EgoActionBench is a new benchmark for dense action captioning in egocentric videos, featuring 1,407 episodes and 41,132 fine‑grained action annotations that specify which hand (left, right, or both) performs each temporally localized atomic action. The benchmark introduces an evaluation protocol that assesses semantic fidelity, temporal overlap, action precision, and recall, thereby measuring what happens, when it happens, and which hand is involved. Twenty baseline models were evaluated, and two backbones (Qwen3‑VL‑4B and LLaVA‑OneVision‑2‑8B) were fine‑tuned and reinforced, achieving significant score improvements and outperforming all open‑source baselines.
arXiv:2603. 09731v3 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) are increasingly considered as a foundation for embodied agents, yet it remains unclear whether they can reliably reason about the long-term physical consequences of actions from an egocentric viewpoint.
arXiv:2610.33685v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models are bottlenecked by the scarcity of robot data. Egocentric human manipulation videos offer an abundant, diverse,...
The paper introduces P³-WAM, a unified world action model for egocentric dexterous manipulation that integrates 3D perception, action-conditioned prediction, and goal-directed planning. It operates in three modes—Perceive, Predict, and Plan—using a mixture-of-transformers architecture that shares base weights and adapts to modality-specific inputs. Experiments demonstrate that unified training improves performance across all three capabilities, outperforming specialized baselines on 3D hand reconstruction and action-conditioned video prediction while producing more instruction‑adherent manipulation videos.
arXiv:2606. 00054v1 Announce Type: cross Abstract: Recent progress in generalizable embodied control has been driven by large-scale pretraining of Vision-Language-Action (VLA) models.
AtomEgo investigates how to integrate large-scale egocentric human interaction data into embodied foundation model pre‑training. The study uses a curated 2,659‑hour corpus and a scalable data pipeline to evaluate three co‑training paradigms across vision‑language‑action and world‑action architectures. Results show that the benefit of egocentric data depends on both its scale and the quality of alignment with robotic embodiment, offering practical guidance for scalable ego‑robot pre‑training.
arXiv:2610.31845v1 Announce Type: cross Abstract: Scaling policy learning requires broad experience and reliable supervision. Robot demonstrations provide high-quality trajectory supervision but are...
LAPT-Scale introduces a latent action pre‑training framework that leverages large‑scale egocentric human videos to train embodied robotic policies. The method comprises LAPT‑Scale‑LAM, which learns continuous latent actions by modeling visual transitions in human videos, and LAPT‑Scale‑Policy, which aligns these latent actions with robot embodiment and fine‑tunes for specific tasks. Using 23.3K hours of video, the approach improves real‑robot success rates from 63.3% to 85.0% and generalizes to unseen visual conditions.
The paper introduces GALA, a Geometry-Aware Latent Action modeling framework that enhances image-based latent actions with 3D end‑effector motion. It proposes the Unified End‑effector Motion Representation (UEMR) to preserve fine‑grained motion while improving cross‑embodiment generalizability. Experiments show GALA effectively models generalizable fine‑grained motions across embodiments, achieving high success rates in RoboCasa-GR1 and real‑world tasks.
Video Prediction Policy 2 (VPP2) is a new world action model that improves zero‑shot generalization for both video prediction and action generation. It achieves this by pretraining a large, diverse manipulation video dataset with event‑level supervision, then distilling the model into a single‑step visual planner and adding a mixture‑of‑transformers action module. Experiments show VPP2 outperforms leading baselines on open‑ended video prediction, real‑world zero‑shot manipulation, and several challenging benchmarks.
ZimaBlue is a scalable framework that learns generalizable World Action Models (WAMs) from large-scale egocentric videos. It follows a three-stage curriculum: causal video pre‑training, video‑action mid‑training with a unified action representation, and final specialization to a target robot. The system employs an asynchronous Slow‑Fast architecture to enable real‑time 30 Hz action prediction, achieving a jump in real‑robot zero‑shot success from 36.1% to 77.8% when leveraging over 120,000 hours of embodied video.
EgoJEPA is a three‑stage framework that learns embodiment‑agnostic representations from egocentric human videos, eliminating the need for hand‑crafted retargeting mappings. It first pre‑trains latent actions via visual prediction conditioned on hand‑motion differences, then models future visual dynamics and latent actions jointly, and finally adapts to specific robots with lightweight action adapters. Experiments on five manipulation tasks across three robot embodiments show a 74% average success rate versus 65% for a retargeting baseline, while cutting cross‑embodiment wall‑clock cost by 47.2%.
Zero-shot cross-task generalization, where a policy must execute manipulation tasks never seen during training, remains a central challenge in robot learning. In large language models, a novel task ca...