Institution profile

Horizon Robotics

Industry researchasia · cn
Official website
Research library114linked papers
Opportunities0open roles
Selected work

Representative Papers

ClearAIR: A Human-Visual-Perception-Inspired All-in-One Image Restoration

Jan 06, 2026arXiv.org

This work addresses the limitations of existing all-in-one image restoration methods, which often suffer from over-smoothing and artifacts due to their reliance on degradation-specific representations and struggle with complex real-world degradations. Inspired by human visual perception mechanisms, we propose a coarse-to-fine hierarchical restoration framework. It first leverages a multimodal large language model for cross-modal image quality assessment, then performs task-adaptive restoration through semantic-guided, region-aware degradation modeling, and finally enhances fine detail recovery via a self-supervised internal cue reuse mechanism. To our knowledge, this is the first approach to integrate human visual perception principles into all-in-one image restoration. Extensive experiments demonstrate state-of-the-art performance across multiple synthetic and real-world datasets, with significant suppression of artifacts and notable improvements in perceptual detail quality.

1 citationsRead paper

Stream Forcing: Constructing Unified Training Trajectory for Robust Streaming Video Generation

Aug 10, 2026

This work addresses the inconsistency between training and inference noise scheduling in streaming video generation by proposing a unified training framework termed Stream Forcing. The method models diffusion sampling as a frame-indexed stochastic process parameterized by noise levels, establishing a continuous training trajectory that bridges independent sampling during training and inference-consistent sampling at test time. Through a joint calibration strategy coupled with a temporally coherent sampling algorithm, the framework ensures smoothness along the training trajectory and strong inter-frame temporal consistency, effectively balancing comprehensive training coverage with efficient inference within a single architecture. Empirically, the approach reduces FVD by 36.6% on UCF-101 and achieves a 27.9% FVD improvement in zero-shot long-horizon video extrapolation tasks.

0 citationsRead paper

DreamWAM: Beyond RGB Future Prediction for World Action Models

Aug 05, 2026

This work addresses the susceptibility of existing RGB-based action models to irrelevant factors such as texture and illumination, which hinders their ability to focus on action-relevant states. To overcome this limitation, the authors propose a structured world modeling approach that jointly represents appearance, motion, geometry, and semantics. During training, a shared attention mechanism guides the action branch to learn relevant features, while at inference time, the model operates solely on RGB inputs. Built upon a shared VideoDiT–ActionDiT architecture, the method integrates an RGB-motion joint latent space denoising process and employs lightweight gated residual branches to encode geometric and semantic information. Evaluated on the LIBERO benchmark, the approach achieves success rates of 98.40% and 98.90% in no-backtrack and joint-inference settings, respectively; on the more challenging LIBERO-Plus, it attains 63.44% and 75.47%, and demonstrates a real-world robot execution success rate of 74.4%, substantially outperforming current baselines.

0 citationsRead paper

Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models

Aug 04, 2026

Existing world action models struggle to balance efficiency and robustness during inference: retaining future-aware representations incurs high computational costs, while more efficient approaches compromise temporal generalization. This work proposes a sparse future conditioning mechanism that leverages a SparseMoT module and an Interval KV-Fusion strategy to selectively reuse multi-depth future representations without increasing attention complexity, thereby avoiding redundant video-action interactions. Seamlessly integrated into the diffusion-based action denoising pipeline, the method boosts success rates on LIBERO-Plus from 49.14% to 73.57%, significantly outperforming Fast-WAM, and achieves a 2.21× faster inference speed than Joint-WAM. It further attains state-of-the-art performance across LIBERO, RoboTwin 2.0, and real-world robotic tasks.

0 citationsRead paper
Recent publications

Latest Papers

Stream Forcing: Constructing Unified Training Trajectory for Robust Streaming Video Generation

Aug 10, 2026

This work addresses the inconsistency between training and inference noise scheduling in streaming video generation by proposing a unified training framework termed Stream Forcing. The method models diffusion sampling as a frame-indexed stochastic process parameterized by noise levels, establishing a continuous training trajectory that bridges independent sampling during training and inference-consistent sampling at test time. Through a joint calibration strategy coupled with a temporally coherent sampling algorithm, the framework ensures smoothness along the training trajectory and strong inter-frame temporal consistency, effectively balancing comprehensive training coverage with efficient inference within a single architecture. Empirically, the approach reduces FVD by 36.6% on UCF-101 and achieves a 27.9% FVD improvement in zero-shot long-horizon video extrapolation tasks.

0 citationsRead paper

DreamWAM: Beyond RGB Future Prediction for World Action Models

Aug 05, 2026

This work addresses the susceptibility of existing RGB-based action models to irrelevant factors such as texture and illumination, which hinders their ability to focus on action-relevant states. To overcome this limitation, the authors propose a structured world modeling approach that jointly represents appearance, motion, geometry, and semantics. During training, a shared attention mechanism guides the action branch to learn relevant features, while at inference time, the model operates solely on RGB inputs. Built upon a shared VideoDiT–ActionDiT architecture, the method integrates an RGB-motion joint latent space denoising process and employs lightweight gated residual branches to encode geometric and semantic information. Evaluated on the LIBERO benchmark, the approach achieves success rates of 98.40% and 98.90% in no-backtrack and joint-inference settings, respectively; on the more challenging LIBERO-Plus, it attains 63.44% and 75.47%, and demonstrates a real-world robot execution success rate of 74.4%, substantially outperforming current baselines.

0 citationsRead paper

Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models

Aug 04, 2026

Existing world action models struggle to balance efficiency and robustness during inference: retaining future-aware representations incurs high computational costs, while more efficient approaches compromise temporal generalization. This work proposes a sparse future conditioning mechanism that leverages a SparseMoT module and an Interval KV-Fusion strategy to selectively reuse multi-depth future representations without increasing attention complexity, thereby avoiding redundant video-action interactions. Seamlessly integrated into the diffusion-based action denoising pipeline, the method boosts success rates on LIBERO-Plus from 49.14% to 73.57%, significantly outperforming Fast-WAM, and achieves a 2.21× faster inference speed than Joint-WAM. It further attains state-of-the-art performance across LIBERO, RoboTwin 2.0, and real-world robotic tasks.

0 citationsRead paper

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Jul 30, 2026

This work presents the first practical implementation of recursive self-improvement (RSI) in machine learning engineering, introducing a unified learn-and-evolve loop framework. The authors develop OpenMLE, a full-stack executable AI4AI system that integrates a verifiable task environment, operator learning, and long-horizon search modules. They apply execution-driven supervised fine-tuning and reinforcement learning to the Frontis-MA1 (35B) model, transforming it into a meta-evolutionary agent that performs closed-loop optimization through four atomic operations: Draft, Improve, Debug, and Crossover. Leveraging deduplicated training data and an asynchronous experience prior mechanism (OpenMLE-Evo-Max), the system achieves a Medal Average of 71.21% on MLE-Bench Lite—surpassing GPT-5.5+Codex—and attains a 70% Match-SOTA score on NatureBench Lite, demonstrating strong generalization and cross-domain capabilities.

0 citationsRead paper