The understanding layer for Physical AI. We turn robot & AV video and trajectory data into training-ready timelines of every action, failure, and edge case. 🤖
San Francisco, CA
Joined June 2024
- Tweets47
- Following60
- Followers265
- Likes39
Pinned Tweet
The future of computer vision is agentic.
1/ We built Nomadic around a gap we kept seeing in video understanding: VLMs generate chain-of-thought that's fluent and often correct in structure, but weakly grounded in what's actually in the video.
This limitation shows up in cases like "rolling stops". Is the stop sign actually associated with the vehicle’s lane? Does the ego vehicle ever come to a full stop, or only decelerate?
Nomadic AI retweeted
"Talk to your customers" is the most repeated advice in startups. @nomadicai cofounder @mstf1bl heard it, agreed with it, and still didn't do enough of it.
"Please don't sit in silos. Please don't sit at your office and just iterate on your product without any feedback whatsoever. Do the absolute bare minimum to be able to talk to your ideal customer."
"You're almost always wrong… when you just go to the person across from you, ask them an honest question… they tell you something that's almost always different from what you thought."
Nomadic AI retweeted
The robot failures that look like successes. A box of baklava. A police car on the shoulder.
The next Study.
Coming soon.
Nomadic AI retweeted
Only @mstf1bl could get from baklava to a one-against-fifteen chess simul to GeoGuessr to the Raas he and his cofounder danced at Harvard — and make all four of them about building a company.
Next episode of The Study soon... still editing.
We're attending & presenting at #ECCV2026 in Malmö, Sweden on September 8-12!
Our Members of Technical Staff Christian (author of "Learn to Rank: Visual Attribution by Learning Importance Ranking") and Akshay, alongside our CEO Mustafa & Business Developer Atahan will be within the present to meet fellow autonomous system & physical AI researchers and developers.
See you there!
Nomadic AI retweeted
Physical AI looks different when the demos are autonomous excavators, radar-camera fusion, robot perception, and teachable general-purpose robots. Bots, Bevs & Devs was a sharp tour of what has to work outside the lab. @nomadicai @svrobo @BedrockRobotics
Our largest & busiest Physical AI event yet! ✅
Big thanks to @svrobo for co-hosting August 2026's Bots/Bevs/Devs event. We loved our speakers' content and look forward to co-hosting the next ones!
Nomadic AI retweeted
Looking forward to talking at this event with @svrobo and @nomadicai tonight!
We are hosting the most brilliant minds in robotics & physical AI with @svrobo ! Join us at our office on Wednesday at 717 Market St where brilliant entrepreneurs and operators like @kevinmpeterson1 (CTO @BedrockRobotics) , Abhishek Bhatia from @ServeRobotics and @Griffin from Zendar will showcase what's their top of mind. Come hungry for engaging content & delicious desserts! 🤖 🧠 🍰
luma.com/6h8hzcez?v=1
We are hosting the most brilliant minds in robotics & physical AI with @svrobo ! Join us at our office on Wednesday at 717 Market St where brilliant entrepreneurs and operators like @kevinmpeterson1 (CTO @BedrockRobotics) , Abhishek Bhatia from @ServeRobotics and @Griffin from Zendar will showcase what's their top of mind. Come hungry for engaging content & delicious desserts! 🤖 🧠 🍰
luma.com/6h8hzcez?v=1
In robotics this is really about distribution. A robot doing the same pick a million times in the same environment doesn't teach you good generalization
We cluster your episodes so that you can you curate better datasets
For those impacted by Amazon AGI team layoffs, Nomadic AI is hiring Members of Technical Staff to build visual AI that evaluates the next physical frontiers of AI. Please reach out!
1/ Introducing Auto Event Clustering 🧵
Nomadic now automatically clusters events from your robotics footage.
- See your dominant failure modes instantly.
- Jump from one representative event to every similar occurrence.
- Prioritize fixes by impact instead of reviewing runs one at a time.
2/ Do you know what’s actually in your robotics data?
- How much coverage do you have?
- Which environments and operating conditions are represented?
- What types of failures keep happening?
- Which behaviors are rare or missing?
- How much of your dataset is repetitive?
- Where do you need to collect more data?
Answering these questions used to mean manually reviewing thousands of runs. You no longer need this manual process
3/ Nomadic’s new Auto Event Clustering helps you explore your data as a map of behaviors.
It automatically groups similar events across video, trajectories, and logs, giving you a clear view of your dataset’s distribution, coverage, and recurring patterns in a fraction of the time.
Explore more at: nomadicai.com/solutions/robo…
1/ We made GPT-5.5, Gemini 3.1, Cosmos 3, and Qwen3-VL better at understanding robot videos.
Our agentic harness cut classification error by ≥10% on every VLM we tested.
Here's how we do it and how it ties into making automated data curation a reality 🧵
3/ Our agent then intelligently selects frames at key points along the robot trajectory to better predict these classes.
We find that feeding the correct frames into the VLMs had a lot of impact.
4/ The result is higher accuracy on every VLM we tested.
This harness is but one piece of our platform, powering automated data curation and failure monitoring for robotics teams today.
Full writeup: nomadicai.com/blog/nomadic-v…
Fine-grained subtask annotations improve VLA performance.
We trained a vision + proprioception model that’s now SOTA at generating them, generalizing across embodiments.
93.1 F1@50 on REASSEMBLE, 98.6 on @Amazon Robotics blade insertion. Details below 👇🧵
3/ The hard part is that many of these transitions happen faster than a human can comfortably label, and faster than a generic VLM can reliably interpret from sampled frames.
A gripper closes, a peg slips, a blade hits resistance, or the robot pauses and recovers. These are tiny physical events, but they often determine whether the behavior succeeded or failed.
4/ Nomadic is built for that level of structure.
We combine robot telemetry with visual context so the system can identify both the timing of each action and what the robot is interacting with.
The result is a searchable timeline of physical behavior: the kind of data layer robotics teams need for training, debugging, and evaluation.
Full benchmark:
nomadicai.com/blog/action-se…