Rerun released version 0.35 of its robotics visualization and logging platform this week, targeting performance and reliability for data‑heavy workflows. The update introduces windowed MCAP conversion that processes large robot log files in time‑bounded segments instead of loading entire recordings into memory.
On a 20 GB test file, this approach reportedly cut peak memory usage from about 26 GB to 1.4 GB and reduced processing time from 14.3 seconds to 5.8 seconds. The converter can also recover corrupted MCAP files by rebuilding missing summaries and indexes, aiming to keep partially damaged datasets usable.
Rerun 0.35 adds an Hdf5Reader that streams HDF5 files as lazy chunk streams, improving handling of large and diverse datasets. The release also introduces preliminary support for LeRobot v3 datasets, further integrating the platform into modern robotics machine learning pipelines.
The company expanded its 3D visualization capabilities with a new VoxelGridMap archetype for sparse voxel grids, supporting anisotropic voxel sizes, pose offsets, and per‑voxel colors or values. MCAP importer improvements automatically convert certain dense voxel messages, including nav2_msgs/VoxelGrid and Foxglove VoxelGrid, into the new sparse format.
Experimental gamepad support for 3D views enables analog navigation with variable speed in large scenes, enhancing usability for developers working with complex spatial data. In parallel, Rerun is promoting a hands‑on robotics machine learning course and new agent skills tooling to support end‑to‑end data collection, refinement, and training.
Collectively, these enhancements reinforce Rerun’s positioning in robotics data infrastructure by improving performance, data recovery, interoperability, and ML tooling. The week’s developments suggest continued maturation of the platform for enterprise and research users, with potential to support stronger engagement and long‑term adoption in data‑intensive robotics and automation applications.