Munari
Munari provides quality-scoring software for robotics teams collecting human demonstration data. The platform analyzes teleoperation, UMI gripper, and egocentric video sessions to deliver per-operator reliability scores, trendlines, and evidence-based issue identification, helping teams distinguish execution problems from capture hardware issues. It accepts MCAP, ROS bags, LeRobot, and raw video formats, with the first report delivered within 24 hours and no integration required.
- Robotics
- Data & Analytics
- Artificial Intelligence
- Software Only
Funding
Funding not disclosed
Founders
Founder details are not available yet.
Product
Problem
Robotics teams collecting human demonstration data for training models often cannot objectively assess the quality of individual demonstrations or reliably compare performance across multiple collectors. Without systematic scoring, issues arising from operator technique or capture hardware go unnoticed until they degrade downstream model performance.
Solution
Munari provides quality-scoring software that evaluates human demonstrations captured via teleoperation, handheld UMI grippers, and egocentric cameras. The platform analyzes each session and delivers per-operator reliability scores, trendlines that flag quality drops before bad episodes accumulate, and evidence-based issue identification. Teams can distinguish execution issues from capture problems, enabling targeted feedback for improved future collection. The service accepts MCAP, ROS bags, LeRobot, and raw video plus control log formats, with the first report delivered within 24 hours and no integration required.
Target Audience
Primary customers are robotics companies and AI research teams collecting human demonstrations to build physical AI training datasets, including those using teleoperation, UMI, or egocentric capture workflows.
Features
- Per-operator reliability scoring and 7-day trendline monitoring to catch quality degradation early
- Evidence-based issue identification that distinguishes execution problems from capture hardware issues
- Support for multiple collection methods: teleoperation, UMI handheld grippers, and egocentric video
- Ingestion of MCAP, ROS bags, LeRobot, and raw video with control logs
- No-integration workflow with reports delivered within 24 hours
- Data and operations infrastructure purpose-built for Physical AI training data pipelines