We introduce ImageWAM, a simple world-action model that repurposes pretrained image editing models for robot action prediction, using editing KV caches as compact world-action context to reduce inference cost while improving policy performance.

We introduce MaskWAM, an object-centric world-action model that unifies mask prompting and mask prediction to provide spatially grounded supervision and robust policy generalization in language-clear and language-ambiguous manipulation tasks.

We introduce Humanoid-GPT, a GPT-style Transformer with causal attention trained on a 2B-frame retargeted motion corpus for whole-body control, achieving unprecedented zero-shot generalization to unseen motions and control tasks.

We introduce LIMMT, a data-centric framework for humanoid motion tracking that curates motion data through physics feasibility, action diversity, and action complexity, showing that training on just 3% of curated data outperforms using full corpora.

We introduce VLA-JEPA, a JEPA-style pretraining framework that learns action-relevant transition semantics by predicting future latent states without pixel reconstruction or information leakage, achieving consistent gains in generalization and robustness over existing methods.

We introduce ReWorld, a framework that employs reinforcement learning to align video-based embodied world models with physical realism, task completion capability, embodiment plausibility, and visual quality through a hierarchical reward model trained on a large-scale video preference dataset.

We propose DeFI, decoupling visual forward and inverse dynamics pretraining with a General Forward Dynamics Model (GFDM) for future prediction and a General Inverse Dynamics Model (GIDM) for latent actions from video, then unified finetuning for robot manipulation.

Based on cognitive psychology, we introduce a comprehensive and complex spatial reasoning benchmark, including 50 detailed categories and 1.5K manual labeled QA pairs.

We propose DeFI, decoupling visual forward and inverse dynamics pretraining with a General Forward Dynamics Model (GFDM) for future prediction and a General Inverse Dynamics Model (GIDM) for latent actions from video, then unified finetuning for robot manipulation.

We present MM-Nav a multi-view VLA system with 360° perception. The model is trained on large-scale expert navigation data collected from multiple reinforcement learning agents, demonstrating robust generalization in complex navigation scenarios.

We recast the vision–language–action model as a perception–prediction–action model and make the model explicitly predict a compact set of dynamic, spatial and high- level semantic information, supplying concise yet comprehensive look-ahead cues for planning.

We introduce the concept of semantic orientation, representing the object orientation condition on open vocabulary language.

we propose Hybrid-depth, a novel framework that systematically integrates foundation models (CLIP and DINO) to extract visual priors and acquire sufficient contextual information for self-supervised depth estimation methods. This method achieve SOTA accuracy on KITTI and boost downstream perception.

We introduce DexVLG, a vision‑language‑grasp model trained on the 170M‑pose, 174k‑object dataset that can generate instruction‑aligned dexterous grasp poses and achieves SOTA success and part‑grasp accuracy.

We propose a framework that disentangles scene graphs into semantic components and recomposes them to achieve complex, generalizable image generation.

We propose a framework that transfers VLMs to downstream tasks by designing visual prompts from an attention perspective that reduces the transfer solution space.

We introduce CL-Dis, a closed-loop unsupervised disentanglement framework that integrates β-VAE distillation with diffusion-based feedback to learn semantically disentangled representations without labels.

We introduce HTCL, a hierarchical temporal context learning paradigm for camera-based 3D semantic scene completion.

We propose an on-board mobility prediction approach using an articulated wheeled bevameter that consists of a force-controlled arm and an instrumented bevameter (with force and vision sensors) as its end-effector.