I invite readers to write their specific use‑case questions in comments or DMs so I can help find the best possible approach for data‑quality verification.
I noted three practical pointers for building a computer‑vision AI capability: curate the most relevant data, ensure detailed annotation guidelines from the smallest most representative data, and start small by building and deploying the first ML model to a test environment with feedback collection built‑in.
I prioritize relevance over volume when curating data for vision tasks.
I create granular annotation guidelines based on the smallest, most representative data samples.
I adopt a ‘start small’ approach, deploying an initial model to a sandbox for rapid feedback.
The biggest type of data over the next couple of decades will be computer‑vision data.
If we want to tap the value of AI on it, we need to build data refineries for computer‑vision data.
I decided to adopt influence‑based sample selection rather than random or size‑based curation because it directly targets policy performance gains.
The DataMIL team includes Shivin Dass, Alaa Khaddaj, Logan Engstrom, Aleksander Madry, Andrew Ilyas, and Roberto Martín‑Martín.
I am prioritising pipelines that capture richer visual data – both 2‑D and 3‑D – with action markings, keypoint locations, and depth information to make robots safer and more useful.
so that just means you have to gather more and more visual data with 2d and 3d markings of actions and keypoints and depth to really make safer and useful robots!
I am exploring ways to scale egocentric video annotation, as expressed in the note “Egocentric Video annotation at scale!”.
I announced the initiative “Egocentric Video annotation at scale!”.
I am interested in identifying the most popular and relevant data export formats in robotics and Physical AI, as this influences how we design our data pipelines and sharing mechanisms.
A timestamp in the corner of an egocentric video shouldn't matter. Until it does. I recently pulled a frame from a 10‑minute egocentric recording—a hand working over a manufacturing line, captured on proprietary hardware—where a burned‑in timestamp overlay in the bottom‑left corner became a spurious feature, caused a train/deploy mismatch, occluded critical contact points, and polluted OCR/VLM pipelines, all without any playback errors.
To address this, I decided to build a full‑stack data supply chain that automatically detects and removes timestamp overlays, bad crops, and occluded frames, turning ad‑hoc fire‑drill fixes into a repeatable, standardized stage applied consistently across every clip in the fleet.
I link this concern to broader questions about data formats and standards raised in robotics gold rush and geopolitics about which data formats and labeling standards will serve robot learning across warehouses, farms, and hospitals.