https://chaoyue0307.github.io/awesome-egocentric-atlas/feed.xml 2026-07-20T00:00:00Z Recently added egocentric AI resources, newest first. https://huggingface.co/datasets/yyyyywv/egocentric 2026-07-01T00:00:00Z

About 32.8 GB and 1,019 robot episodes totaling 994,459 frames across seven LeRobot task groups: 800 gripper episodes plus plate, shoe, tea, wash, gift-in-hand, and flower-picking sets, with eye or head and wrist cameras, state, and action streams at 20-25 fps

https://arxiv.org/abs/2607.11643 2026-07-01T00:00:00Z

38B unified embodied-synthesis model spanning multi-view scene generation, embodiment transfer, editing, and video generation; reports improving pi0.5 out-of-distribution manipulation success from 36.9% to 63.2%

https://worldscape-moe.com/ 2026-07-01T00:00:00Z

Mixture-of-experts diffusion world model unifying camera trajectories, robot actions, and hand-joint action maps across locomotion, manipulation, and egocentric hand-control experiments

https://arxiv.org/abs/2607.11689 2026-07-01T00:00:00Z

Roadmap organizing WAM gaps across model roles and representations, objectives and standardization, and system composition, then proposing embodied brains, physical harnesses, shared contracts, and closed-loop post-training

https://jacobchalk.github.io/Whareformer/ 2026-07-01T00:00:00Z

First learning-based solution for online Out of Sight, Not out of Mind tracking, trained on 56 videos and evaluated on 260 long videos from EPIC-KITCHENS-100, IT3DEgo, and HD-EPIC with updatable appearance and 3D-location memory

https://doi.org/10.1038/s41597-026-07657-7 2026-07-01T00:00:00Z

Wearable motion-capture dataset for gait analysis using IMUs and shank-mounted egocentric cameras

https://wanda.lecar-lab.org/ 2026-07-01T00:00:00Z

Synthetic-data engine expanding one real demonstration per task into 16,922 LeRobot episodes and about 26.8M timesteps (roughly 248 hours at 30 fps) across five long-horizon mobile-manipulation tasks and diverse generated worlds

https://arxiv.org/abs/2607.06988 2026-07-01T00:00:00Z

Test-time training framework that steers frozen world-action models from unlabeled raw human videos by adapting a lightweight memory through self-supervised video prediction and paired human-robot meta-training

https://liujiahao2077.github.io/WALA.github.io 2026-07-01T00:00:00Z

Learns executable latent actions jointly from action-labeled demonstrations and action-free human or robot videos using DINOv3 feature deltas, dense depth, and a trainable latent world model; reports 75.2% RoboCasa success

https://arxiv.org/abs/2607.12356 2026-07-01T00:00:00Z

Grounds geometry- and semantics-aware 3D Gaussian primitives into compact VLA context tokens; reports 99% token reduction, +22.8% success over seven real tasks, and +30.0% over VLA-Adapter on OOD tasks

https://sitonggong.github.io/EgoServe-page/ 2026-07-01T00:00:00Z

3,000+ proactive-service instances across 11 service categories and four temporal memory horizons, with the training-free EgoMemo agent combining multi-scale summaries, a semantic graph, and visual retrieval archives

https://github.com/CeeZh/VTS 2026-07-01T00:00:00Z

Self-correcting temporal-tree search agent for grounded long-video QA, with SFT and RL code, an 8B checkpoint, and released trajectories/evaluation data; gains 7.4 T-F1 on Haystack-Ego4D

https://umishra.me/temporal-ratio/ 2026-07-01T00:00:00Z

Systematic video-action model study introducing Temporal Ratio to measure reliance on future latent rollouts, plus inference-time adaptive guidance that narrows compositional OOD gaps on LIBERO and real-world robot tasks

https://huggingface.co/VerboseTechLabs 2026-07-01T00:00:00Z

Six public industrial sample repositories totaling 53 first-person clips, 17,513 seconds (4.86 hours), and about 17.0 GB across textile, metal, electronics assembly, skilled commercial work, and manufacturing-unit operations

https://huggingface.co/datasets/VerboseTechLabs/household-cleaning-egocentric 2026-07-01T00:00:00Z

12 head- or chest-mounted 1080p/30 fps cleaning clips totaling 6,259 seconds (104.3 minutes), distributed as an ungated 5.74 GB ZIP with audio and clip metadata across three sub-activities

https://huggingface.co/datasets/VerboseTechLabs/cooking-chopping-egocentric 2026-07-01T00:00:00Z

11 head- or chest-mounted kitchen clips totaling 6,104 seconds (101.7 minutes) across cooking and chopping, distributed as an ungated 6.09 GB ZIP with clip-level metadata

https://arxiv.org/abs/2607.14514 2026-07-01T00:00:00Z

Training-free cross-episode object-goal navigation with persistent room- and object-level visual-topological memory, coarse-to-fine retrieval, and a conservative execution guard, evaluated on HM3D v0.1/v0.2 and MP3D

https://vsi-super-wild.github.io/ 2026-07-01T00:00:00Z

6,980 human-verified spatial QA pairs over 442 continuous panoramic in-the-wild videos totaling 284.52 hours across eight scene categories, including recordings longer than four hours and four observer-object-environment task families

https://arxiv.org/abs/2607.01804 2026-07-01T00:00:00Z

Lightweight detect-and-correct inference layer for action-chunked VLA policies, using visual feature evolution to detect open-loop deviations and adapt action horizons

https://arxiv.org/abs/2607.06706 2026-07-01T00:00:00Z

Survey of 183 VLA contributions from 2017-2026 across architectures, training recipes, action representations, bimanual coordination, UAV navigation/control, language grounding, memory, and world models

https://github.com/MCG-NJU/VIABench 2026-07-01T00:00:00Z

First-person video benchmark recorded or shared by blind and visually impaired people, evaluating proactive navigation reminders, visual question answering, and vision-guided interaction in online and offline settings

https://arxiv.org/abs/2607.08489 2026-07-01T00:00:00Z

Training-free Video caption Evaluation via GAze Score metric plus an evaluation dataset of egocentric activities and instructional slides paired with synchronized gaze and reference captions; the abstract does not report the dataset size

https://huggingface.co/datasets/Grably/usa-egocentric 2026-07-01T00:00:00Z

Three iPhone kitchen-manipulation episodes with normalized RGB videos, raw videos, camera calibration and trajectory files, MediaPipe and HaMeR hand outputs, action segments, contact/object tracks on the strongest run, review manifests, and quality reports

https://arxiv.org/abs/2607.13621 2026-07-01T00:00:00Z

Unified language-guided human seeking and persistent following benchmark requiring semantic exploration, delayed identity grounding, phase switching, and recovery in single- and multi-person environments

https://arxiv.org/abs/2607.12571 2026-07-01T00:00:00Z

Retraining-free defense that detects abnormal VLA evidence evolution, localizes compact visual-trigger regions by counterfactual score drop, and recovers observations through localized inpainting

https://arxiv.org/abs/2607.07287 2026-07-01T00:00:00Z

Predictive-and-reactive tactile foundation model for dexterous manipulation that separates vision-language planning, tactile world-model prediction, visuo-tactile goal-conditioned action generation, and high-frequency tactile residual refinement

https://arxiv.org/abs/2607.15621 2026-07-01T00:00:00Z

Asynchronous driving VLA with a frozen 7B slow reasoner and fast action expert, producing fresh control every 50 ms and raising CARLA LangAuto-Short route completion from 37.0 to 94.0

https://arxiv.org/abs/2607.02322 2026-07-01T00:00:00Z

Hybrid dynamic data-collection strategy for VLA spatial generalization, using a dual-arm setup where one arm manipulates while the other acts as a moving environmental camera

https://bigai-dex.github.io/blog/teledexter/ 2026-07-01T00:00:00Z

Hand-object co-tracking controller trained with single-stage RL and transferred zero-shot across two dexterous hands; reports 75.2% average success over seven in-hand and tool-use tasks and uses 50 demonstrations per task for behavior cloning

https://tactidex.github.io/ 2026-07-01T00:00:00Z

Real-world human-demonstration benchmark synchronizing whole-hand tactile pressure, multi-granularity hand kinematics, object 6D states, language descriptions, and task phases, with TactiSkill transfer across single-hand and bimanual tasks