HilbertA: Hilbert Attention for Image Generation with Diffusion Models

Designing sparse attention for diffusion transformers requires reconciling two-dimensional spatial locality with GPU efficiency, a trade-off that current methods struggle to achieve. Existing approaches enforce two-dimensional spatial locality but often incur uncoalesced memory access. We present HilbertA, a 2D-aware and GPU-efficient sparse attention mechanism. HilbertA reorders image tokens along Hilbert curves to achieve a contiguous memory layout while preserving spatial neighborhoods, and employs a sliding schedule across layers to enable long-range information propagation without repeated or uncoalesced memory access. To further enhance cross-tile communication and positional awareness, HilbertA introduces a small central shared region. Implemented in Triton, HilbertA delivers comparable image quality with significant acceleration over prior methods on Flux.1-dev, demonstrating the feasibility of hardware-aligned two-dimensional sparse attention for high-resolution image generation. HilbertA delivers attention speedups of $2.3\times$ when generating $1024\times 1024$ images, and up to $4.17\times$ at $2048\times 2048$, while achieving image quality comparable to or surpassing baselines.

Reveal and Release: Iterative LLM Unlearning with Self-generated Data

Large language model (LLM) unlearning has demonstrated effectiveness in removing the influence of undesirable data (also known as forget data). Existing approaches typically assume full access to the forget dataset, overlooking two key challenges: (1) Forget data is often privacy-sensitive, rare, or legally regulated, making it expensive or impractical to obtain (2) The distribution of available forget data may not align with how that information is represented within the model. To address these limitations, we propose a ``Reveal-and-Release'' method to unlearn with self-generated data, where we prompt the model to reveal what it knows using optimized instructions. To fully utilize the self-generated forget data, we propose an iterative unlearning framework, where we make incremental adjustments to the model's weight space with parameter-efficient modules trained on the forget data. Experimental results demonstrate that our method balances the tradeoff between forget quality and utility preservation.

Kimi K2: Open Agentic Intelligence

We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.

Mastering Agile Tasks with Limited Trials

Embodied robots nowadays can already handle many real-world manipulation tasks. However, certain other real-world tasks (e.g., shooting a basketball into a hoop) are highly agile and require high execution precision, presenting additional challenges for methods primarily designed for quasi-static manipulation tasks. This leads to increased efforts in costly data collection, laborious reward design, or complex motion planning. Such tasks, however, are far less challenging for humans. Say a novice basketball player typically needs only $\sim$10 attempts to make their first successful shot, by roughly imitating a motion prior and then iteratively adjusting their motion based on the past outcomes. Inspired by this human learning paradigm, we propose the Adaptive Diffusion Action Plannin (ADAP) algorithm, a simple & scalable approach which iteratively refines its action plan by few real-world trials within a learned prior motion pattern, until reaching a specific goal. Experiments demonstrated that ADAP can learn and accomplish a wide range of goal-conditioned agile dynamic tasks with human-level precision and efficiency directly in real-world, such as throwing a basketball into the hoop in fewer than 10 trials. Project website:https://adap-robotics.github.io/ .

EasyInsert: A Data-Efficient and Generalizable Insertion Policy

Insertion task is highly challenging that requires robots to operate with exceptional precision in cluttered environments. Existing methods often have poor generalization capabilities. They typically function in restricted and structured environments, and frequently fail when the plug and socket are far apart, when the scene is densely cluttered, or when handling novel objects. They also rely on strong assumptions such as access to CAD models or a digital twin in simulation. To address this, we propose EasyInsert, a framework which leverages the human intuition that relative pose (delta pose) between plug and socket is sufficient for successful insertion, and employs efficient and automated real-world data collection with minimal human labor to train a generalizable model for relative pose prediction. During execution, EasyInsert follows a coarse-to-fine execution procedure based on predicted delta pose, and successfully performs various insertion tasks. EasyInsert demonstrates strong zero-shot generalization capability for unseen objects in cluttered environments, handling cases with significant initial pose deviations while maintaining high sample efficiency and requiring little human effort. In real-world experiments, with just 5 hours of training data, EasyInsert achieves over 90% success in zero-shot insertion for 13 out of 15 unseen novel objects, including challenging objects like Type-C cables, HDMI cables, and Ethernet cables. Furthermore, with only one human demonstration and 4 minutes of automatically collected data for fine-tuning, it reaches over 90% success rate for all 15 objects.

SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation

Real-world tasks such as garment manipulation and table rearrangement demand robots to perform generalizable, highly precise, and long-horizon actions. Although imitation learning has proven to be an effective approach for teaching robots new skills, large amounts of expert demonstration data are still indispensible for these complex tasks, resulting in high sample complexity and costly data collection. To address this, we propose Semantic Keypoint Imitation Learning (SKIL), a framework which automatically obtain semantic keypoints with help of vision foundation models, and forms the descriptor of semantic keypoints that enables effecient imitation learning of complex robotic tasks with significantly lower sample complexity. In real world experiments, SKIL doubles the performance of baseline methods in tasks such as picking a cup or mouse, while demonstrating exceptional robustness to variations in objects, environmental changes, and distractors. For long-horizon tasks like hanging a towel on a rack where previous methods fail completely, SKIL achieves a mean success rate of 70\% with as few as 30 demonstrations. Furthermore, SKIL naturally supports cross-embodiment learning due to its semantic keypoints abstraction, our experiments demonstrate that even human videos bring considerable improvement to the learning performance. All these results demonstrate the great success of SKIL in achieving data-efficint generalizable robotic learning. Visualizations and code are available at: https://skil-robotics.github.io/SKIL-robotics/.

Navi2Gaze: Leveraging Foundation Models for Navigation and Target Gazing

Task-aware navigation continues to be a challenging area of research, especially in scenarios involving open vocabulary. Previous studies primarily focus on finding suitable locations for task completion, often overlooking the importance of the robot's pose. However, the robot's orientation is crucial for successfully completing tasks because of how objects are arranged (e.g., to open a refrigerator door). Humans intuitively navigate to objects with the right orientation using semantics and common sense. For instance, when opening a refrigerator, we naturally stand in front of it rather than to the side. Recent advances suggest that Vision-Language Models (VLMs) can provide robots with similar common sense. Therefore, we develop a VLM-driven method called Navigation-to-Gaze (Navi2Gaze) for efficient navigation and object gazing based on task descriptions. This method uses the VLM to score and select the best pose from numerous candidates automatically. In evaluations on multiple photorealistic simulation benchmarks, Navi2Gaze significantly outperforms existing approaches and precisely determines the optimal orientation relative to target objects.

CoPa: General Robotic Manipulation through Spatial Constraints of Parts with Foundation Models

Foundation models pre-trained on web-scale data are shown to encapsulate extensive world knowledge beneficial for robotic manipulation in the form of task planning. However, the actual physical implementation of these plans often relies on task-specific learning methods, which require significant data collection and struggle with generalizability. In this work, we introduce Robotic Manipulation through Spatial Constraints of Parts (CoPa), a novel framework that leverages the common sense knowledge embedded within foundation models to generate a sequence of 6-DoF end-effector poses for open-world robotic manipulation. Specifically, we decompose the manipulation process into two phases: task-oriented grasping and task-aware motion planning. In the task-oriented grasping phase, we employ foundation vision-language models (VLMs) to select the object's grasping part through a novel coarse-to-fine grounding mechanism. During the task-aware motion planning phase, VLMs are utilized again to identify the spatial geometry constraints of task-relevant object parts, which are then used to derive post-grasp poses. We also demonstrate how CoPa can be seamlessly integrated with existing robotic planning algorithms to accomplish complex, long-horizon tasks. Our comprehensive real-world experiments show that CoPa possesses a fine-grained physical understanding of scenes, capable of handling open-set instructions and objects with minimal prompt engineering and without additional training. Project page: https://copa-2024.github.io/

SpaceOctopus: An Octopus-inspired Motion Planning Framework for Multi-arm Space Robot

Space robots have played a critical role in autonomous maintenance and space junk removal. Multi-arm space robots can efficiently complete the target capture and base reorientation tasks due to their flexibility and the collaborative capabilities between the arms. However, the complex coupling properties arising from both the multiple arms and the free-floating base present challenges to the motion planning problems of multi-arm space robots. We observe that the octopus elegantly achieves similar goals when grabbing prey and escaping from danger. Inspired by the distributed control of octopuses' limbs, we develop a multi-level decentralized motion planning framework to manage the movement of different arms of space robots. This motion planning framework integrates naturally with the multi-agent reinforcement learning (MARL) paradigm. The results indicate that our method outperforms the previous method (centralized training). Leveraging the flexibility of the decentralized framework, we reassemble policies trained for different tasks, enabling the space robot to complete trajectory planning tasks while adjusting the base attitude without further learning. Furthermore, our experiments confirm the superior robustness of our method in the face of external disturbances, changing base masses, and even the failure of one arm.

EfficientZero V2: Mastering Discrete and Continuous Control with Limited Data

Sample efficiency remains a crucial challenge in applying Reinforcement Learning (RL) to real-world tasks. While recent algorithms have made significant strides in improving sample efficiency, none have achieved consistently superior performance across diverse domains. In this paper, we introduce EfficientZero V2, a general framework designed for sample-efficient RL algorithms. We have expanded the performance of EfficientZero to multiple domains, encompassing both continuous and discrete actions, as well as visual and low-dimensional inputs. With a series of improvements we propose, EfficientZero V2 outperforms the current state-of-the-art (SOTA) by a significant margin in diverse tasks under the limited data setting. EfficientZero V2 exhibits a notable advancement over the prevailing general algorithm, DreamerV3, achieving superior outcomes in 50 of 66 evaluated tasks across diverse benchmarks, such as Atari 100k, Proprio Control, and Vision Control.