Voxel51
End-to-end multimodal data platform for physical AI
115 followers
End-to-end multimodal data platform for physical AI
115 followers
Stop wasting training runs on bad data. Search, visualize, curate, and label the multimodal data behind your physical AI models. Voxel51 turns raw recordings into the data your models actually need.
This is the 3rd launch from Voxel51. View more
Voxel51 Multimodal
Launched this week
The real world is messy. Robots encounter unexpected situations and behave in ways you never anticipated. But every failure is a chance to teach your robot something new.
The hard part? Sifting through hours of video and sensor data to figure out why your robot failed. Voxel51 is an end-to-end multimodal data platform that accelerates robot learning. Find the failures that matter, understand why they happened, and fill gaps in your training data to improve models faster.
Try Voxel51 for free.
Free
Launch Team
Hey Product Hunt 👋
I’m Brian, Co-Founder and CEO of Voxel51.
Robotics teams need to iterate faster. But too much time goes into preparing data, and too much compute goes into training on episodes that don’t improve downstream behaviors.
We built Voxel51 Multimodal to make robot learning faster. It’s an end-to-end multimodal data platform that helps teams turn real-world failures into better training data and smarter robots.
💬 Here’s how it works:
Find the failures that matter: Upload your robot episodes and let Voxel51 index your data. Ask the Voxel51 Agent questions like “Where is task success rate lowest?” or “Find episodes where the robot failed to complete a grasp.”
Understand what went wrong: Visualize synchronized camera feeds, robot states, actions, and sensor data side-by-side. Replay episodes to see exactly what happened before, during, and after a failure.
Build better training datasets: Curate the most valuable episodes, surface edge case scenarios, and prioritize the data your robot needs to learn from.
Train, evaluate, repeat: Evaluate your next policy, see where performance improved (or didn't), and feed those insights back into your next training run.
🚀 Why Voxel51
Learn faster: Accelerate your data flywheel, save 5+ months of development and iterate on data up to 50× faster.
Make training runs count: Focus your time and compute on data that can make a difference.
Keep everything in one place: Search, investigation, curation, and model evaluation, with a system of record.
Improve model performance: Customers report increases of up to 30%.
Better data. Smarter robots. Repeat. Try Voxel51 on your data for free.
What’s the hardest part of figuring out why your robot failed? We’d love to hear what your team is running into.
@brian_e_moore Yes Brian! Such a powerful release for Physical AI builders....
Digging through house long recordings to find that 45-second clip that I need.
@mt_whirled ^now trivial with auto-indexing in Voxel51!
@mt_whirled This is a problem everyone has Monica. Completely! Finding the right data matters.
Voxel51 has been a great way to build a data flywheel for visual AI, excited to see it applied to robotics now!
@eric_hofesmann absolutely! It natively supports MCAP, LeRobot, URDF, and more
Garbage in, garbage out is especially true for physical AI.
@layla_yun1 yep, data is the new gold
See my review below, but basically it allows you to quickly build intuition about your datasets, keep them organized, and select the most unique images/videos/rosbags to train/eval/test on.
I used it for some customer demos/proof of concept work regarding robot perception. You can very quickly bootstrap a model and dataset. It makes for great reports and summaries to non-engineer types, too.
@j_a_duncan thanks so much for sharing your use cases!
This is honestly one of the very best tools for finding and correcting anomalies in your training data.
@timmmmmm definitely, just upload your data and open the Embeddings panel
It's like ripping off a blindfold and actually being able to explore what's in my sensor data.