Tag
RL can assist robot data collection during demonstrations through a technique called Data Inoculation, improving data quality for both behavior cloning and RL, as shown in a new paper on stacking cubes.
Axonnic 推出前臂肌肉感应袖套，通过 EMG、视频与手部姿态数据构建机器人操作训练数据集，用于缝纫、工具操作和精密装配等任务。
Marko Kraemer endorses Raphael and the team at Alexandria Inc, a new tool for building internal human 'datalabs' to capture knowledge at scale and enable the transition from human to superintelligence (SI) workforce.
Maxinsights announced a $30M seed round led by Peak XV Partners for a platform that captures real-world human work via camera-wearing operators to generate organized training datasets for robotic models. Two new products were also announced alongside the funding.
A practitioner describes building a Claude-based crypto news trading agent that skips almost everything, scores its refusals against a random baseline, and rewrites the prompt to force low-confidence actions for data collection.
A parent reports that the AI assistant Tomo asked their daughter about their black cat despite never being told about it, raising concerns about whether the tool scraped her social media data. Tomo admitted it could not rule out obtaining information from her social media accounts, prompting privacy worries.
Google is reportedly linking users' phone numbers across all signed-in accounts without clear disclosure, raising privacy concerns about cross-account data tracking and consolidation.
Construction workers are using head-mounted cameras to capture first-person footage, which helps train AI systems and robots to learn how skilled humans perform real-world tasks.
WindBorne Systems won a U.S. Air Force contract worth up to $49 million for atmospheric data collected by its fleet of ~700 autonomous weather balloons, which also feeds its WeatherMesh AI forecasting model. The deal was awarded without competitive bidding and marks the startup's largest government agreement to date.
A tweet speculating that all past iMessages could eventually be harvested by personal AI assistants to mimic users' speaking styles, and then sold as data, raising privacy concerns.
Pantheon describes how it built a highly diverse Universal Manipulation Interface (UMI) robotics dataset in eight weeks with 90 operators, custom in-house hardware, and VLM-based freeform annotation, producing over 1 million unique tasks at $10/hour and releasing a 100-hour annotated sample.
A user questions how AI assistants could apparently know personal details, such as the name of a newly adopted cat (Julius), despite claims of not accessing phone data, messages, or microphone. The post highlights concerns about AI data collection and personal information leakage.
A tweet introducing three open-source tools—Agent-Reach, Patchright Enhanced, and Scrapling—that let AI agents scrape data from sites like X, YouTube, and Reddit without needing API access.
Researchers extracted Meta's Muse AI assistant's internal instructions, revealing it creates detailed hourly-updated profiles on friends and family, raising privacy concerns about Meta's access to social data.
Omar (@omarsar0) discusses Kled V3, an AI data-collection platform that lets labs order custom datasets (image, video, audio, text, annotation) from a network of 500,000+ opt-in contributors within 72 hours, enabling tight identify-failure→collect-data→retrain loops.
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.
The post argues that AI companies' rush into physical gadgets like Meta's Muse Charm pendant and OpenAI's screenless device with Jony Ive may be driven less by better UX than by a need for fresh, real-world human behavior data as the open web gets polluted by AI-generated content.
AI agents can now scrape data from nearly the entire internet, including X, YouTube, Reddit, forums, and even websites without APIs, powered by 3 newly released open-source tools on GitHub for research, analysis, and monitoring.
The article satirizes the performative and repetitive AI projects posted on LinkedIn, particularly computer vision demos, and describes the creation of a simple detector to identify such posts.
Meta AI on Facebook is building detailed profiles of children and families from historical posts, raising privacy concerns and prompting Meta to fix the issue after user complaints.