Search MCAP and LeRobot

Ask the Voxel51 Agent for the episodes you need, map them by embedding to find the outliers, and let SAM 2 track objects through driving clips, all in your browser.

What you can do

  • Open MCAP and LeRobot episodes with every camera on one timeline
  • Ask the Voxel51 Agent for “episodes where the robot picks up clothes” and review what it flags
  • Plot every episode by embedding. The outliers are the ones to watch.
  • Found one bad episode? Similarity search pulls the rest like it.
  • One click, one SAM 2 mask. One prompt, drafted boxes to review.
  • Your whole team in one workspace, free
We’ll email you a one-time code. No password, no credit card.
Work or .edu email. Personal email addresses aren’t accepted.

Demo datasets

Try Voxel51 before you upload anything

No data handy? Add any of 20+ demo datasets, including 1,000 DROID episodes as MCAP, 2,378 LeRobot Community episodes, and PandaSet driving data. Each one you add uses one of your 3 dataset slots.
Demo datasets open in your free workspace. Enter your work email and we’ll send a one-time code.

You can’t grep an MCAP. But you can search it.

1,000 DROID episodes as MCAP, three cameras synced on one timeline, in your browser. Ask the Voxel51 Agent for episodes where the robot is picking up clothes, then review the episodes it flags.
Natural language search with the Voxel51 Agent to find DROID episodes where the robot is picking up clothes.

Most demos are never watched.

2,378 episodes from 117 community LeRobot datasets on one embeddings plot. Click a cluster to see the outliers, then pull every episode like the bad one with similarity search.
3D embeddings plot of LeRobot Community dataset, as visualized in Voxel51

Click once. Then let SAM 2 track it.

103 PandaSet driving clips, and frames don’t count as samples. One click makes a SAM 2 mask, and tracking carries it to the end of the clip and back. You fix the drift.
SAM 2 video labeling in Voxel51, with a pedestrian labeled and tracked across video frames.

The Voxel51 Agent drafts. You decide.

9,888 labeled PandaSet driving images. Search “cyclist at dusk” to find the frames, then prompt “cyclists” in Agentic Labeling and the Voxel51 Agent drafts the boxes for you to review.
agentically-labeled cyclists at dusk, created using the Voxel51 Agent annotation capabilities

What’s free with Voxel51

  • 3 datasets, 10,000 samples each, including any demo datasets you add
  • Video frames don’t count as samples
  • 20GB of hosted storage
  • 30,000 model inference requests and 30 hours of compute a month
  • The Voxel51 Agent and SAM 2 included
  • Unlimited teammates, with limits shared across your workspace
  • 20+ demo datasets to add from the app

Voxel51 free plan: frequently asked questions

Can I upload MCAP files to Voxel51?

Yes. Voxel51 opens MCAP files, the default recording format for ROS 2, in your browser and plays every camera stream back on one timeline.