Training generalist policies for robotic manipulation has shown great promise, as they enable language-conditioned, multi-task behaviors across diverse scenarios. However, evaluating these policies remains difficult because real-world testing is expensive, time-consuming, and labor-intensive. It also requires frequent environment resets and carries safety risks when deploying unproven policies on physical robots. Manually creating and populating simulation environments with assets for robotic manipulation has not addressed these issues, primarily due to the significant engineering effort required and the substantial sim-to-real gap, both in terms of physics and rendering. In this paper, we explore the use of action-conditional video generation models as a scalable way to learn world models for policy evaluation. We demonstrate how to incorporate action conditioning into existing pre-trained video generation models. This allows leveraging internet-scale in-the-wild online videos during the pre-training stage and alleviates the need for a large dataset of paired video-action data, which is expensive to collect for robotic manipulation. Our paper examines the effect of dataset diversity, pre-trained weights, and common failure cases for the proposed evaluation pipeline. Our experiments demonstrate that across various metrics, including policy ranking and the correlation between actual policy values and predicted policy values, these models offer a promising approach for evaluating policies without requiring real-world interactions.

We propose a framework for automated policy evaluation using Cosmos-Surg-dVRK, a Cosmos world foundation model (WFM) finetune, to perform simulated surgical policy rollouts and subsequent automated success rate evaluation using a video classifier. After rollout completion, the generated video is automatically labeled for task success or failure using a trained video classifier, enabling objective selection of the most promising surgical policies for real-robot evaluation and deployment. The rise of surgical robots and vision-language-action models has accelerated the development of autonomous surgical policies and efficient assessment strategies. However, evaluating these policies directly on physical robotic platforms such as the da Vinci Research Kit (dVRK) remains hindered by high costs, time demands, reproducibility challenges, and variability in execution. World foundation models (WFM) for physical AI offer a transformative approach to simulate complex real-world surgical tasks, such as soft tissue deformation, with high fidelity. This work introduces Cosmos-Surg-dVRK, a surgical finetune of the Cosmos WFM, which, together with a trained video classifier, enables fully automated online evaluation and benchmarking of surgical policies. On tabletop suture pad tasks, the automated pipeline achieves strong correlation between online rollouts in Cosmos-Surg-dVRK and policy outcomes on the real dVRK Si platform, as well as good agreement between human labelers and the V -JEP A 2-derived video classifier. Additionally, preliminary experiments with ex-vivo porcine cholecystectomy tasks in Cosmos-Surg-dVRK demonstrate promising alignment with real-world evaluations, highlighting the platform's potential for more complex surgical procedures. World models have emerged as a foundational approach for enabling intelligent agents to understand and act in complex simulated environments. Building on early work by Ha & Schmidhuber (2018), they learn compact latent representations of spatial and temporal dynamics and predict the consequences of actions in context. Leveraging diffusion processes, recent efforts have introduced large-scale multi-modal world foundation models (WFMs) (Wan et al., 2025; Agarwal et al., 2025; Kong et al., 2024; Xie et al., 2025; Ball et al., 2025). These video generative models serve as scalable, general-purpose learned simulators that encode and synthesize diverse physical phenomena, approximate scene dynamics, render plausible sensory observations, and facilitate policy evaluation and training, enabling progress on sim-to-real transfer. At the intersection of these advances lies the burgeoning field of physical AI, AI systems equipped with sensors and actuators that enable perception, reasoning, and actuation in the physical world.

Large scale robot learning has recently shown promise in enabling robots to perform complex tasks by integrating perception, control, and optionally, language understanding into a unified framework. However, they continue to struggle with long-horizon, contact-rich manipulation tasks, such as the handling of deformable objects, where supervision from demonstrations is often inconsistent in quality. In such settings, reward modeling offers a natural solution: by providing grounded progress signals, it can transform noisy demonstrations into stable supervision that generalizes across diverse trajectories. In this work, we introduce a stage-aware, video-based reward modeling framework that jointly predicts the high-level task stage and fine-grained progress within each stage. Reward labels are automatically derived from natural language subtask annotations, enabling consistent progress estimation across variable-length and heterogeneous demonstrations. This design overcomes the limitations of frame-index-based labeling, which collapses in long, variable-duration tasks such as folding a T -shirt. Our reward model demonstrates robustness to demonstration variability, generalization to out-of-distribution scenarios, and strong utility for downstream policy training. Building upon this reward model, we propose the Reward-Aligned Behavior Cloning (RA-BC) framework, which selectively filters high-quality data and reweights training samples according to reward estimates. Extensive experiments demonstrate that the reward model outperforms baselines on out-of-distribution real robot policy rollouts and human demonstration validation. Our approach achieves 83% success on folding T -shirts from the flattened state and 67% from the crumpled state--dramatically surpassing vanilla behavior cloning, which attains only 8% and 0% success under the same training dataset, respectively. Overall, our results highlight reward modeling as a key enabler for scalable, annotation-efficient, and robust imitation learning in long-horizon robotic manipulation. The long-standing vision of enabling robots to seamlessly assist humans in household chores has inspired decades of research in robotics. From tidying living spaces to preparing meals, such capabilities hold the promise of freeing up human time, and improving quality of life.

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.

Humans can perform various combinations of physical skills without having to relearn skills from scratch every single time. For example, we can swing a bat when walking without having to re-learn such a policy from scratch by composing the individual skills of walking and bat swinging. Enabling robots to combine or compose skills is essential so they can learn novel skills and tasks faster with fewer real world samples. To this end, we propose a novel compositional approach called DSE- Diffusion Score Equilibrium that enables few-shot learning for novel skills by utilizing a combination of base policy priors. Our method is based on probabilistically composing diffusion policies to better model the few-shot demonstration data-distribution than any individual policy. Our goal here is to learn robot motions few-shot and not necessarily goal oriented trajectories. Unfortunately we lack a general purpose metric to evaluate the error between a skill or motion and the provided demonstrations. Hence, we propose a probabilistic measure - Maximum Mean Discrepancy on the Forward Kinematics Kernel (MMD-FK), that is task and action space agnostic. By using our few-shot learning approach DSE, we show that we are able to achieve a reduction of over 30% in MMD-FK across skills and number of demonstrations. Moreover, we show the utility of our approach through real world experiments by teaching novel trajectories to a robot in 5 demonstrations.

This paper proposes a framework for MAP-estimation based imitation learning, which can be seen as adding to the standard supervised learning loss a cost of deviating from the observed state distribution. The key idea for optimizing such a loss is a gradient expression for the stationary distribution of a given policy that can be estimated online using policy rollouts. This idea is an extension of a previous result by Morimura (2010). I find the approach original, and potentially interesting to the NIPS community. The consequences of deviating from the demonstrated states in imitation learning have been recognized earlier, but this paper proposes a novel approach to this problem.

With the rise of stochastic generative models in robot policy learning, end-to-end visuomotor policies are increasingly successful at solving complex tasks by learning from human demonstrations. Nevertheless, since real-world evaluation costs afford users only a small number of policy rollouts, it remains a challenge to accurately gauge the performance of such policies. This is exacerbated by distribution shifts causing unpredictable changes in performance during deployment. To rigorously evaluate behavior cloning policies, we present a framework that provides a tight lower-bound on robot performance in an arbitrary environment, using a minimal number of experimental policy rollouts. Notably, by applying the standard stochastic ordering to robot performance distributions, we provide a worst-case bound on the entire distribution of performance (via bounds on the cumulative distribution function) for a given task. We build upon established statistical results to ensure that the bounds hold with a user-specified confidence level and tightness, and are constructed from as few policy rollouts as possible. In experiments we evaluate policies for visuomotor manipulation in both simulation and hardware. Specifically, we (i) empirically validate the guarantees of the bounds in simulated manipulation settings, (ii) find the degree to which a learned policy deployed on hardware generalizes to new real-world environments, and (iii) rigorously compare two policies tested in out-of-distribution settings. Our experimental data, code, and implementation of confidence bounds are open-source.

There have recently been large advances both in pre-training visual representations for robotic control and segmenting unknown category objects in general images. To leverage these for improved robot learning, we propose $\textbf{POCR}$, a new framework for building pre-trained object-centric representations for robotic control. Building on theories of "what-where" representations in psychology and computer vision, we use segmentations from a pre-trained model to stably locate across timesteps, various entities in the scene, capturing "where" information. To each such segmented entity, we apply other pre-trained models that build vector descriptions suitable for robotic control tasks, thus capturing "what" the entity is. Thus, our pre-trained object-centric representations for control are constructed by appropriately combining the outputs of off-the-shelf pre-trained models, with no new training. On various simulated and real robotic tasks, we show that imitation policies for robotic manipulators trained on POCR achieve better performance and systematic generalization than state of the art pre-trained representations for robotics, as well as prior object-centric representations that are typically trained from scratch.

Decentralized policy optimization has been commonly used in cooperative multi-agent tasks. However, since all agents are updating their policies simultaneously, from the perspective of individual agents, the environment is non-stationary, resulting in it being hard to guarantee monotonic policy improvement. To help the policy improvement be stable and monotonic, we propose model-based decentralized policy optimization (MDPO), which incorporates a latent variable function to help construct the transition and reward function from an individual perspective. We theoretically analyze that the policy optimization of MDPO is more stable than model-free decentralized policy optimization. Moreover, due to non-stationarity, the latent variable function is varying and hard to be modeled. We further propose a latent variable prediction method to reduce the error of the latent variable function, which theoretically contributes to the monotonic policy improvement. Empirically, MDPO can indeed obtain superior performance than model-free decentralized policy optimization in a variety of cooperative multi-agent tasks.

Recent progress in end-to-end Imitation Learning approaches has shown promising results and generalization capabilities on mobile manipulation tasks. Such models are seeing increasing deployment in real-world settings, where scaling up requires robots to be able to operate with high autonomy, i.e. requiring as little human supervision as possible. In order to avoid the need for one-on-one human supervision, robots need to be able to detect and prevent policy failures ahead of time, and ask for help, allowing a remote operator to supervise multiple robots and help when needed. However, the black-box nature of end-to-end Imitation Learning models such as Behavioral Cloning, as well as the lack of an explicit state-value representation, make it difficult to predict failures. To this end, we introduce Behavioral Cloning Value Approximation (BCVA), an approach to learning a state value function based on and trained jointly with a Behavioral Cloning policy that can be used to predict failures. We demonstrate the effectiveness of BCVA by applying it to the challenging mobile manipulation task of latched-door opening, showing that we can identify failure scenarios with with 86% precision and 81% recall, evaluated on over 2000 real world runs, improving upon the baseline of simple failure classification by 10 percentage-points.