Tag
A new Meta paper introduces ProPerMem, a year-long egocentric video memory benchmark (~618K tokens) showing that proactive AI assistants fail primarily at deciding when to speak, not what to say—frontier models like GPT-5.3-Chat score 96% with oracle evidence but drop to 15-38% when choosing their own timing.
ProPerAgent: introduces ProPerMem, a benchmark for memory-driven proactive assistance pairing egocentric video with year-long user memories, and ProPerAgent, a graph-organized memory harness with RL post-training that lifts strict accuracy from under 10% to 42.4%, showing content and timing are separately necessary for proactive AI assistants.
ScaleNav通过在互联网第一人称视频上预训练目标导向视觉动态，结合动作对齐与联合微调，实现语言条件下的局部导航，在仿真与真实四足机器人部署上取得高成功率并超越基线。
该论文提出 HumanVerse——一个500小时的开放世界第一视角人体全身操作数据集，以及基于此训练的人形基础模型 λ_0，实现了无需目标任务机器人演示的零样本迁移到 Unitree G1，在 SIMPLE 上达到 90.0% 成功率。
TouchScale 引入了一个 500 小时的人类视觉-触觉交互数据集，使用统一的可穿戴设备同步采集 egocentric RGB-D 视频、手腕视频与全手双触觉测量，用于视觉-触觉预训练和机器人操作策略学习，在零样本触觉预测和真实机器人操作成功率上均取得显著提升。
Preload (YC F26) announced synchronized egocentric video, EMG, and tactile data collection for robot manipulation, VLAs, and world models, offering end-to-end robotic data collection at scale.
EgoTools introduces a 100-hour egocentric tool-use corpus and a 1,000-question diagnostic benchmark, showing current multimodal models (e.g., Gemini-3.1-Pro at 51.7% on Perception & Grounding) still struggle with tool-centric embodied reasoning, while fine-tuning Qwen3-VL-8B on the data lifts accuracy from 50.0% to 60.9%.
MemLife introduces a multimodal memory system for long-term egocentric videos that builds entity-grounded text episodes and retrieves them with a time-indexed agentic reader, improving over training-free baselines by 4.6-12.0% on long-horizon benchmarks. A reinforcement learning framework called MemOpt further optimizes the memory writer for faithfulness and retrievability, yielding consistent 2.7-5.0% gains.
InfiniHand is an end-to-end streaming framework for world-space hand motion estimation from egocentric video, jointly estimating hand parameters, camera trajectories, and locations to reduce errors and improve performance over state-of-the-art baselines.
CapMem is a human-annotated benchmark for episodic memory in egocentric video using captions, showing that caption-based QA outperforms direct video QA on long videos.
A 2B vision-language model distilled from a tool-using agent achieves first-place performance on long egocentric video question answering by pruning its multilingual embedding table to meet parameter limits, reaching 89% accuracy of the larger pipeline with only 1.1% of parameters.
SCOUT introduces a recovery-aware agentic framework with an adaptive exploration-exploitation policy for ultra-long egocentric video reasoning, trained via UPS-GRPO, a uncertainty-prioritized RL method with turn-level advantage decomposition. It achieves state-of-the-art results on ultra-long egocentric benchmarks and remains competitive on shorter long-video settings.
Presents SERUM, a multi-pass framework that extracts structured behavioral models of user actions and intents from raw egocentric video using hierarchical VLM annotation, reducing hallucinations and producing interpretable process models without manual annotation.
ReViV is a unified framework for holistic egocentric 4D reconstruction that simultaneously reconstructs viewer (body, hand, gaze) and view (depth, camera trajectory) dynamics from a single monocular RGB video using a Masked Generative Egocentric Transformer, achieving state-of-the-art accuracy and efficiency.
Vinci2 presents a proactive assistance system for continuous egocentric videos, introducing a new benchmark EgoServe and a memory-augmented agent EgoMemo that determines when to intervene based on temporal context.
Reka AI Labs releases CS2-10k, a large dataset of over 600,000 egocentric gameplay videos with 10,000+ hours of frame-level keyboard, mouse, and 3D position data, available on Hugging Face for world models, action-conditioned video generation, and egocentric navigation research.
This paper finds that egocentric human video, when processed with a filtering and labeling pipeline, can outperform teleoperated real-robot data for pretraining embodied foundation models, achieving lower validation loss and higher success rates on real-robot tasks.
ACE-EGO-0 is a unified Vision-Language-Action pretraining framework that leverages egocentric human videos and robot trajectories via a reliability-aware training objective, achieving state-of-the-art on embodied AI benchmarks.
EgoPhys introduces a framework to construct deformable physical digital twins from egocentric RGB video using generalizable priors and a compact codebook, enabling zero-shot generalization to unseen objects without per-spring optimization. The system is demonstrated on a real robot, showing that egocentric human play video can serve as internal world representation for deformable-object planning.
This paper introduces V-RAGBench, a benchmark for evaluating retrieval-augmented generation over long egocentric videos, and CARVE, a method that adaptively selects retrieval configurations per chunk to improve VideoRAG performance.