictable future outcomes. We demonstrate that such representations can enable the learning of complex exploratory x in locomotion, manipulation, and embodied-AI tasks, revealing capabilities and behaviors that traditionally require extrinsic rewards. Unlike approaches that rely on explicit distance learning or episodic memory mechanisms (e.g., quasimetric-based methods), our method builds directly on temporal similarities, yielding a simpler yet effective strategy for exploration.

Effective exploration in reinforcement learning requires not only tracking where an agent has been, but also understanding how the agent perceives and represents the world.

Diffusion-based policies have achieved remarkable results in robotic manipulation but often struggle to adapt rapidly in dynamic scenarios, leading to delayed responses or task failures. We present DCDP, a Dynamic Closed-Loop Diffusion Policy framework that integrates chunk-based action generation with real-time correction. DCDP integrates a self-supervised dynamic feature encoder, cross-attention fusion, and an asymmetric action encoder-decoder to inject environmental dynamics before action execution, achieving real-time closed-loop action correction and enhancing the system's adaptability in dynamic scenarios. In dynamic PushT simulations, DCDP improves adaptability by 19\% without retraining while requiring only 5\% additional computation. Its modular design enables plug-and-play integration, achieving both temporal coherence and real-time responsiveness in dynamic robotic scenarios, including real-world manipulation tasks. The project page is at: https://github.com/wupengyuan/dcdp

Diffusion-based policies have achieved remarkable results in robotic manipulation but often struggle to adapt rapidly in dynamic scenarios, leading to delayed responses or task failures.

State-of-the-art text-to-image diffusion models (DMs) achieve remarkable quality, yet their massive parameter scale (8-11B) poses significant challenges for inferences on resource-constrained devices. In this paper, we present HierarchicalPrune, a novel compression framework grounded in a key observation: DM blocks exhibit distinct functional hierarchies, where early blocks establish semantic structures while later blocks handle texture refinements. HierarchicalPrune synergistically combines three techniques: (1) Hierarchical Position Pruning, which identifies and removes less essential later blocks based on position hierarchy; (2) Positional Weight Preservation, which systematically protects early model portions that are essential for semantic structural integrity; and (3) Sensitivity-Guided Distillation, which adjusts knowledge-transfer intensity based on our discovery of block-wise sensitivity variations. As a result, our framework brings billion-scale diffusion models into a range more suitable for on-device inference, while preserving the quality of the output images. Specifically, combined with INT4 weight quantisation, HierarchicalPrune achieves 77.5-80.4% memory footprint reduction (e.g., from 15.8 GB to 3.2 GB) and 27.9-38.0% latency reduction, measured on server and consumer grade GPUs, with the minimum drop of 2.6% in GenEval score and 7% in HPSv2 score compared to the original model. Finally, our comprehensive user study with 85 participants demonstrates that HierarchicalPrune maintains perceptual quality comparable to the original model while significantly outperforming prior works.

State-of-the-art text-to-image diffusion models (DMs) achieve remarkable quality, yet their massive parameter scale (8-11B) poses significant challenges for inferences on resource-constrained devices.

Reinforcement learning (RL) has become a key paradigm for training software engineering (SWE) agents, but existing pipelines typically rely on per-task containers for isolation. At scale, pre-built container images incur substantial storage overhead, slow environment setup, and require container-management privileges. We propose SWE-MiniSandbox, a lightweight, container-free method that enables scalable RL training of SWE agents without sacrificing isolation. Instead of relying on per-instance containers, SWE-MiniSandbox executes each task in an isolated workspace backed by kernel-level mechanisms, substantially reducing system overhead. It leverages lightweight environment pre-caching techniques to eliminate the need for bulky container images. As a result, our approach lowers disk usage to approximately 5\% of that required by container-based pipelines and reduces environment preparation time to about 25\% of the container baseline. Empirical results demonstrate that SWE-MiniSandbox achieves evaluation performance comparable to standard container-based pipelines. By removing the dependency on heavy container infrastructure, SWE-MiniSandbox offers a practical and accessible foundation for scaling RL-based SWE agents, particularly in resource-constrained research environments.

Reinforcement learning (RL) has become a key paradigm for training software engineering (SWE) agents, but existing pipelines typically rely on per-task containers for isolation.

Lossless compression is essential for efficient data storage and transmission. Although learning-based lossless compressors achieve strong results, most of them are designed for a single modality, leading to redundant compressor deployments in multi-modal settings. Designing a unified multi-modal compressor is critical yet challenging, as different data types vary largely in format, dimension, and statistics. Multi-modal large language models offer a promising resolution but remain too complex for practical use. Thus, we propose \textbf{OmniZip}, \textbf{a unified and lightweight lossless compressor for multi-modal data (like image, text, speech, tactile, database, and gene sequence)}. Built on a lightweight backbone, OmniZip incorporates three key components to enable efficient multi-modal lossless compression: a modality-unified tokenizer that reversibly transforms diverse data into tokens, a modality-routing context learning mechanism that enables flexible multi-modal context modeling, and a modality-routing feedforward design that further enhances the model's nonlinear representation flexibility. A reparameterization training strategy is used to enhance model capacity. OmniZip outperforms or matches other state-of-the-art compressors on multiple modalities, achieving 42\%, 57\%, 62\% and 42\%, 53\% higher compression efficiency than gzip on CLIC-M, TouchandGo, enwik9, LibriSpeech, and WikiSQL datasets, respectively. It also supports near real-time inference on resource-constrained edge devices, reaching about 1MB/s on MacBook CPUs and iPhone NPUs. Our code is released at https://github.com/adminasmi/OmniZip-CVPR2026.

Lossless compression is essential for efficient data storage and transmission.

Despite the rapid progress of Vision-Language-Action (VLA) models, the prevailing paradigm of predicting discrete waypoints remains fundamentally misaligned with the intrinsic continuity of physical motion. This discretization imposes rigid sampling rates, lacks high-order differentiability, and introduces quantization artifacts that hinder precise, compliant interaction. We propose Neural Implicit Action Fields (NIAF), a paradigm shift that reformulates action prediction from discrete waypoints to continuous action function regression. By utilizing an MLLM as a hierarchical spectral modulator over a learnable motion prior, NIAF synthesizes infinite-resolution trajectories as continuous-time manifolds. This formulation enables analytical differentiability, allowing for explicit supervision of velocity, acceleration, and jerk to ensure mathematical consistency and physical plausibility. Our approach achieves state-of-the-art results on CALVIN and LIBERO benchmarks across diverse backbones. Furthermore, real-world experiments demonstrate that NIAF enables stable impedance control, bridging the gap between high-level semantic understanding and low-level dynamic execution.

Despite the rapid progress of Vision-Language-Action (VLA) models, the prevailing paradigm of predicting discrete waypoints remains fundamentally misaligned with the intrinsic continuity of physical motion.

Agentic artificial intelligence (AI) presents a promising pathway toward realizing autonomous and self-improving wireless network services. However, resource-constrained, widely distributed, and data-heterogeneous nature of wireless networks poses significant challenges to existing agentic AI that relies on centralized architectures, leading to high communication overhead, privacy risks, and non-independent and identically distributed (non-IID) data. Federated learning (FL) has the potential to improve the overall loop of agentic AI through collaborative local learning and parameter sharing without exchanging raw data. This paper proposes new federated agentic AI approaches for wireless networks. We first summarize fundamentals of agentic AI and mainstream FL types. Then, we illustrate how each FL type can strengthen a specific component of agentic AI's loop. Moreover, we conduct a case study on using FRL to improve the performance of agentic AI's action decision in low-altitude wireless networks (LAWNs). Finally, we provide a conclusion and discuss future research directions.

Agentic artificial intelligence (AI) presents a promising pathway toward realizing autonomous and self-improving wireless network services.

Motivation: Sparse autoencoders (SAEs) decompose foundation model activations into interpretable features, but causal feature-to-feature interactions across network depth remain unknown for biological foundation models. Results: We introduce causal circuit tracing by ablating SAE features and measuring downstream responses, and apply it to Geneformer V2-316M and scGPT whole-human across four conditions (96,892 edges, 80,191 forward passes). Both models show approximately 53 percent biological coherence and 65 to 89 percent inhibitory dominance, invariant to architecture and cell type. scGPT produces stronger effects (mean absolute d = 1.40 vs. 1.05) with more balanced dynamics. Cross-model consensus yields 1,142 conserved domain pairs (10.6x enrichment, p < 0.001). Disease-associated domains are 3.59x more likely to be consensus. Gene-level CRISPRi validation shows 56.4 percent directional accuracy, confirming co-expression rather than causal encoding.

Motivation: Sparse autoencoders (SAEs) decompose foundation model activations into interpretable features, but causal feature-to-feature interactions across network depth remain unknown for biological foundation models.

Federated Learning (FL) faces major challenges in real-world deployments due to statistical heterogeneity across clients and system heterogeneity arising from resource-constrained devices. While clustering-based approaches mitigate statistical heterogeneity and pruning techniques improve memory and communication efficiency, these strategies are typically studied in isolation. We propose CA-AFP, a unified framework that jointly addresses both challenges by performing cluster-specific model pruning. In CA-AFP, clients are first grouped into clusters, and a separate model for each cluster is adaptively pruned during training. The framework introduces two key innovations: (1) a cluster-aware importance scoring mechanism that combines weight magnitude, intra-cluster coherence, and gradient consistency to identify parameters for pruning, and (2) an iterative pruning schedule that progressively removes parameters while enabling model self-healing through weight regrowth. We evaluate CA-AFP on two widely used human activity recognition benchmarks, UCI HAR and WISDM, under natural user-based federated partitions. Experimental results demonstrate that CA-AFP achieves a favorable balance between predictive accuracy, inter-client fairness, and communication efficiency. Compared to pruning-based baselines, CA-AFP consistently improves accuracy and lower performance disparity across clients with limited fine-tuning, while requiring substantially less communication than dense clustering-based methods. It also shows robustness to different Non-IID levels of data. Finally, ablation studies analyze the impact of clustering, pruning schedules and scoring mechanism offering practical insights into the design of efficient and adaptive FL systems.

Federated Learning (FL) faces major challenges in real-world deployments due to statistical heterogeneity across clients and system heterogeneity arising from resource-constrained devices.

Siamese visual trackers have recently advanced through increasingly sophisticated fusion mechanisms built on convolutional or Transformer architectures. However, both struggle to deliver pixel-level interactions efficiently on resource-constrained hardware, leading to a persistent accuracy-efficiency imbalance. Motivated by this limitation, we redesign the Siamese neck with a simple yet effective Multilayer Perception (MLP)-based fusion module that enables pixel-level interaction with minimal structural overhead. Nevertheless, naively stacking MLP blocks introduces a new challenge: computational cost can scale quadratically with channel width. To overcome this, we construct a hierarchical search space of carefully designed MLP modules and introduce a customized relaxation strategy that enables differentiable neural architecture search (DNAS) to decouple channel-width optimization from other architectural choices. This targeted decoupling automatically balances channel width and depth, yielding a low-complexity architecture. The resulting tracker achieves state-of-the-art accuracy-efficiency trade-offs. It ranks among the top performers on four general-purpose and three aerial tracking benchmarks, while maintaining real-time performance on both resource-constrained Graphics Processing Units (GPUs) and Neural Processing Units (NPUs).

Siamese visual trackers have recently advanced through increasingly sophisticated fusion mechanisms built on convolutional or Transformer architectures.

In visually ambiguous manipulation such as detecting button click tactile feedback is often the sole source of ground truth. However, fusing tactile data poses a significant challenge due to a spatiotemporal mismatch: tactile perception requires high-frequency processing with long-horizon memory (System 1), whereas visual policies operate at low control frequencies (System 2). Existing architectures struggle to bridge this gap: Transformers are computationally prohibitive for high-frequency loops (>100Hz), while LSTMs suffer from forgetting over extended interaction histories. In this paper, we introduce TacMamba, a hierarchical architecture that aligns high-bandwidth tactile reflexes with low-frequency visual planning. Our approach comprises three core contributions: (1) a custom high-frequency tactile interface designed for flexible integration; (2) a Mamba-based Tactile History Compressor that encodes continuous force history into a compact state with O(1) inference latency (0.45 ms), enabling plug-and-play fusion with VLA models without joint pre-training and (3) a Tactile-Guided Dual-Stage Training strategy that leverages temporal discrimination for self-supervised representation learning and phase-uniform sampling to mitigate data sparsity. Experiments on discrete counting and implicit state switching demonstrate that TacMamba achieves 100% success rates, significantly outperforming the visual-only pi_0.5 baseline, while strictly satisfying hard real-time constraints.

In visually ambiguous manipulation such as detecting button click tactile feedback is often the sole source of ground truth.

Vision-language-action (VLA) models have shown strong generalization for robotic action prediction through large-scale vision-language pretraining. However, most existing models rely solely on RGB cameras, limiting their perception and, consequently, manipulation capabilities. We present OmniVLA, an omni-modality VLA model that integrates novel sensing modalities for physically-grounded spatial intelligence beyond RGB perception. The core of our approach is the sensor-masked image, a unified representation that overlays spatially grounded and physically meaningful masks onto the RGB images, derived from sensors including an infrared camera, a mmWave radar, and a microphone array. This image-native unification keeps sensor input close to RGB statistics to facilitate training, provides a uniform interface across sensor hardware, and enables data-efficient learning with lightweight per-sensor projectors. Built on this, we present a multisensory vision-language-action model architecture and train the model based on an RGB-pretrained VLA backbone. We evaluate OmniVLA on challenging real-world tasks where sensor-modality perception guides the robotic manipulation. OmniVLA achieves an average task success rate of 84%, significantly outperforms both RGB-only and raw-sensor-input baseline models by 59% and 28% respectively, meanwhile showing higher learning efficiency and stronger generalization capability.

Vision-language-action (VLA) models have shown strong generalization for robotic action prediction through large-scale vision-language pretraining.

Shared L1-memory clusters of streamlined instruction processors (processing elements - PEs) are commonly used as building blocks in modern, massively parallel computing architectures (e.g. GP-GPUs). Scaling out these architectures by increasing the number of clusters incurs computational and power overhead, caused by the requirement to split and merge large data structures in chunks and move chunks across memory hierarchies via the high-latency global interconnect. Scaling up the cluster reduces buffering, copy, and synchronization overheads. However, the complexity of a fully connected cores-to-L1-memory crossbar grows quadratically with PE-count, posing a major physical implementation challenge. We present TeraPool, a physically implementable, >1000 floating-point-capable RISC-V PEs scaled-up cluster design, sharing a Multi-MegaByte >4000-banked L1 memory via a low latency hierarchical interconnect (1-7/9/11 cycles, depending on target frequency). Implemented in 12nm FinFET technology, TeraPool achieves near-gigahertz frequencies (910MHz) typical, 0.80 V/25C. The energy-efficient hierarchical PE-to-L1-memory interconnect consumes only 9-13.5pJ for memory bank accesses, just 0.74-1.1x the cost of a FP32 FMA. A high-bandwidth main memory link is designed to manage data transfers in/out of the shared L1, sustaining transfers at the full bandwidth of an HBM2E main memory. At 910MHz, the cluster delivers up to 1.89 single precision TFLOP/s peak performance and up to 200GFLOP/s/W energy efficiency (at a high IPC/PE of 0.8 on average) in benchmark kernels, demonstrating the feasibility of scaling a shared-L1 cluster to a thousand PEs, four times the PE count of the largest clusters reported in literature.

Shared L1-memory clusters of streamlined instruction processors (processing elements - PEs) are commonly used as building blocks in modern, massively parallel computing architectures (e.g.

Urban embodied AI agents, ranging from delivery robots to quadrupeds, are increasingly populating our cities, navigating chaotic streets to provide last-mile connectivity. Training such agents requires diverse, high-fidelity urban environments to scale, yet existing human-crafted or procedurally generated simulation scenes either lack scalability or fail to capture real-world complexity. We introduce UrbanVerse, a data-driven real-to-sim system that converts crowd-sourced city-tour videos into physics-aware, interactive simulation scenes. UrbanVerse consists of: (i) UrbanVerse-100K, a repository of 100k+ annotated urban 3D assets with semantic and physical attributes, and (ii) UrbanVerse-Gen, an automatic pipeline that extracts scene layouts from video and instantiates metric-scale 3D simulations using retrieved assets. Running in IsaacSim, UrbanVerse offers 160 high-quality constructed scenes from 24 countries, along with a curated benchmark of 10 artist-designed test scenes. Experiments show that UrbanVerse scenes preserve real-world semantics and layouts, achieving human-evaluated realism comparable to manually crafted scenes. In urban navigation, policies trained in UrbanVerse exhibit scaling power laws and strong generalization, improving success by +6.3% in simulation and +30.1% in zero-shot sim-to-real transfer comparing to prior methods, accomplishing a 300 m real-world mission with only two interventions.

Urban embodied AI agents, ranging from delivery robots to quadrupeds, are increasingly populating our cities, navigating chaotic streets to provide last-mile connectivity.

Due to their efficiency and small size, decision trees and random forests are popular machine learning models used for classification on resource-constrained systems. In such systems, the available execution time for inference in a random forest might not be sufficient for a complete model execution. Ideally, the already gained prediction confidence should be retained. An anytime algorithm is designed to be able to be aborted anytime, while giving a result with an increasing quality over time. Previous approaches have realized random forests as anytime algorithms on the granularity of trees, stopping after some but not all trees of a forest have been executed. However, due to the way decision trees subdivide the sample space in every step, an increase in prediction quality is achieved with every additional step in one tree. In this paper, we realize decision trees and random forest as anytime algorithms on the granularity of single steps in trees. This approach opens a design space to define the step order in a forest, which has the potential to optimize the mean accuracy. We propose the Optimal Order, which finds a step order with a maximal mean accuracy in exponential runtime and the polynomial runtime heuristics Forward Squirrel Order and Backward Squirrel Order, which greedily maximize the accuracy for each additional step taken down and up the trees, respectively. Our evaluation shows, that the Backward Squirrel Order performs $\sim94\%$ as well as the Optimal Order and $\sim99\%$ as well as all other step orders.

Due to their efficiency and small size, decision trees and random forests are popular machine learning models used for classification on resource-constrained systems.

Vision-Language-Action (VLA) models build a token-domain robot control paradigm, yet suffer from low speed. Speculative Decoding (SD) is an optimization strategy that can boost inference speed. Two key issues emerge when integrating VLA and SD: first, SD relies on re-inference to address token errors, which is computationally expensive; second, to mitigate token errors, the acceptance threshold in SD requires careful adjustment. Existing works fail to address the above two issues effectively. Meanwhile, as the bridge between AI and the physical world, existing embodied intelligence has overlooked the application of robotic kinematics. To address these issues, we innovatively combine token-domain VLA models with kinematic-domain prediction for SD, proposing a kinematic-rectified SD framework named KERV. We employ a kinematics-based Kalman Filter to predict actions and compensate for SD errors, avoiding costly re-inference. Moreover, we design a kinematics-based adjustment strategy to dynamically rectify the acceptance threshold, addressing the difficulty of threshold determination. Experimental results across diverse tasks and environments demonstrate that KERV achieves 27%~37% acceleration with nearly no Success Rate loss.

Vision-Language-Action (VLA) models build a token-domain robot control paradigm, yet suffer from low speed.

Humans learn not only how their bodies move, but also how the surrounding world responds to their actions. In contrast, while recent Vision-Language-Action (VLA) models exhibit impressive semantic understanding, they often fail to capture the spatiotemporal dynamics governing physical interaction. In this paper, we introduce Pri4R, a simple yet effective approach that endows VLA models with an implicit understanding of world dynamics by leveraging privileged 4D information during training. Specifically, Pri4R augments VLAs with a lightweight point track head that predicts 3D point tracks. By injecting VLA features into this head to jointly predict future 3D trajectories, the model learns to incorporate evolving scene geometry within its shared representation space, enabling more physically aware context for precise control. Due to its architectural simplicity, Pri4R is compatible with dominant VLA design patterns with minimal changes. During inference, we run the model using the original VLA architecture unchanged; Pri4R adds no extra inputs, outputs, or computational overhead. Across simulation and real-world evaluations, Pri4R significantly improves performance on challenging manipulation tasks, including a +10% gain on LIBERO-Long and a +40% gain on RoboCasa. We further show that 3D point track prediction is an effective supervision target for learning action-world dynamics, and validate our design choices through extensive ablations.

Humans learn not only how their bodies move, but also how the surrounding world responds to their actions.

Autonomous robots are increasingly prevalent in our society, emerging in medical care, transportation vehicles, and home assistance. These robots rely on motion planning and collision detection to identify a sequence of movements allowing them to navigate to an end goal without colliding with the surrounding environment. While many specialized accelerators have been proposed to meet the real-time requirements of robotics planning tasks, they often lack the flexibility to adapt to the rapidly changing landscape of robotics and support future advancements. However, GPUs are well-positioned for robotics and we find that they can also tackle collision detection algorithms with enhancements to existing ray tracing accelerator (RTA) units. Unlike intersection tests in ray tracing, collision queries in robotics require control flow mechanisms to avoid unnecessary computations in each query. In this work, we explore and compare different architectural modifications to address the gaps of existing GPU RTAs. Our proposed RoboGPU architecture introduces a RoboCore that computes collision queries 3.1$\times$ faster than RTA implementations and 14.8$\times$ faster than a CUDA baseline. RoboCore is also useful for other robotics tasks, achieving 3.6$\times$ speedup on a state-of-the-art neural motion planner and 1.1$\times$ speedup on Monte Carlo Localization compared to a baseline GPU. RoboGPU matches the performance of dedicated hardware accelerators while being able to adapt to evolving motion planning algorithms and support classical algorithms.

Autonomous robots are increasingly prevalent in our society, emerging in medical care, transportation vehicles, and home assistance.

Diffusion-based planners have shown strong potential for autonomous driving by capturing multi-modal driving behaviors. A key challenge is how to effectively guide these models for safe and reactive planning in closed-loop settings, where the ego vehicle's actions influence future states. Recent work leverages typical expert driving behaviors (i.e., anchors) to guide diffusion planners but relies on a truncated diffusion schedule that introduces an asymmetry between the forward and denoising processes, diverging from the core principles of diffusion models. To address this, we introduce BridgeDrive, a novel anchor-guided diffusion bridge policy for closed-loop trajectory planning. Our approach formulates planning as a diffusion bridge that directly transforms coarse anchor trajectories into refined, context-aware plans, ensuring theoretical consistency between the forward and reverse processes. BridgeDrive is compatible with efficient ODE solvers, enabling real-time deployment. We achieve state-of-the-art performance on the Bench2Drive closed-loop evaluation benchmark, improving the success rate by 7.72% over prior arts. Project page: https://github.com/shuliu-ethz/BridgeDrive.

Diffusion-based planners have shown strong potential for autonomous driving by capturing multi-modal driving behaviors.

Vision-Language-Action (VLA) models rely on current observations, including images, language instructions, and robot states, to predict actions and complete tasks. While accurate visual perception is crucial for precise action prediction and execution, recent work has attempted to further improve performance by introducing explicit reasoning during inference. However, such approaches face significant limitations. They often depend on data-intensive resources such as Chain-of-Thought (CoT) style annotations to decompose tasks into step-by-step reasoning, and in many cases require additional visual grounding annotations (e.g., bounding boxes or masks) to highlight relevant image regions. Moreover, they involve time-consuming dataset construction, labeling, and retraining, which ultimately results in longer inference sequences and reduced efficiency. To address these challenges, we propose ATA, a novel training-free framework that introduces implicit reasoning into VLA inference through complementary attention-guided and action-guided strategies. Unlike CoT or explicit visual-grounding methods, ATA formulates reasoning implicitly by integrating attention maps with an action-based region of interest (RoI), thereby adaptively refining visual inputs without requiring extra training or annotations. ATA is a plug-and-play implicit reasoning approach for VLA models, lightweight yet effective. Extensive experiments show that it consistently improves task success and robustness while preserving, and even enhancing, inference efficiency.

Vision-Language-Action (VLA) models rely on current observations, including images, language instructions, and robot states, to predict actions and complete tasks.

Recent advances in large vision-language models (VLMs) and large language models (LLMs) have enabled zero-shot approaches to visual language navigation (VLN), where an agent follows natural language instructions using only ego perception and reasoning. However, existing zero-shot methods typically construct a naive observation graph and perform per-step VLM-LLM inference on it, resulting in high latency and computation costs that limit real-time deployment. To address this, we present SFCo-Nav, an efficient zero-shot VLN framework inspired by the principle of slow-fast cognitive collaboration. SFCo-Nav integrates three key modules: 1) a slow LLM-based planner that produces a strategic chain of subgoals, each linked to an imagined object graph; 2) a fast reactive navigator for real-time object graph construction and subgoal execution; and 3) a lightweight asynchronous slow-fast bridge aligns advanced structured, attributed imagined and perceived graphs to estimate navigation confidence, triggering the slow LLM planner only when necessary. To the best of our knowledge, SFCo-Nav is the first slow-fast collaboration zero-shot VLN system supporting asynchronous LLM triggering according to the internal confidence. Evaluated on the public R2R and REVERIE benchmarks, SFCo-Nav matches or exceeds prior state-of-the-art zero-shot VLN success rates while cutting total token consumption per trajectory by over 50% and running more than 3.5 times faster. Finally, we demonstrate SFCo-Nav on a legged robot in a hotel suite, showcasing its efficiency and practicality in indoor environments.

Recent advances in large vision-language models (VLMs) and large language models (LLMs) have enabled zero-shot approaches to visual language navigation (VLN), where an agent follows natural language instructions using only ego perception and reasoning.

Large Multimodal Models (LMMs) are inherently modular, consisting of vision and audio encoders, projectors, and large language models. Yet, they are almost always executed monolithically, which underutilizes the heterogeneous accelerators (NPUs, GPUs, DSPs) in modern SoCs and leads to high end-to-end latency. In this paper, we present NANOMIND, a hardware--software co-design inference framework for Large Multimodal Models (LMMs) that breaks large models into modular ``bricks'' (vision, language, audio, etc.) and maps each to its ideal accelerator. The key insight is that large models can be broken into modular components and scheduled to run on the most appropriate compute units. It performs module-level dynamic offloading across accelerators on unified-memory SoCs. By combining customized hardware design, system-level scheduling, and optimized low-bit computation kernels, we demonstrate our framework with a compact, battery-powered device capable of running LMMs entirely on device. This prototype functions as a self-contained intelligent assistant that requires no network connectivity, while achieving higher throughput and superior power efficiency under strict resource constraints. The design further bypasses CPU bottlenecks and reduces redundant memory usage through token-aware buffer management and module-level coordination. Our system outperforms existing implementations in resource efficiency, cutting energy consumption by 42.3\% and GPU memory usage by 11.2\%. This enables a battery-powered device to run LLaVA-OneVision with a camera for nearly 20.8 hours.

Large Multimodal Models (LMMs) are inherently modular, consisting of vision and audio encoders, projectors, and large language models.

Recent advances in FlowMatching-based Vision-Language-Action (VLA) frameworks have demonstrated remarkable advantages in generating high-frequency action chunks, particularly for highly dexterous robotic manipulation tasks. Despite these notable achievements, their practical applications are constrained by prolonged generation latency, which stems from inherent iterative sampling requirements and architectural limitations. To address this critical bottleneck, we propose a Mean-Flow based One-Step VLA approach. Specifically, we resolve the noise-induced issues in the action generation process, thereby eliminating the consistency constraints inherent to conventional Flow-Matching methods. This significantly enhances generation efficiency and enables one-step action generation. Real-world robotic experiments show that the generation speed of the proposed Mean-Flow based One-Step VLA is 8.7 times and 83.9 times faster than that of SmolVLA and Diffusion Policy, respectively. These results elucidate its great potential as a high-efficiency backbone for VLA-based robotic manipulation.

Recent advances in FlowMatching-based Vision-Language-Action (VLA) frameworks have demonstrated remarkable advantages in generating high-frequency action chunks, particularly for highly dexterous robotic manipulation tasks.

Existing Vision-Language-Action (VLA) models often struggle to generalize to long-horizon tasks due to their heavy reliance on immediate observations. While recent studies incorporate retrieval mechanisms or extend context windows to handle procedural tasks, they often struggle to capture Non-Markovian dependencies, where optimal actions rely solely on specific past states rather than the current observation. To address this, we introduce Keyframe-Chaining VLA, a framework that extracts and links key historical frames to model long-horizon dependencies. Specifically, we propose an automatic keyframe selector that learns a discriminative embedding space, effectively identifying distinct state transitions. To capture task-critical information, we design a progress-aware query mechanism that dynamically retrieves historical frames based on their temporal relevance to the current execution phase. These selected keyframes are integrated into the VLA as interleaved visual tokens, explicitly grounding the policy in the long-horizon temporal context. Finally, we introduce a suite of four Non-Markovian manipulation tasks built upon the ManiSkill simulator to measure task success rates. Experimental results demonstrate that our method achieves superior performance, effectively tackling robot manipulation tasks characterized by long-horizon temporal dependencies. Code is available at https://github.com/cytoplastm/KC-VLA.

Existing Vision-Language-Action (VLA) models often struggle to generalize to long-horizon tasks due to their heavy reliance on immediate observations.

Developing generalist robots capable of mastering diverse skills remains a central challenge in embodied AI. While recent progress emphasizes scaling model parameters and offline datasets, such approaches are limited in robotics, where learning requires active interaction. We argue that effective online learning should scale the \emph{number of tasks}, rather than the number of samples per task. This regime reveals a structural advantage of model-based reinforcement learning (MBRL). Because physical dynamics are invariant across tasks, a shared world model can aggregate multi-task experience to learn robust, task-agnostic representations. In contrast, model-free methods suffer from gradient interference when tasks demand conflicting actions in similar states. Task diversity therefore acts as a regularizer for MBRL, improving dynamics learning and sample efficiency. We instantiate this idea with \textbf{EfficientZero-Multitask (EZ-M)}, a sample-efficient multi-task MBRL algorithm for online learning. Evaluated on \textbf{HumanoidBench}, a challenging whole-body control benchmark, EZ-M achieves state-of-the-art performance with significantly higher sample efficiency than strong baselines, without extreme parameter scaling. These results establish task scaling as a critical axis for scalable robotic learning. The project website is available \href{https://yewr.github.io/ez_m/}{here}.

Developing generalist robots capable of mastering diverse skills remains a central challenge in embodied AI.

Vision-Language-Action (VLA) models are emerging as a promising paradigm for end-to-end autonomous driving, valued for their potential to leverage world knowledge and reason about complex driving scenes. However, existing methods suffer from two critical limitations: a persistent misalignment between language instructions and action outputs, and the inherent inefficiency of typical auto-regressive action generation. In this paper, we introduce LinkVLA, a novel architecture that directly addresses these challenges to enhance both alignment and efficiency. First, we establish a structural link by unifying language and action tokens into a shared discrete codebook, processed within a single multi-modal model. This structurally enforces cross-modal consistency from the ground up. Second, to create a deep semantic link, we introduce an auxiliary action understanding objective that trains the model to generate descriptive captions from trajectories, fostering a bidirectional language-action mapping. Finally, we replace the slow, step-by-step generation with a two-step coarse-to-fine generation method C2F that efficiently decodes the action sequence, saving 86% inference time. Experiments on closed-loop driving benchmarks show consistent gains in instruction following accuracy and driving performance, alongside reduced inference latency.

Vision-Language-Action (VLA) models are emerging as a promising paradigm for end-to-end autonomous driving, valued for their potential to leverage world knowledge and reason about complex driving scenes.