Embodied intelligence systems, which enhance agent capabilities through continuous environment interactions, have garnered significant attention from both academia and industry. Vision-Language-Action models, inspired by advancements in large foundation models, serve as universal robotic control frameworks that substantially improve agent-environment interaction capabilities in embodied intelligence systems. This expansion has broadened application scenarios for embodied AI robots. This survey comprehensively reviews VLA models for embodied manipulation. Firstly, it chronicles the developmental trajectory of VLA architectures. Subsequently, we conduct a detailed analysis of current research across 5 critical dimensions: VLA model structures, training datasets, pre-training methods, post-training methods, and model evaluation. Finally, we synthesize key challenges in VLA development and real-world deployment, while outlining promising future research directions.

Whole-body loco-manipulation for quadruped robots with arms remains a challenging problem, particularly in achieving multi-task control. To address this, we propose MLM, a reinforcement learning framework driven by both real-world and simulation data. It enables a six-DoF robotic arm-equipped quadruped robot to perform whole-body loco-manipulation for multiple tasks autonomously or under human teleoperation. To address the problem of balancing multiple tasks during the learning of loco-manipulation, we introduce a trajectory library with an adaptive, curriculum-based sampling mechanism. This approach allows the policy to efficiently leverage real-world collected trajectories for learning multi-task loco-manipulation. To address deployment scenarios with only historical observations and to enhance the performance of policy execution across tasks with different spatial ranges, we propose a Trajectory-Velocity Prediction policy network. It predicts unobservable future trajectories and velocities. By leveraging extensive simulation data and curriculum-based rewards, our controller achieves whole-body behaviors in simulation and zero-shot transfer to real-world deployment. Ablation studies in simulation verify the necessity and effectiveness of our approach, while real-world experiments on a Go2 robot with an Airbot robotic arm demonstrate the policy's good performance in multi-task execution.

Vision-language-action models (VLAs) trained on large-scale robotic datasets have demonstrated strong performance on manipulation tasks, including bimanual tasks. However, because most public datasets focus on single-arm demonstrations, adapting VLAs for bimanual tasks typically requires substantial additional bimanual data and fine-tuning. To address this challenge, we introduce TwinVLA, a modular framework that composes two copies of a pretrained single-arm VLA into a coordinated bimanual VLA. Unlike monolithic cross-embodiment models trained on mixtures of single-arm and bimanual data, TwinVLA improves both data efficiency and performance by composing pretrained single-arm policies. Across diverse bimanual tasks in real-world and simulation settings, TwinVLA outperforms a comparably-sized monolithic RDT-1B model without requiring any bimanual pretraining. Furthermore, it narrows the gap to state-of-the-art model, $π_0$ which rely on extensive proprietary bimanual data and compute cost. These results establish our modular composition approach as a data-efficient and scalable path toward high-performance bimanual manipulation, leveraging public single-arm data.

Phys2Real is a real-to-sim-to-real pipeline for robotic manipulation that combines VLM-based physical parameter estimation with interaction-based adaptation through uncertainty-aware fusion. It comprises three stages: (I) real-to-sim: object reconstruction from segmented Gaussian Splats into simulation-ready meshes, (II) policy learning: reinforcement learning of policies conditioned on physical parameters such as the center of mass (CoM) of an object, and (III) sim-to-real transfer: uncertainty-aware fusion of VLM priors and interaction-based estimates for online adaptation. Abstract-- 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. T o 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. 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. Deploying robotic manipulation policies trained in simulation to the real world remains a fundamental challenge, especially for tasks requiring fine-grained physical dynamics. Robots must adapt to varying object properties such as friction, mass distribution, and compliance, which significantly affect manipulation outcomes but are difficult to model precisely. While learning from demonstrations has shown significant promise, it often lacks the physical grounding and reasoning needed to adapt to novel objects.

Imitation learning (IL) with generative models, such as diffusion and flow matching, has enabled robots to perform complex, long-horizon tasks. However, distribution shifts from unseen environments or compounding action errors can still cause unpredictable and unsafe behavior, leading to task failure. Early failure prediction during runtime is therefore essential for deploying robots in human-centered and safety-critical environments. We propose FIPER, a general framework for Failure Prediction at Runtime for generative IL policies that does not require failure data. FIPER identifies two key indicators of impending failure: (i) out-of-distribution (OOD) observations detected via random network distillation in the policy's embedding space, and (ii) high uncertainty in generated actions measured by a novel action-chunk entropy score. Both failure prediction scores are calibrated using a small set of successful rollouts via conformal prediction. A failure alarm is triggered when both indicators, aggregated over short time windows, exceed their thresholds. We evaluate FIPER across five simulation and real-world environments involving diverse failure modes. Our results demonstrate that FIPER better distinguishes actual failures from benign OOD situations and predicts failures more accurately and earlier than existing methods. We thus consider this work an important step towards more interpretable and safer generative robot policies. Code, data and videos are available at https://tum-lsy.github.io/fiper_website.

PhysHSI can also learn (e) stylized locomotion, such as dinosaur-like walking and high-knee stepping. Abstract-- Deploying humanoid robots to interact with real-world environments--such as carrying objects or sitting on chairs--requires generalizable, lifelike motions and robust scene perception. Although prior approaches have advanced each capability individually, combining them in a unified system is still an ongoing challenge. In this work, we present a physical-world humanoid-scene interaction system, PhysHSI, that enables humanoids to autonomously perform diverse interaction tasks while maintaining natural and lifelike behaviors. PhysHSI comprises a simulation training pipeline and a real-world deployment system. In simulation, we adopt adversarial motion prior-based policy learning to imitate natural humanoid-scene interaction data across diverse scenarios, achieving both generalization and lifelike behaviors. For real-world deployment, we introduce a coarse-to-fine object localization module that combines LiDAR and camera inputs to provide continuous and robust scene perception. Imagine deploying humanoid robots in everyday environments--carrying boxes into diverse places or sitting naturally on a chair. Building such a humanoid-scene interaction (HSI) system is considered more sophisticated than executing whole-body skills such as standing up [1, 2], dancing [3, 4], or performing agile motions [5-7].

The top row shows the starting object configuration with its goal rendered as a transparent overlay, while the bottom row shows the result after the robot's motion. GD2P synthesizes diverse hand poses conditioned on object geometry, handling flat (left), volumetric (middle), and tall (right) objects. Grey arrows represent the transporting direction, whereas white volumetric dots mark the estimated fingertip contact with the object. Abstract-- Nonprehensile manipulation, such as pushing and pulling, enables robots to move, align, or reposition objects that may be difficult to grasp due to their geometry, size, or relationship to the robot or the environment. Much of the existing work in nonprehensile manipulation relies on parallel-jaw grippers or tools such as rods and spatulas. In contrast, multi-fingered dexterous hands offer richer contact modes and versatility for handling diverse objects to provide stable support over the objects, which compensates for the difficulty of modeling the dynamics of nonprehensile manipulation. Therefore, we propose G eometry-aware D exterous P ushing and P ulling (GD2P) for nonprehensile manipulation with dexterous robotic hands. We study pushing and pulling by framing the problem as synthesizing and learning pre-contact dexterous hand poses that lead to effective manipulation. We generate diverse hand poses via contact-guided sampling, filter them using physics simulation, and train a diffusion model conditioned on object geometry to predict viable poses. At test time, we sample hand poses and use standard motion planners to select and execute pushing and pulling actions. GSS holds concurrent appointments as a Professor at USC and as an Amazon Scholar. This paper describes work performed at USC and is not associated with Amazon. All authors are with the Thomas Lord Department of Computer Science at the University of Southern California, USA.

Imitation learning has shown promise in learning from large-scale real-world datasets. However, pretrained policies usually perform poorly without sufficient in-domain data. Besides, human-collected demonstrations entail substantial labour and tend to encompass mixed-quality data and redundant information. As a workaround, human-in-the-loop systems gather domain-specific data for policy post-training, and exploit closed-loop policy feedback to offer informative guidance, but usually require full-time human surveillance during policy rollout. In this work, we devise ARMADA, a multi-robot deployment and adaptation system with human-in-the-loop shared control, featuring an autonomous online failure detection method named FLOAT. Thanks to FLOAT, ARMADA enables paralleled policy rollout and requests human intervention only when necessary, significantly reducing reliance on human supervision. Hence, ARMADA enables efficient acquisition of in-domain data, and leads to more scalable deployment and faster adaptation to new scenarios. We evaluate the performance of ARMADA on four real-world tasks. FLOAT achieves nearly 95% accuracy on average, surpassing prior state-of-the-art failure detection approaches by over 20%. Besides, ARMADA manifests more than 4$\times$ increase in success rate and greater than 2$\times$ reduction in human intervention rate over multiple rounds of policy rollout and post-training, compared to previous human-in-the-loop learning methods.

We study view-invariant imitation learning by explicitly conditioning policies on camera extrinsics. Using Plucker embeddings of per-pixel rays, we show that conditioning on extrinsics significantly improves generalization across viewpoints for standard behavior cloning policies, including ACT, Diffusion Policy, and SmolVLA. To evaluate policy robustness under realistic viewpoint shifts, we introduce six manipulation tasks in RoboSuite and ManiSkill that pair "fixed" and "randomized" scene variants, decoupling background cues from camera pose. Our analysis reveals that policies without extrinsics often infer camera pose using visual cues from static backgrounds in fixed scenes; this shortcut collapses when workspace geometry or camera placement shifts. Conditioning on extrinsics restores performance and yields robust RGB-only control without depth. We release the tasks, demonstrations, and code at https://ripl.github.io/know_your_camera/ .

Non-prehensile manipulation, encompassing ungraspable actions such as pushing, poking, and pivoting, represents a critical yet underexplored domain in robotics due to its contact-rich and analytically intractable nature. In this work, we revisit this problem from two novel perspectives. First, we move beyond the usual single-arm setup and the strong assumption of favorable external dexterity such as walls, ramps, or edges. Instead, we advocate a generalizable dual-arm configuration and establish a suite of Bimanual Non-prehensile Manipulation Primitives (BiNoMaP). Second, we depart from the prevailing RL-based paradigm and propose a three-stage, RL-free framework to learn non-prehensile skills. Specifically, we begin by extracting bimanual hand motion trajectories from video demonstrations. Due to visual inaccuracies and morphological gaps, these coarse trajectories are difficult to transfer directly to robotic end-effectors. To address this, we propose a geometry-aware post-optimization algorithm that refines raw motions into executable manipulation primitives that conform to specific motion patterns. Beyond instance-level reproduction, we further enable category-level generalization by parameterizing the learned primitives with object-relevant geometric attributes, particularly size, resulting in adaptable and general parameterized manipulation primitives. We validate BiNoMaP across a range of representative bimanual tasks and diverse object categories, demonstrating its effectiveness, efficiency, versatility, and superior generalization capability. Non-prehensile manipulation refers to a class of robotic actions that do not rely on firm grasping but instead leverage physical interactions such as poking, or pivoting, or pushing to achieve manipulation goals Zhou et al. (2019); Hogan & Rodriguez (2020); Sun et al. (2020); Zhou & Held (2023); Zhang et al. (2023). These skills are not merely complementary to traditional grasp-based tasks; they are often essential in scenarios where grasping is physically infeasible or inefficient. In dual-arm robotic systems Liu et al. (2022); Wu & Kruse (2024); Y amada et al. (2025); Lu et al. (2025), non-prehensile manipulation becomes especially relevant when dealing with objects that are too fragile, too flat, or lack sufficient geometry for reliable grasping. Despite its importance, current non-prehensile manipulation faces two core bottlenecks.