Yilu Wu

Hanlin Wang, Yilu Wu, Sheng Guo et al.

In this paper, we study the problem of procedure planning in instructional videos, which aims to make a plan (i.e. a sequence of actions) given the current visual observation and the desired goal. Previous works cast this as a sequence modeling problem and leverage either intermediate visual observations or language instructions as supervision to make autoregressive planning, resulting in complex learning schemes and expensive annotation costs. To avoid intermediate supervision annotation and error accumulation caused by planning autoregressively, we propose a diffusion-based framework, coined as PDPP, to directly model the whole action sequence distribution with task label as supervision instead. Our core idea is to treat procedure planning as a distribution fitting problem under the given observations, thus transform the planning problem to a sampling process from this distribution during inference. The diffusion-based modeling approach also effectively addresses the uncertainty issue in procedure planning. Based on PDPP, we further apply joint training to our framework to generate plans with varying horizon lengths using a single model and reduce the number of training parameters required. We instantiate our PDPP with three popular diffusion models and investigate a series of condition-introducing methods in our framework, including condition embeddings, MoEs, two-stage prediction and Classifier-Free Guidance strategy. Finally, we apply our PDPP to the Visual Planners for human Assistance problem which requires the goal specified in natural language rather than visual observation. We conduct experiments on challenging datasets of different scales and our PDPP model achieves the state-of-the-art performance on multiple metrics, even compared with those strongly-supervised counterparts. These results further demonstrates the effectiveness and generalization ability of our model.

Yilu Wu, Chenhui Zhu, Shuai Wang et al.

We are committed to learning human skill generators at key-step levels. The generation of skills is a challenging endeavor, but its successful implementation could greatly facilitate human skill learning and provide more experience for embodied intelligence. Although current video generation models can synthesis simple and atomic human operations, they struggle with human skills due to their complex procedure process. Human skills involve multi-step, long-duration actions and complex scene transitions, so the existing naive auto-regressive methods for synthesizing long videos cannot generate human skills. To address this, we propose a novel task, the Key-step Skill Generation (KS-Gen), aimed at reducing the complexity of generating human skill videos. Given the initial state and a skill description, the task is to generate video clips of key steps to complete the skill, rather than a full-length video. To support this task, we introduce a carefully curated dataset and define multiple evaluation metrics to assess performance. Considering the complexity of KS-Gen, we propose a new framework for this task. First, a multimodal large language model (MLLM) generates descriptions for key steps using retrieval argument. Subsequently, we use a Key-step Image Generator (KIG) to address the discontinuity between key steps in skill videos. Finally, a video generation model uses these descriptions and key-step images to generate video clips of the key steps with high temporal consistency. We offer a detailed analysis of the results, hoping to provide more insights on human skill generation. All models and data are available at https://github.com/MCG-NJU/KS-Gen.

Yilu Wu, Hanlin Wang, Jing Wang et al.

Given the current visual observations, the traditional procedure planning task in instructional videos requires a model to generate goal-directed plans within a given action space. All previous methods for this task conduct training and inference under the same action space, and they can only plan for pre-defined events in the training set. We argue this setting is not applicable for human assistance in real lives and aim to propose a more general and practical planning paradigm. Specifically, in this paper, we introduce a new task named Open-event Procedure Planning (OEPP), which extends the traditional procedure planning to the open-event setting. OEPP aims to verify whether a planner can transfer the learned knowledge to similar events that have not been seen during training. We rebuild a new benchmark of OpenEvent for this task based on existing datasets and divide the events involved into base and novel parts. During the data collection process, we carefully ensure the transfer ability of procedural knowledge for base and novel events by evaluating the similarity between the descriptions of different event steps with multiple stages. Based on the collected data, we further propose a simple and general framework specifically designed for OEPP, and conduct extensive study with various baseline methods, providing a detailed and insightful analysis on the results for this task.

Chenhui Zhu, Yilu Wu, Shuai Wang et al.

Image-to-video generation has made remarkable progress with the advancements in diffusion models, yet generating videos with realistic motion remains highly challenging. This difficulty arises from the complexity of accurately modeling motion, which involves capturing physical constraints, object interactions, and domain-specific dynamics that are not easily generalized across diverse scenarios. To address this, we propose MotionRAG, a retrieval-augmented framework that enhances motion realism by adapting motion priors from relevant reference videos through Context-Aware Motion Adaptation (CAMA). The key technical innovations include: (i) a retrieval-based pipeline extracting high-level motion features using video encoder and specialized resamplers to distill semantic motion representations; (ii) an in-context learning approach for motion adaptation implemented through a causal transformer architecture; (iii) an attention-based motion injection adapter that seamlessly integrates transferred motion features into pretrained video diffusion models. Extensive experiments demonstrate that our method achieves significant improvements across multiple domains and various base models, all with negligible computational overhead during inference. Furthermore, our modular design enables zero-shot generalization to new domains by simply updating the retrieval database without retraining any components. This research enhances the core capability of video generation systems by enabling the effective retrieval and transfer of motion priors, facilitating the synthesis of realistic motion dynamics.