VLA 研究日報VLA 研究日报

Vision-Language-Action (VLA) models enable robots to perform manipulation tasks directly from natural language instructions and are increasingly viewed as a foundation for generalist robotic policies. However, their reliability under Out-of-Distribution (OOD) instructions remains underexplored. In this paper, we reveal a critical failure mode in which VLA policies continue executing visually plausible actions even when the language instruction contradicts the scene. We refer to this phenomenon as linguistic blindness, where VLA policies prioritize visual priors over instruction semantics during action generation. To systematically analyze this issue, we introduce ICBench, a diagnostic benchmark constructed from the LIBERO dataset that probes language-action coupling by injecting controlled OOD instruction contradictions while keeping the visual environment unchanged. Evaluations on three representative VLA architectures, including Pi0, Pi0.5 and OpenVLA OFT, show that these models frequently succeed at tasks despite logically impossible instructions, revealing a strong visual bias in action generation. To mitigate this issue, we propose Instruction-Guided Attention Recalibration (IGAR), a train-free inference-time mechanism that rebalances attention distributions to restore the influence of language instructions. IGAR operates without retraining or architectural modification and can be directly applied to existing VLA models. Experiments across 30 LIBERO tasks demonstrate that IGAR substantially reduces erroneous execution under OOD contradictory instructions while preserving baseline task performance. We additionally validate the approach on a real Franka robotic arm, where IGAR effectively prevents manipulation triggered by inconsistent instructions.

[清华]

Vision-Language-Action (VLA) models like OpenVLA demonstrate impressive zero-shot generalization across robotic manipulation tasks but struggle to adapt to specific deployment environments where consistent high performance on a limited set of tasks is more valuable than broad generalization. We present EXPierence replayed, REtrieval augmented, Specialized VLA (ExpReS-VLA), a method that enables rapid on-device adaptation of pre-trained VLAs to target domains while preventing catastrophic forgetting through compressed experience replay and retrieval-augmented generation. Our approach maintains a memory-efficient buffer by storing extracted embeddings from OpenVLA's frozen vision backbone, reducing storage requirements by 97% compared to raw image-action pairs. During deployment, ExpReS-VLA retrieves the $k$ most similar past experiences using cosine similarity to augment training batches, while a prioritized experience replay buffer preserves recently successful trajectories. To leverage failed attempts, we introduce Thresholded Hybrid Contrastive Loss (THCL), enabling the model to learn from both successful and unsuccessful demonstrations. Experiments on the LIBERO benchmark show improvements from 82.6% to 93.1% on spatial reasoning and 61% to 72.3% on long-horizon tasks over base OpenVLA, with gains across architectures including $π_0$ (+3.2 points) and OpenVLA-OFT (+1.7 points). Physical robot experiments across five tasks demonstrate 98% success on both in-distribution and out-of-distribution conditions, improving from 84.7% and 32% respectively for naive fine-tuning. Adaptation completes in 31 seconds using 12 demonstrations on a single RTX 5090.

[CMU]

Learning robotic manipulation policies directly in the real world can be expensive and time-consuming. While reinforcement learning (RL) policies trained in simulation present a scalable alternative, effective sim-to-real transfer remains challenging, particularly for tasks that require precise dynamics. To address this, we propose Phys2Real, a real-to-sim-to-real RL pipeline that combines vision-language model (VLM)-inferred physical parameter estimates with interactive adaptation through uncertainty-aware fusion. Our approach consists of three core components: (1) high-fidelity geometric reconstruction with 3D Gaussian splatting, (2) VLM-inferred prior distributions over physical parameters, and (3) online physical parameter estimation from interaction data. Phys2Real conditions policies on interpretable physical parameters, refining VLM predictions with online estimates via ensemble-based uncertainty quantification. On planar pushing tasks of a T-block with varying center of mass (CoM) and a hammer with an off-center mass distribution, Phys2Real achieves substantial improvements over a domain randomization baseline: 100% vs 79% success rate for the bottom-weighted T-block, 57% vs 23% in the challenging top-weighted T-block, and 15% faster average task completion for hammer pushing. Ablation studies indicate that the combination of VLM and interaction information is essential for success. Project website: https://phys2real.github.io/.

[Stanford]

Vision-Language Navigation (VLN) enables robots to follow natural-language instructions in visually grounded environments, serving as a key capability for embodied robotic systems. Recent Vision-Language-Action (VLA) models have demonstrated strong navigation performance, but their high computational cost introduces latency that limits real-time deployment. We propose a training-free spatio-temporal vision token pruning framework tailored to VLA-based VLN. We apply spatial token selection to the current view, alongside spatio-temporal compression for historical memories, enabling efficient long-horizon inference while reducing redundant computation. Leveraging attention-based token importance and query-guided spatio-temporal filtering, the proposed approach preserves navigation-relevant information without retraining or modifying pretrained models, allowing plug-and-play integration into existing VLA systems. Through experiments on standard VLN benchmarks, we confirm that our method significantly outperforms existing pruning strategies. It successfully preserves superior navigation accuracy under extreme pruning scenarios, all while maintaining the highly competitive inference efficiency. Real-world deployment on a Unitree Go2 quadruped robot further validates reliable and low-latency instruction-following navigation under practical robotic constraints. We hope this work helps bridge the gap between large-scale multimodal modeling and efficient, real-time embodied deployment in robotic navigation systems.

[UMD]

Despite remarkable progress in Vision-Language-Action models (VLAs) for robot manipulation, these large pre-trained models require fine-tuning to be deployed in specific environments. These fine-tuned models are highly sensitive to camera viewpoint changes that frequently occur in unstructured environments. In this paper, we propose a zero-shot camera adaptation framework without additional demonstration data, policy fine-tuning, or architectural modification. Our key idea is to virtually adjust test-time camera observations to match the training camera configuration in real-time. For that, we use a recent feed-forward novel view synthesis model which outputs high-quality target view images, handling both extrinsic and intrinsic parameters. This plug-and-play approach preserves the pre-trained capabilities of VLAs and applies to any RGB-based policy. Through extensive experiments on the LIBERO benchmark, our method consistently outperforms baselines that use data augmentation for policy fine-tuning or additional 3D-aware features for visual input. We further validate that our approach constantly enhances viewpoint robustness in real-world robotic manipulation scenarios, including settings with varying camera extrinsics, intrinsics, and freely moving handheld cameras.

[KAIST]

Video generative models (VGMs) pretrained on large-scale internet data can produce temporally coherent rollout videos that capture rich object dynamics, offering a compelling foundation for zero-shot robotic manipulation. However, VGMs often produce physically implausible rollouts, and converting their pixel-space motion into robot actions through geometric retargeting further introduces cumulative errors from imperfect depth estimation and keypoint tracking. To address these challenges, we present \method{}, a data-free framework that aligns VGM outputs with compositional constraints generated by vision-language models (VLMs) at inference time. The key insight is that VLMs offer a capability complementary to VGMs: structured spatial reasoning that can identify the physical constraints critical to the success and safety of manipulation execution. Given a language instruction, \method{} uses a VLM to automatically extract a set of compositional constraints capturing task-specific requirements, which are then applied at two stages: (1) constraint-guided rollout selection, which scores and filters a batch of VGM rollouts to retain the most physically plausible candidate, and (2) constraint-based trajectory optimization, which uses the selected rollout as initialization and refines the robot trajectory under the same constraint set to correct retargeting errors. We evaluate \method{} on six real-robot manipulation tasks requiring precise, constraint-sensitive execution, improving the overall success rate by 43.3\% points over the strongest baseline without any task-specific training data.

[Northwestern]

Situated reasoning often relies on active exploration, yet in many real-world scenarios such exploration is infeasible due to physical constraints of robots or safety concerns of visually impaired users. Given only a limited observation, can an agent mentally simulate a future trajectory toward a target situation and answer spatial what-if questions? We introduce WanderDream, the first large-scale dataset designed for the emulative simulation of mental exploration, enabling models to reason without active exploration. WanderDream-Gen comprises 15.8K panoramic videos across 1,088 real scenes from HM3D, ScanNet++, and real-world captures, depicting imagined trajectories from current viewpoints to target situations. WanderDream-QA contains 158K question-answer pairs, covering starting states, paths, and end states along each trajectory to comprehensively evaluate exploration-based reasoning. Extensive experiments with world models and MLLMs demonstrate (1) that mental exploration is essential for situated reasoning, (2) that world models achieve compelling performance on WanderDream-Gen, (3) that imagination substantially facilitates reasoning on WanderDream-QA, and (4) that WanderDream data exhibit remarkable transferability to real-world scenarios. The source code and all data will be released.

[清华]

Diffusion-based world models have shown strong potential for unified world simulation, but the iterative denoising remains too costly for interactive use and long-horizon rollouts. While feature caching can accelerate inference without training, we find that policies designed for single-modal diffusion transfer poorly to world models due to two world-model-specific obstacles: \emph{token heterogeneity} from multi-modal coupling and spatial variation, and \emph{non-uniform temporal dynamics} where a small set of hard tokens drives error growth, making uniform skipping either unstable or overly conservative. We propose \textbf{WorldCache}, a caching framework tailored to diffusion world models. We introduce \textit{Curvature-guided Heterogeneous Token Prediction}, which uses a physics-grounded curvature score to estimate token predictability and applies a Hermite-guided damped predictor for chaotic tokens with abrupt direction changes. We also design \textit{Chaotic-prioritized Adaptive Skipping}, which accumulates a curvature-normalized, dimensionless drift signal and recomputes only when bottleneck tokens begin to drift. Experiments on diffusion world models show that WorldCache delivers up to \textbf{3.7$\times$} end-to-end speedups while maintaining \textbf{98\%} rollout quality, demonstrating the vast advantages and practicality of WorldCache in resource-constrained scenarios. Our code is released in https://github.com/FofGofx/WorldCache.

[清华]

Deriving compact and temporally aware visual representations from dynamic scenes is essential for successful execution of sequential scene understanding tasks such as visual tracking and robotic manipulation. In this paper, we introduce Token Bottleneck (ToBo), a simple yet intuitive self-supervised learning pipeline that squeezes a scene into a bottleneck token and predicts the subsequent scene using minimal patches as hints. The ToBo pipeline facilitates the learning of sequential scene representations by conservatively encoding the reference scene into a compact bottleneck token during the squeeze step. In the reconstruction step, we guide the model to capture temporal dynamics by predicting the target scene using the bottleneck token along with few target patches as hints. This design encourages the vision backbone to embed temporal dependencies, thereby enabling understanding of dynamic transitions across scenes. Extensive experiments in diverse sequential tasks, including video label propagation and robot manipulation in simulated environments demonstrate the superiority of \ours~over baselines. Moreover, deploying our pre-trained model on physical robots confirms its robustness and effectiveness in real-world environments. We further validate the scalability of ToBo across different model scales. Code is available at https://github.com/naver-ai/tobo.

[NAVER]

Recent progress in Vision-Language-Action (VLA) models has enabled embodied agents to interpret multimodal instructions and perform complex tasks. However, existing VLAs are mostly confined to short-horizon, table-top manipulation, lacking the memory and reasoning capability required for mobile manipulation, where agents must coordinate navigation and manipulation under changing spatial contexts. In this work, we present EchoVLA, a memory-aware VLA model for mobile manipulation. EchoVLA incorporates a synergistic declarative memory inspired by the human brain, consisting of a scene memory that maintains a collection of spatial-semantic maps and an episodic memory that stores task-level experiences with multimodal contextual features. The two memories are individually stored, updated, and retrieved based on current observations, task history, and instructions, and their retrieved representations are fused via coarse- and fine-grained attention to guide base-arm diffusion policies. To support large-scale training, we further introduce MoMani, an automated benchmark that generates expert-level trajectories through multimodal large language model (MLLM)-guided planning and feedback-driven refinement, supplemented with real-robot demonstrations. Comprehensive simulated and real-world results demonstrate that EchoVLA substantially improves overall performance, e.g., it achieves the highest success rates of 0.52 on manipulation/navigation tasks and 0.31 on mobile manipulation tasks in simulation, exceeding the strong baseline $π_{0.5}$ by +0.20 and +0.11, respectively.

[浙大]

Underwater robotic manipulation remains challenging because lighting variation, color attenuation, scattering, and reduced visibility can severely degrade visuomotor policies. We present Bi-AQUA, the first underwater bilateral control-based imitation learning framework for robot arms that explicitly models lighting within the policy. Bi-AQUA integrates transformer-based bilateral action chunking with a hierarchical lighting-aware design composed of a label-free Lighting Encoder, FiLM-based visual feature modulation, and a lighting token for action conditioning. This design enables adaptation to static and dynamically changing underwater illumination while preserving the force-sensitive advantages of bilateral control, which are particularly important in long-horizon and contact-rich manipulation. Real-world experiments on underwater pick-and-place, drawer closing, and peg extraction tasks show that Bi-AQUA outperforms a bilateral baseline without lighting modeling and achieves robust performance under seen, unseen, and changing lighting conditions. These results highlight the importance of combining explicit lighting modeling with force-aware bilateral imitation learning for reliable underwater manipulation. For additional material, please check: https://mertcookimg.github.io/bi-aqua

[东大]

Multimodal audiovisual perception can enable new avenues for robotic manipulation, from better material classification to the imitation of demonstrations for which only audio signals are available (e.g., playing a tune by ear). However, to unlock such multimodal potential, robots need to learn the correlations between an object's visual appearance and the sound it generates when they interact with it. Such an active sensorimotor experience requires new interaction capabilities, representations, and exploration methods to guide the robot in efficiently building increasingly rich audiovisual knowledge. In this work, we present CAVER, a novel robot that builds and utilizes rich audiovisual representations of objects. CAVER includes three novel contributions: 1) a novel 3D printed end-effector, attachable to parallel grippers, that excites objects' audio responses, 2) an audiovisual representation that combines local and global appearance information with sound features, and 3) an exploration algorithm that uses and builds the audiovisual representation in a curiosity-driven manner that prioritizes interacting with high uncertainty objects to obtain good coverage of surprising audio with fewer interactions. We demonstrate that CAVER builds rich representations in different scenarios more efficiently than several exploration baselines, and that the learned audiovisual representation leads to significant improvements in material classification and the imitation of audio-only human demonstrations. https://caver-bot.github.io/

[CMU]

High-quality, long-horizon demonstrations are essential for embodied AI, yet acquiring such data for tightly coupled wheeled mobile manipulators remains a fundamental bottleneck. Unlike fixed-base systems, mobile manipulators require continuous coordination between $SE(2)$ locomotion and precise manipulation, exposing limitations in existing teleoperation and wearable interfaces. We present \textbf{SuperSuit}, a bimodal data acquisition framework that supports both robot-in-the-loop teleoperation and active demonstration under a shared kinematic interface. Both modalities produce structurally identical joint-space trajectories, enabling direct data mixing without modifying downstream policies. For locomotion, SuperSuit maps natural human stepping to continuous planar base velocities, eliminating discrete command switches. For manipulation, it employs a strictly isomorphic wearable arm in both modes, while policy training is formulated in a shift-invariant delta-joint representation to mitigate calibration offsets and structural compliance without inverse kinematics. Real-world experiments on long-horizon mobile manipulation tasks show 2.6$\times$ higher demonstration throughput in active mode compared to a teleoperation baseline, comparable policy performance when substituting teleoperation data with active demonstrations at fixed dataset size, and monotonic performance improvement as active data volume increases. These results indicate that consistent kinematic representations across collection modalities enable scalable data acquisition for long-horizon mobile manipulation.

[清华]

Accurate physics simulation is essential for robotic learning and control, yet analytical simulators often fail to capture complex contact dynamics, while learning-based simulators typically require large amounts of costly real-world data. To bridge this gap, we propose a few-shot real-to-sim approach that combines the physical consistency of analytical formulations with the representational capacity of graph neural network (GNN)-based models. Using only a small amount of real-world data, our method calibrates analytical simulators to generate large-scale synthetic datasets that capture diverse contact interactions. On this foundation, we introduce a mesh-based GNN that implicitly models rigid-body forward dynamics and derive surrogate gradients for collision detection, achieving full differentiability. Experimental results demonstrate that our approach enables learning-based simulators to outperform differentiable baselines in replicating real-world trajectories. In addition, the differentiable design supports gradient-based optimization, which we validate through simulation-based policy learning in multi-object interaction scenarios. Extensive experiments show that our framework not only improves simulation fidelity with minimal supervision but also increases the efficiency of policy learning. Taken together, these findings suggest that differentiable simulation with few-shot real-world grounding provides a powerful direction for advancing future robotic manipulation and control.

[HKUST]

Current Vision-Language-Action (VLA) models rely primarily on RGB perception, preventing them from capturing modalities such as thermal signals that are imperceptible to conventional visual sensors. Moreover, end-to-end generative policies lack explicit safety constraints, making them fragile when encountering obstacles and novel scenarios outside the training distribution. To address these limitations, we propose Safe-Night VLA, a multimodal manipulation framework that enables robots to see the unseen while enforcing rigorous safety constraints for thermal-aware manipulation in unstructured environments. Specifically, Safe-Night VLA integrates long-wave infrared thermal perception into a pre-trained vision-language backbone, enabling semantic reasoning grounded in thermodynamic properties. To ensure safe execution under out-of-distribution conditions, we incorporate a safety filter via control barrier functions, which provide deterministic workspace constraint enforcement during policy execution. We validate our framework through real-world experiments on a Franka manipulator, introducing a novel evaluation paradigm featuring temperature-conditioned manipulation, subsurface target localization, and reflection disambiguation, while maintaining constrained execution at inference time. Results demonstrate that Safe-Night VLA outperforms RGB-only baselines and provide empirical evidence that foundation models can effectively leverage non-visible physical modalities for robust manipulation.

[清华]

This paper presents PRISM: an instruction-conditioned refinement method for imitation policies in robotic manipulation. This approach bridges Imitation Learning (IL) and Reinforcement Learning (RL) frameworks into a seamless pipeline, such that an imitation policy on a broad generic task, generated from a set of user-guided demonstrations, can be refined through reinforcement to generate new unseen fine-grain behaviours. The refinement process follows the Eureka paradigm, where reward functions for RL are iteratively generated from an initial natural-language task description. Presented approach, builds on top of this mechanism to adapt a refined IL policy of a generic task to new goal configurations and the introduction of constraints by adding also human feedback correction on intermediate rollouts, enabling policy reusability and therefore data efficiency. Results for a pick-and-place task in a simulated scenario show that proposed method outperforms policies without human feedback, improving robustness on deployment and reducing computational burden.

[UPC]