What actually breaks in field data collection
Sync drift, a loose stereo rig, and a battery that dies at hour six. The nine automated checks we run before a human ever opens an episode.
Blog
Notes from the field and the pipeline
What we learn collecting real-world demonstration data at scale — protocol design, quality failures, sensor trade-offs, and what actually happens on a factory floor at two in the afternoon.
Research
Ten thousand clean hours beat a hundred thousand noisy ones
Everyone quotes hours. Almost nobody quotes what fraction of those hours a policy can actually learn from. Here is how we measure usable yield, and why it changed how we price collection.
Read the post
Recording inside a working factory
You cannot stop the line for data collection. What we changed about our protocol after the first week on an auto-components floor in Pune.
Egocentric or third-person?
Both views answer different questions. What we capture when a customer can only afford one, and why we usually argue for both on a subset.
From a one-line task to a runnable protocol
A customer says "fold laundry". Turning that into trial structure, camera placement, edge-case coverage and an acceptance spec takes about a week.
Informed consent across twelve languages
Consent is not a signature. What it takes to genuinely explain data use to a participant who has never heard of a robotics company.
Failure is the data you are missing
Most collected datasets quietly exclude the attempts that went wrong. That removes exactly the recovery behaviour a deployed policy needs most.
Why we are building this in India
Industrial density, climatic range and urban–rural contrast inside one jurisdiction. The case for India as the physical AI data layer.
No posts in this category yet.
New posts, roughly twice a month
Protocol design, quality failures and what the field teaches us. No product announcements.